SlideShare a Scribd company logo

Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction

•
•
Estevam Hruschka
Estevam Hruschka

The Web is inundated with information in many different formats including semi-structured and unstructured data. Machine Reading is a research area that aims to build systems that can read natural-language-based information, extracting knowledge and storing it into knowledge bases. Thus, Machine Reading systems are developed to produce language- understanding technology that will automatically process text in affordable time. In this tutorial the idea of automatically reading the Web using Machine Reading techniques will be explored. Four of the most successful Machine Reading approaches in- tended to Read the Web (namely KnowItAll, Yago, NELL and DBPedia systems) will be presented and discussed. The principles, the subtleties, as well as current results of each approach will be addressed. On-line resources (from each approach) will be explored and the future directions in each system will be pointed out. YAGO, KnowItAll, NELL and DBPedia are not the only research efforts focusing on Reading the Web. They were selected, to be presented in this tutorial, because they show four different and very relevant approaches to this problem, but it does not mean they are the only relevant approaches at all. In spite of mainly focusing on the four aforementioned systems, some other independent contributions on the Read the Web idea will be mentioned and pointed out as related works.

Data & Analytics
1 of 343
Download to read offline
Estevam  R.  Hruschka  Jr.   Federal  University  of  São  Carlos   Machine Reading the Web: Beyond Named Entity Recognition and Relation Extraction  
Disclaimers   •  Previous  versions  of  this  tutorial  were  presented  at   IBERAMIA2012  (h@p://iberamia2012.dsic.upv.es/tutorials/)   and  WWW2013  (h@p://www2013.org/program/machine-­‐ reading-­‐the-­‐web/).  Also,  a  short  version  was  presented  at   ECMLPKDD2015  Summer  School(h@p:// www.ecmlpkdd2015.org/summer-­‐school/ss-­‐schedule).   •  Feel  free  to  e-­‐mail  me  (estevam.hruschka@gmail.com)  with   quesTons  about  this  tutorial  or  any  feedback/suggesTons/ criTcisms.  Your  feedback  can  help  improving  the  quality  of   these  slides,  thus,  they  are  very  welcome.   •  As  in  many  tutorials’  slides,  these  slides  were  prepared  to  be   presented,  and  la@er  studied.  Thus,  they  are  meant  to  be   more  self-­‐contained  than  slides  from  a  paper  presentaTon.  
Disclaimers   •  Due  to  Tme  constraints,  I  do  not  intend  to  cover  all  the   algorithms  and  publicaTons  related  to  YAGO,  KnowItAll,  NELL   and  DBPedia.  What  I  do  intend,  instead,  is  to  give  an  overview   of  all  four  projects  and  what  is  the  main  approach  to  “Read   the  Web”,  used  in  each  project.     •  YAGO,  KnowItAll,  NELL  and  DBPedia  are  not  the  only  research   efforts  focusing  on  “Reading  the  Web”.  They  were  selected,   to  be  presented  in  this  tutorial,  because  they  represent  four   different  and  very  relevant  approaches  to  this  problem,  but  it   does  not  mean  they  are  the  best  (or  the  only  relevant)  ones   at  all.  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – YAGO   – KnowItAll   – NELL   – DBPedia  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – YAGO   – KnowItAll   – NELL   – DBPedia  
Picture  taken  from  [Fern,  2008]    
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – YAGO   – KnowItAll   – NELL   – DBPedia  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – YAGO   – KnowItAll   – NELL   – DBPedia  
Picture  taken  from  [DARPA,  2012]    
Picture  taken  from  [DARPA,  2012]    
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
DBPedia  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
The  YAGO-­‐NAGA  Project:   HarvesAng,  Searching,  and  Ranking   Knowledge  from  the  Web    
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
KnowItAll  
KnowItAll:  Open  InformaTon  ExtracTon  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
NELL  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
Machine  Learning   •  What  is  Machine  Learning?   The  field  of  Machine  Learning  seeks  to  answer   the  quesTon   “How  can  we  build  computer  systems  that   automaTcally  improve  with  experience,  and   what  are  the  fundamental  laws  that  govern  all   learning  processes?”  [Mitchell,  2006]  
Machine  Learning   •  What  is  Machine  Learning?   a  machine  learns  with  respect  to  a  parTcular:   -­‐  task  T     -­‐  performance  metric  P   -­‐  type  of  experience  E       if  the  system  reliably  improves  its  performance  P  at   task  T,  following  experience  E.  [Mitchell,  1997]  
Machine  Learning   •  Examples  of  Machine  Learning  approaches   for  different  tasks  (T),  performance  metrics   (P)  an  experiences  (E)   -­‐  data  mining   -­‐  autonomous  discovery   -­‐  database  updaTng   -­‐  programming  by  example   -­‐  Pa@ern  recogniTon    
Machine  Learning   •  Supervised  Learning;   •  Unsupervised  Learning   •  Semi-­‐Supervised  Learning  
Supervised  Learning  
Supervised  Learning   (one  simple  anecdotal  approach)   0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)   ?????????   What  model   should  be   chosen?   ?????????  
0   5   10   15   20   25   0   5   10   15   20   25   Unsupervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Unsupervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Unsupervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Unsupervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   ?????????   What  model   should  be   chosen?   ?????????   Unsupervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)   ?????????   What  model   should  be   chosen?   ?????????  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
Machine  Reading   •  “The  autonomous  understanding  of   text”  [Etzioni  et  al.,  2007]   •  “One  of  the  most  important  methods  by  which   human  beings  learn  is  by  reading”  [Clark  et  al.,   2007],  thus  why  not  building  machines  capable   of  learning  by  reading?  
Machine  Reading   •  “The  problem  of  deciding  what  was  implied  by  a   wri@en  text,  of  reading  between  the  lines  is  the   problem  of  inference.”  [Norvig,  2007]     •  Typically,  Machine  Reading  is  different  from   Natural  Language  Processing  alone  
It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                                                                   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
Machine  Reading                                                       same                                               This  slide  was  adapted  from  [Hady  et  al.,  2011]     It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill."
It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                                               This  slide  was  adapted  from  [Hady  et  al.,  2011]    
It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                               uncleOf           owns   hires           headOf   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                               uncleOf           owns   hires           headOf   affairWith   affairWith   enemyOf   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
Machine  Reading   •  One  important  (ini6al)  approach  to  machine   reading  is  to  extract  facts  from  text  and  store   them  in  a  structured  form.   •  Facts  can  be  seen  as  enTTes  and  their   relaTons   •  Ontology  is  one  of  the  most  common   representaTon  for  the  extracted  facts    
Machine  Reading  
Machine  Reading  
Machine  Reading  
Machine  Reading  
Machine  Reading  
Machine  Reading  
Machine  Reading  
Machine  Reading  
Machine  Reading   •  Named  EnTty  ResoluTon/RecogniTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Polysemy  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
Machine  Reading   •  Named  EnTty  ResoluTon/RecogniTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Polysemy  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
Machine  Reading   •  Named  EnTty  ResoluTon/RecogniTon   –  Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.     –  Free  text   •  Hearst-­‐pa@erns;  clustering  by  verbal  phrases   •  Natural-­‐language  processing   •  Advanced  pa@erns  &  iteraTve  bootstrapping    (“Dual  IteraTve  Pa@ern  RelaTon  ExtracTon”)  
Named  EnTty  RecogniTon   •  Named  EnTty  RecogniTon  [Nadeau  &  Sekine,   2007]   – term  “Named  EnTty”  coined  for  the  Sixth  Message   Understanding  Conference  (MUC-­‐6)  (R.  Grishman   &  Sundheim  1996).     – important  sub-­‐tasks  of  IE  called  “Named  EnTty   RecogniTon  and  ClassificaTon  (NERC)”.  
•  recognize  informaTon  units  like  names,   including  person,  organizaAon  and  locaAon   names,  and  numeric  expressions  including   Ame,  date,  money  and  percent  expressions.   •  In  Machine  Reading,  many  other  enTTes:   product,  kitchen  item,  sport,  etc.   Named  EnTty  RecogniTon   [Nadeau  &  Sekine,  2007]  
•  Named  EnTty  ResoluTon  [Theobald  &  Weikum,   2012]   – Which  individual  enTTes  belong  to  which  classes?   •  instanceOf  (Surajit  Chaudhuri,  computer  scien6sts),   •  instanceOf  (BarbaraLiskov,  computer  scien6sts),   •  instanceOf  (Barbara  Liskov,  female  humans),  …   Named  EnTty  ResoluTon  
•  Named  EnTTes  RecogniTon  as  a  machine   learning  task.   – Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   Named  EnTty  RecogniTon  
•  Named  EnTty  RecogniTon  as  a  Machine  Learning  task.   –  Supervised  Learning   –  Possible  features  [RaTnov  &  Roth,  2009],  [Khambhatla,   2004],  [Zhou  et.  al.  2005]   •   Words  “around”  and  including  enTTes     •  POS  (Part-­‐Of-­‐Speech)   •  Prefixes  and  suffixes   •  CapitalizaTon   •  Number  of  words   •  Number  of  characters   •  First  word,  last  word   •  gaze@eer  matches       Named  EnTty  RecogniTon  
•  Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   Named  EnTty  RecogniTon  
•  Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   Kernels   Named  EnTty  RecogniTon  
•  Supervised  Learning  using  Kernels   – A  Kernel  defines  similarity  implicitly  in  a  higher   dimensional  space     – Can  be  based  on  Strings,  Word  Sequences,  Parse   Trees,  etc.   •  For  strings  similarity∝  number  of  common  substrings   (or  subsequences)     •  Recommended  reading  on  string  kernels  [Lodhi  et.  al.,   2002]       Named  EnTty  RecogniTon  [Bach  &  Badaskar,  2007]  
[Bach  &  Badaskar,  2007]   Named  EnTty  RecogniTon  
•  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     Set  of   labeled  Pa@ern   Examples   Named  EnTty  RecogniTon  
•  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon  
•  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon  
Set  of  labeled  Instances     •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon  
Set  of  labeled  Instances     •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon   Google   Apple  
  Set  of  labeled  Instances       •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon   Google   Apple   NE   Pa@ern   Classifier  
  Set  of  labeled  Instances       •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon   Google   Apple   NE   Pa@ern   Classifier   What  about   unsupervised?  
  Set  of  labeled  Instances       •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   Named  EnTty  RecogniTon   NE   Pa@ern   Classifier   What  about   unsupervised?  
  Set  of  labeled  Instances       •  Unsupervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   Named  EnTty  RecogniTon   NE   Pa@ern   Classifier  
•  [RaTnov  &  Roth,  2009]   Named  EnTty  RecogniTon  
•  [Pennington  &  Socher  &  Manning,  2014]   Named  EnTty  RecogniTon  
Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Polysemy  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
Machine  Reading   •  RelaTon  ExtracTon   –  Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.     –  Free  text   •  Hearst-­‐pa@erns;  clustering  by  verbal  phrases   •  Natural-­‐language  processing   •  Advanced  pa@erns  &  iteraTve  bootstrapping    (“Dual  IteraTve  Pa@ern  RelaTon  ExtracTon”)  
Machine  Reading   •  RelaTon  ExtracTon  [Theobald  &  Weikum,  2012]   –  Which  instances  (pairs  of  individual  enTTes)  are  there   for  given  binary  relaTons  with  specific  type   signatures?   •  hasAdvisor  (JimGray,  MikeHarrison)   •  hasAdvisor  (HectorGarcia-­‐Molina,  Gio  Wiederhold)   •  hasAdvisor  (Susan  Davidson,  Hector  Garcia-­‐Molina)   •  graduatedAt  (JimGray,  Berkeley)   •  graduatedAt  (HectorGarcia-­‐Molina,  Stanford)   •  hasWonPrize  (JimGray,  TuringAward)   •  bornOn  (JohnLennon,  9Oct1940)   •  diedOn  (JohnLennon,  8Dec1980)   •  marriedTo  (JohnLennon,  YokoOno)  
RelaTon  ExtracTon   •  ExtracTng  semanTc  relaTons  between  enTTes   in  text   •  RelaTon  extracTon  as  a  Machine  Learning  task.   – Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   [Bach  &  Badaskar,  2007]  
RelaTon  ExtracTon   •  RelaTon  extracTon  as  a  Machine  Learning  task.   –  Supervised  Learning   –  Possible  features  [Khambhatla,  2004],  [Zhou  et.  al.   2005]   •   Words  between  and  including  enTTes     •  Types  of  enTTes  (person,  locaTon,  etc)     •  Number  of  enTTes  between  the  two  enTTes,  whether  both   enTTes  belong  to  same  chunk     •  #  words  separaTng  the  two  enTTes     •  Path  between  the  two  enTTes  in  a  parse  tree       [Bach  &  Badaskar,  2007]  
RelaTon  ExtracTon   •  ExtracTng  semanTc  relaTons  between  enTTes   in  text   •  RelaTon  extracTon  as  a  classificaTon  task.   – Supervised  Learning     NLP  tools   (POS,  Parse   Trees,  NER)   text   Features   ExtracTon   Classifier   [Bach  &  Badaskar,  2007]  
RelaTon  ExtracTon   •  ExtracTng  semanTc  relaTons  between  enTTes   in  text   •  RelaTon  extracTon  as  a  classificaTon  task.   – Supervised  Learning     NLP  tools   (POS,  Parse   Trees,  NER)   text   Features   ExtracTon   Classifier   Kernels   [Bach  &  Badaskar,  2007]  
RelaTon  ExtracTon   •  Supervised  Learning  using  Kernels   – A  Kernel  defines  similarity  implicitly  in  a  higher   dimensional  space     – Can  be  based  on  Strings,  Word  Sequences,  Parse   Trees,  etc.   •  For  strings,  similarity∝  number  of  common  substrings   (or  subsequences)     •  Recommended  reading  on  string  kernels  [Lodhi  et.  al.,   2002]       [Bach  &  Badaskar,  2007]  
RelaTon  ExtracTon  [Bach  &  Badaskar,  2007]  
RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples  
RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X    
RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Pair  of   Instances   Classifier  
RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Pair  of   Instances   Classifier  
RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Pair  of   Instances   Classifier     Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Google-­‐Mountain  View   Apple-­‐CuperAno  
RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Google-­‐Mountain  View   Apple-­‐CuperAno   Pair  of   Instances   Classifier  
RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Google-­‐Mountain  View   Apple-­‐CuperAno   Pair  of   Instances   Classifier   What  about   unsupervised?  
RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   What  about   unsupervised?  
RelaTon  ExtracTon   •  Unsupervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier  
•  Supervised  learning  [Bunescu  &  Mooney,  2005]   •  Distant  and  ParTal  Supervised  [Angeli  &   Tibshirani  &  Wu  &  Manning,  2014]   RelaTon  ExtracTon  
Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Polysemy  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  taken  from:  h@p://nlp.stanford.edu/projects/coref.shtml    
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  taken  from:  h@p://nlp.stanford.edu/projects/coref.shtml     within-document co-reference
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  taken  from:  h@p://nlp.stanford.edu/projects/coref.shtml     within-document co-reference
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   computer     Apple   Computer    
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   apple   computer     Apple   Computer    
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   apple   computer     Apple   Computer     cross-document co-reference
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the  same   enTty   •  Which  names  denote  which  enTTes?  [Theobald   &  Weikum,  2012]   –  means  (“Lady  Di“,  Diana  Spencer),   –  means  (“Diana  Frances  Mountba@en-­‐Windsor”,  Diana   Spencer),  …   –  means  (“Madonna“,  Madonna  Louise  Ciccone),   –  means  (“Madonna“,  Madonna(painTng  by  Edward   Munch)),  …   cross-document co-reference
Co-­‐Reference  and  Polysemy  ResoluTon   •  Polysemy:  is  the  capacity  for  a  sign  (such  as  a   word,  phrase,  or  symbol)  to  have  mulTple   meanings  [Wikipedia]  
Co-­‐Reference  and  Polysemy  ResoluTon   •  Polysemy:  is  the  capacity  for  a  sign  (such  as  a   word,  phrase,  or  symbol)  to  have  mulTple   meanings  [Wikipedia]   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   apple     (the  fruit)   Apple   Computer    
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐Reference  and  Polysemy   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   apple   computer     apple     (the  fruit)   Apple   Computer    
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference  and  Polysemy:     – Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier  
•  Co-­‐Reference   ResoluTon.   – Supervised   Learning   – Possible   features   [Bengtson  &   Roth,  2008]     Co-­‐Reference  and  Polysemy  ResoluTon  
•  Co-­‐Reference   ResoluTon.   – Supervised   Learning   – Possible   features   [Bengtson  &   Roth,  2008]     Co-­‐Reference  and  Polysemy  ResoluTon  
Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference  and  Polysemy:     – Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   Kernels  
•  Supervised  Learning  using  Kernels   – A  Kernel  defines  similarity  implicitly  in  a  higher   dimensional  space     – Can  be  based  on  Strings,  Word  Sequences,  Parse   Trees,  etc.   •  For  strings  similarity∝  number  of  common  substrings   (or  subsequences)     •  Recommended  reading  on  string  kernels  [Lodhi  et.  al.,   2002]       Co-­‐Reference  and  Polysemy  ResoluTon  
•  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   Co-­‐Reference  and  Polysemy  ResoluTon  
•  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   X  also  know  as  Y     Co-­‐Reference  and  Polysemy  ResoluTon  
•  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   Pair  of   Instances   Classifier   Co-­‐Reference  and  Polysemy  ResoluTon   X  also  know  as  Y    
  Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   X  also  know  as  Y    
Pair  of   Instances   Classifier     Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Apple  Computer  -­‐   Apple   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   X  also  know  as  Y    
  Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   Apple  Computer  -­‐   Apple   X  also  know  as  Y    
  Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   Apple  Computer  -­‐   Apple   X  also  know  as  Y     What  about   unsupervised?  
  Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   What  about   unsupervised?  
  Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon  
•  Co-­‐Reference  ResoluTon:     [Singh  et  al.,  2011],  [Krishnamurthy  &  Mitchell,   2011],[Du@a  &  Weikum,  2015]   •  Polysemy  ResoluTon:   [Krishnamurthy  &  Mitchell,  2011],  [Galárraga  et   al.,    2014]   Co-­‐Reference  and  Polysemy  ResoluTon  
Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Synonym  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
Machine  Reading   •  RelaTon  Discovery   – Which  new  relaTons  are  there  for  given  pair  of   enTTes?   •  hasAdvisor  (JimGray,  MikeHarrison)  
Machine  Reading   •  RelaTon  Discovery   – Which  new  relaTons  are  there  for  given  pair  of   enTTes?   •  hasAdvisor  (JimGray,  MikeHarrison)   •  hasCoAuthor(HectorGarcia-­‐Molina,  Gio  Wiederhold)  
Machine  Reading   •  RelaTon  Discovery   – Which  new  relaTons  are  there  for  given  pair  of   enTTes?   •  hasAdvisor  (JimGray,  MikeHarrison)   •  hasCoAuthor(HectorGarcia-­‐Molina,  Gio  Wiederhold)   •  graduatedAt  (JimGray,  Berkeley)  
Machine  Reading   •  RelaTon  Discovery   – Which  new  relaTons  are  there  for  given  pair  of   enTTes?   •  hasAdvisor  (JimGray,  MikeHarrison)   •  hasCoAuthor(HectorGarcia-­‐Molina,  Gio  Wiederhold)   •  graduatedAt  (JimGray,  Berkeley)   •  studiedAt  (HectorGarcia-­‐Molina,  Stanford)   •  bornOn  (JohnLennon,  9Oct1940)   •  releasedAlbum  (JohnLennon,  10Dec1965)  
  Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   RelaTon  Discovery   Clustering   Algorithm  
Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Synonym  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
Inference   •  Inference  is  the  act  or  process  of  deriving   logical  conclusions  from  premises  known  or   assumed  to  be  true  [Wikipedia]  
Inference   •  Manually  craved  inference  rules   •  AutomaTcally  learned  inference  rules   •  Data  mining  the  Knowledge  Base  
Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Synonym  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
Machine  Reading   •  Ontology  RepresentaTon    Facts  (RDF  triples)   1:  (Jim,  hasAdvisor,  Mike)   2:  (Surajit,  hasAdvisor,  Jeff)   3:  (Madonna,  marriedTo,  GuyRitchie)   4:  (Nicolas,  marriedTo,  Carla)   5:  (ManchesterU,  wonCup,  ChampionsLeague)   ReificaTon:   “Facts  about  Facts”:   6:      (1,  inYear,  1968)   7:      (2,  inYear,  2006)   8:      (3,  validFrom,  22-­‐Dec-­‐2000)     9:      (3,  validUnTl,  Nov-­‐2008)   10:  (4,  validFrom,  2-­‐Feb-­‐2008)   11:  (2,  source,  SigmodRecord)   12:  (5,  inYear,  1999)   13:  (5,  locaTon,  CampNou)   14:  (5,  source,  Wikipedia)  
Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Synonym  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
Document/Sentence  UnderstanTng     (MicroRead)   •  “The  scienTst  observed  the  bu@erfly  with  the   blue  circle”      
Document/Sentence  UnderstanTng     (MicroRead)   •  “The  scienTst  observed  the  bu[erfly  with  the   blue  circle”      
Document/Sentence  UnderstanTng     (MicroRead)   •  “The  scienTst  observed  the  bu[erfly  with  the   blue  circle”       •  “The  scienTst  observed  the  bu@erfly  with  the   blue  microscope”  
Document/Sentence  UnderstanTng     (MicroRead)   •  “The  scienTst  observed  the  bu[erfly  with  the   blue  circle”       •  “The  scienAst  observed  the  bu@erfly  with  the   blue  microscope”  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
DBPedia   h@p://wiki.dbpedia.org/  
DBPedia   Mapping  Wikipedia  semi-­‐structured  data  into  RDF  triples  
DBPedia   Mapping  Wikipedia  semi-­‐structured  data  into  RDF  triples   Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”  
DBPedia   •  How  to  Read  Wikipedia  Semi-­‐structured  data?   [Lehmann  et  al.,  2014]   – Parse  Wikipedia  Markup  language   – Overcome  the  lack  of  standard  problem   •  Same  properTes  might  have  different  names   •  “Datebirth”  and  “Birth_date”   •  “Birthplace”  and  “Birth_place”   – Instead  of  “Modeling  the  World”,  try  to  structure   the  available  informaTon  
DBPedia  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
The  YAGO-­‐NAGA  Project:   HarvesAng,  Searching,  and  Ranking  Knowledge   from  the  Web    
The  YAGO-­‐NAGA  Project:   HarvesAng,  Searching,  and  Ranking   Knowledge  from  the  Web    
YAGO   •  Yet  Another  Great  Ontology  -­‐  YAGO   •  Main  Goal:  building  a  conveniently  searchable,   large-­‐scale,  highly  accurate  knowledge  base  of   common  facts  in  a  machine-­‐processable   representaTon  
YAGO  
YAGO   •  Turn  Web  into  Knowledge  Base  [Weikum  et   al.,  2009]   – Building  a  comprehensive  Knowledge  Base  of   human  knowledge   – knowledge  from  Wikipedia  and  WordNet   – the  ontology  check  itself  for  precision    
YAGO   •  The  knowledge  base  is  automaTcally   constructed  from  Wikipedia   •  Each  arTcle  in  Wikipedia  becomes  an  enTty  in   the  kb  (e.g.,  since  Leonard  Cohen  has  an   arTcle  in  Wikipedia,  LeonardCohen  becomes   an  enTty  in  YAGO).    
YAGO  
YAGO   Free   Text  
YAGO   Free   Text  
YAGO   Free   Text   InfoBox  
YAGO   Wikipedia  InfoBox  
YAGO   Wikipedia  InfoBox   Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”  
YAGO   Wikipedia  InfoBox   Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”  
YAGO   •  Certain  categories  are  exploited  to  deliver   type  informaTon  (e.g.,  the  arTcle  about   Leonard  Cohen  is  in  the  category  Canadian   male  poets,  so  he  becomes  a  Canadian  poet).    
YAGO  
YAGO  
YAGO   •  For  each  category  of  a  page  [Hoffart  et  al.,  2012]   –  Using  shallow  parsing,  determine  the  head  word  of  the   category  name.  In  the  example  of  Canadian  poets,  the   head  word  is  poets.     –  If  the  head  word  is  in  plural,  then  proposes  the  category  as   a  class  and  the  arTcle  enTty  as  an  instance     –  Link  the  class  to  the  WordNet  taxonomy  (most  frequent   sense  of  the  head  word  in  WordNet)   •  only  countable  nouns  can  appear  in  plural  form   •  only  countable  nouns  can  be  ontological  classes   •  themaTc  categories  (such  as  Canadian  poetry)  are   different  from  conceptual  Categories  
YAGO   •  head  words  that  are  not  conceptual  even  though   they  appear  in  plural  (such  as  stubs  in  Canadian   poetry  stubs)  are  in  the  first  list  of  excepTons.     •  words  that  do  not  map  to  their  most  frequent   sense,  but  to  a  different  sense  are  in  the  second   excepTon  list   –  The  word  capital,  e.g.,  refers  to  the  main  city  of  a   country  in  the  majority  of  cases  and  not  to  the   financial  amount,  which  is  the  most  frequent  sense  in   WordNet.  
YAGO   •  About  100  manually  defined  relaTons   –  wasBornOnDate     –  locatedIn     –  hasPopulaTon     •  Categories  and  infoboxes  are  exploited  to  deliver  facts   (instances  of  relaTons).     •  Manually  defined  pa@erns  that  map  categories  and   infobox  a@ributes  to  fact  templates   –  infobox  a@ribute  born=Montreal,  thus   wasBornIn(LeonardCohen,  Montreal)     •  Pa@ern-­‐based  extracTons  resulted  in  2  million   extracted  enTTes  and  20  million  facts  
YAGO   •  Based  on  declaraTve  rules  (stored  in  text  files)   •  The  rules  take  the  form  of  subject-­‐  predicate-­‐ object  triples,  so  that  they  are  basically   addiTonal  facts   •  There  are  different  types  of  rules  
YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.   Knowledge   RepresentaTon  
YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.     Inference    
YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  Knowledge   RepresentaTon  
YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.   InformaTon   ExtracTon  
YAGO   •  AutomaTcally  verifies  consistency   – Check  uniqueness  of  funcTonal  arguments   •  spouse(x,y)  ∧  diff(y,z)  ⇒  ¬spouse(x,z)   – Check  domains  and  ranges  of  relaTons   •  spouse(x,y)  ⇒  female(x)   •  spouse(x,y)  ⇒  male(y)   •  spouse(x,y)  ⇒  (f(x)∧m(y))  ∨  (m(x)∧f(y))      
YAGO   •  AutomaTcally  verifies  consistency   – Hard  Constraint   •  hasAdvisor(x,y)  ∧  graduatedInYear(x,t)  ∧  graduatedInYear(y,s)  ⇒  s  <  t   – Sov  Constraint     •  firstPaper(x,p)  ∧  firstPaper(y,q)  ∧  author(p,x)  ∧  author(p,y)  )  ∧      inYear(p)  >  inYear(q)  +  5years  ⇒  hasAdvisor(x,y)  [0.6]    
YAGO   •  AutomaTcally  verifies  consistency   – Hard  Constraint   •  hasAdvisor(x,y)  ∧  graduatedInYear(x,t)  ∧  graduatedInYear(y,s)  ⇒  s  <  t   – Sov  Constraint     •  firstPaper(x,p)  ∧  firstPaper(y,q)  ∧  author(p,x)  ∧  author(p,y)  )  ∧      inYear(p)  >  inYear(q)  +  5years  ⇒  hasAdvisor(x,y)  [0.6]     Inference  
YAGO   •  Ontology  RepresentaTon   – EnTTes  and  RelaTons  of  public  interest   – Format:  TSV,  RDF,  XML,  N3,  Web  Interface   – Learns   •  Instances  and  pa@erns  from  Wikipedia;   •  Taxonomy  from  WordNet;   •  Geotags  informaTon  from  Geonames.  
YAGO   •  Named  EnTty  ResoluTon/ExtracTon  [Theobald  &   Weikum,  2012]   – Based  on  rules  and  pa@erns  extracted  from   Wikipedia   – DisambiguaTon  is  a  relevant  issue   – Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.  
YAGO   •  Named  EnTty  ResoluTon/ExtracTon  [Theobald  &   Weikum,  2012]   – Based  on  rules  and  pa@erns  extracted  from   Wikipedia   – DisambiguaTon  is  a  relevant  issue   – Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.   Natural  Language   Processing   Machine   Learning  
It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                                                       This  slide  was  adapted  from  [Hady  et  al.,  2011]    
It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                                                                   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
YAGO   •  RelaTon  ExtracTon  [Theobald  &  Weikum,  2012]   – Based  on  rules  and  pa@erns  extracted  from   Wikipedia   – Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.  
YAGO   •  RelaTon  ExtracTon  [Theobald  &  Weikum,  2012]   – Based  on  rules  and  pa@erns  extracted  from   Wikipedia   – Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.   Natural  Language   Processing   Machine   Learning  
It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                                                                   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
Machine  Reading                                                       same                                               This  slide  was  adapted  from  [Hady  et  al.,  2011]     It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill."
It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                                               This  slide  was  adapted  from  [Hady  et  al.,  2011]    
It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                               uncleOf           owns   hires           headOf   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
Machine  Reading   This  slide  was  taken  from  [Hoffart  et  al.,  2015]    
YAGO   •  YAGO2:  Exploring  and  Querying  World   Knowledge  in  Time,  Space,  Context,  and  Many   Languages   – New  relaTons  specifically  designed  to  cover  Tme,   space  and  context   – Wikipedia  translated  pages  as  sources  for  other   languages  
YAGO   •  YAGO3  [Mahdisoltani  &  Biega  &  Suchanek,  2015]   –  an  extension  of  the  YAGO  knowledge  base;   –  built  from  the  Wikipedias  in  mulTple  languages.     –  fuses  the  mulTlingual  informaTon  with  the  English  WordNet   –  categories,  infoboxes,  and  Wikidata,  to  learn  the  meaning  of   infobox  a@ributes  across  languages   –  10  different  languages   –  precision  of  95%-­‐100%  in  the  a@ribute  mapping   –  enlarges  YAGO  by  1m  new  enTTes  and  7m  new  facts.    
YAGO   •  More  on  YAGO:   –  Very  nice  tutorials:   •  “Knowledge  Bases  for  Web  Content  AnalyTcs”  at  WWW   2015,  Florence,  May  2015.   •  "SemanTc  Knowledge  Bases  from  Web  Sources"  at  IJCAI   2011,  Barcelona,  July  2011   "HarvesTng  Knowledge  from  Web  Data  and  Text"  at  CIKM   2010,  Toronto,  October  2010   "From  InformaTon  to  Knowledge:  HarvesTng  EnTTes  and   RelaTonships  from  Web  Sources"  at  PODS  2010,   Indianapolis,  June  2010   –  Project  Website:   •  h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/  
YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)  
YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)  
YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)  
YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)   ?X  <hasChild>  ?C  ?Y  <hasChild>  ?C  =>  ?X  <isMarriedTo>  ?Y  
YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)   ?X  <hasChild>  ?C  ?Y  <hasChild>  ?C  =>  ?X  <isMarriedTo>  ?Y   Machine   Learning  
YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)   ?X  <hasChild>  ?C  ?Y  <hasChild>  ?C  =>  ?X  <isMarriedTo>  ?Y   Machine   Learning   Inference  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
KnowItAll  
KnowItAll:  Open  InformaTon  ExtracTon  
KnowItAll:  Open  InformaTon  ExtracTon  
KnowItAll   •  MoTvaTon:  New  Paradigm  for  Search  [Etzioni,  2008]   –  The  future  of  Web  Search   –  Read  the  Web  instead  of  retrieving  Web  pages  to   perform  Web  Search  
KnowItAll   •  InformaTon  ExtracTon  (IE)  +  tractable   inference     –  IE(sentence)  =  who  did  what?   •  speaker(P.  Smith,  ECMLPKDD2012)   –  Inference  =  uncover  implicit  informaTon   •  Will  Pi@sburgh  Steelers  be  champions  again?     •  Open  InformaTon  ExtracTon  [Banko  et  al.,  2007]  
Open  InformaTon  ExtracTon     [Banko  et  al.,  2007]   •  Open  IE  systems  avoid  specific  nouns  and   verbs     •  Extractors  are  unlexicalized—formulated  only   in  terms  of:   –   syntacTc  tokens  (e.g.,  part-­‐of-­‐speech  tags)     –  closed-­‐word  classes  (e.g.,  of,  in,  such  as).     •  Open  IE  extractors  focus  on  generic  ways  in   which  relaTonships  are  expressed  in  English   –  naturally  generalizing  across  domains.  
Open  InformaTon  ExtracTon     [Banko  et  al.,  2007]   •  Open  IE  extractors  focus  on  generic  ways  in   which  relaTonships  are  expressed  in  English   –  naturally  generalizing  across  domains.   RelaTon   Discovery  
Open  InformaTon  ExtracTon     •  Open  IE  systems  are  tradiTonally  based  on     three  steps  [Etzioni  et  al.,  2011]:   –  1.  Label:  Sentences  are  automaTcally  labeled  with   extracTons  using  heurisTcs  or  distant  supervision.   Unsupervised   Learning  
Open  InformaTon  ExtracTon     •  Open  IE  systems  are  tradiTonally  based  on     three  steps  [Etzioni  et  al.,  2011]:   –  1.  Label:  Sentences  are  automaTcally  labeled  with   extracTons  using  heurisTcs  or  distant  supervision.   –  2.  Learn:  A  relaTon  phrase  extractor  is  learned  using  a   sequence-­‐labeling  graphical  model  (e.g.,  CRF).   Supervised   Learning  
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction
Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction

Recommended

Machine Reading the Web
Machine Reading the WebMachine Reading the Web
Machine Reading the WebEstevam Hruschka
 
Towards Machine Comprehension of Spoken Content
Towards Machine Comprehension of Spoken ContentTowards Machine Comprehension of Spoken Content
Towards Machine Comprehension of Spoken ContentNVIDIA Taiwan
 
Learning from Dashcam Videos
Learning from Dashcam VideosLearning from Dashcam Videos
Learning from Dashcam VideosNVIDIA Taiwan
 
Deep Learning Models for Question Answering
Deep Learning Models for Question AnsweringDeep Learning Models for Question Answering
Deep Learning Models for Question AnsweringSujit Pal
 
Intro to Deep Learning for Question Answering
Intro to Deep Learning for Question AnsweringIntro to Deep Learning for Question Answering
Intro to Deep Learning for Question AnsweringTraian Rebedea
 
A012270104
A012270104A012270104
A012270104IOSR Journals
 
I1803065873
I1803065873I1803065873
I1803065873IOSR Journals
 
Q01725110114
Q01725110114Q01725110114
Q01725110114IOSR Journals
 

Recommended

Machine Reading the Web
Machine Reading the WebMachine Reading the Web
Machine Reading the WebEstevam Hruschka
 
Towards Machine Comprehension of Spoken Content
Towards Machine Comprehension of Spoken ContentTowards Machine Comprehension of Spoken Content
Towards Machine Comprehension of Spoken ContentNVIDIA Taiwan
 
Learning from Dashcam Videos
Learning from Dashcam VideosLearning from Dashcam Videos
Learning from Dashcam VideosNVIDIA Taiwan
 
Deep Learning Models for Question Answering
Deep Learning Models for Question AnsweringDeep Learning Models for Question Answering
Deep Learning Models for Question AnsweringSujit Pal
 
Intro to Deep Learning for Question Answering
Intro to Deep Learning for Question AnsweringIntro to Deep Learning for Question Answering
Intro to Deep Learning for Question AnsweringTraian Rebedea
 
A012270104
A012270104A012270104
A012270104IOSR Journals
 
I1803065873
I1803065873I1803065873
I1803065873IOSR Journals
 
Q01725110114
Q01725110114Q01725110114
Q01725110114IOSR Journals
 
L0102189100
L0102189100L0102189100
L0102189100IOSR Journals
 
L012228185
L012228185L012228185
L012228185IOSR Journals
 
F1303042428
F1303042428F1303042428
F1303042428IOSR Journals
 
Endla soostikku matk
Endla soostikku matkEndla soostikku matk
Endla soostikku matkKaspar Kask
 
Greence
GreenceGreence
Greenceplatonkaterinis
 
vega rochat cv 2014
vega rochat cv 2014vega rochat cv 2014
vega rochat cv 2014vega rochat
 
curriculam vitae
curriculam vitaecurriculam vitae
curriculam vitaevipan dhadwal
 
C012431215
C012431215C012431215
C012431215IOSR Journals
 
E1103013743
E1103013743E1103013743
E1103013743IOSR Journals
 
Service graphics 2015
Service graphics 2015Service graphics 2015
Service graphics 2015John Cornfield
 
B010421824
B010421824B010421824
B010421824IOSR Journals
 
J012626269
J012626269J012626269
J012626269IOSR Journals
 
I012265155
I012265155I012265155
I012265155IOSR Journals
 
E010323842
E010323842E010323842
E010323842IOSR Journals
 
R01741124127
R01741124127R01741124127
R01741124127IOSR Journals
 
H011114758
H011114758H011114758
H011114758IOSR Journals
 
vega rochat portfolio 2014
vega rochat portfolio 2014vega rochat portfolio 2014
vega rochat portfolio 2014vega rochat
 
Stories About Renraku — the new Quality Model of Pharo (esug2016)
Stories About Renraku — the new Quality Model of Pharo (esug2016)Stories About Renraku — the new Quality Model of Pharo (esug2016)
Stories About Renraku — the new Quality Model of Pharo (esug2016)Yuriy Tymchuk
 
Stories About Renraku — the new Quality Model of Pharo
Stories About Renraku — the new Quality Model of PharoStories About Renraku — the new Quality Model of Pharo
Stories About Renraku — the new Quality Model of PharoESUG
 
Troubleshooting Deep Neural Networks - Full Stack Deep Learning
Troubleshooting Deep Neural Networks - Full Stack Deep LearningTroubleshooting Deep Neural Networks - Full Stack Deep Learning
Troubleshooting Deep Neural Networks - Full Stack Deep LearningSergey Karayev
 
Kanban in Action
Kanban in ActionKanban in Action
Kanban in ActionMarcus Hammarberg
 
Hpai class 16 - learning - 041320
Hpai class 16 - learning - 041320Hpai class 16 - learning - 041320
Hpai class 16 - learning - 041320melendez321
 

More Related Content

Viewers also liked

Endla soostikku matk
Endla soostikku matkEndla soostikku matk
Endla soostikku matkKaspar Kask
 
vega rochat cv 2014
vega rochat cv 2014vega rochat cv 2014
vega rochat cv 2014vega rochat
 
vega rochat portfolio 2014
vega rochat portfolio 2014vega rochat portfolio 2014
vega rochat portfolio 2014vega rochat
 

Viewers also liked (17)

L0102189100
L0102189100L0102189100
L0102189100
 
L012228185
L012228185L012228185
L012228185
 
F1303042428
F1303042428F1303042428
F1303042428
 
Endla soostikku matk
Endla soostikku matkEndla soostikku matk
Endla soostikku matk
 
Greence
GreenceGreence
Greence
 
vega rochat cv 2014
vega rochat cv 2014vega rochat cv 2014
vega rochat cv 2014
 
curriculam vitae
curriculam vitaecurriculam vitae
curriculam vitae
 
C012431215
C012431215C012431215
C012431215
 
E1103013743
E1103013743E1103013743
E1103013743
 
Service graphics 2015
Service graphics 2015Service graphics 2015
Service graphics 2015
 
B010421824
B010421824B010421824
B010421824
 
J012626269
J012626269J012626269
J012626269
 
I012265155
I012265155I012265155
I012265155
 
E010323842
E010323842E010323842
E010323842
 
R01741124127
R01741124127R01741124127
R01741124127
 
H011114758
H011114758H011114758
H011114758
 
vega rochat portfolio 2014
vega rochat portfolio 2014vega rochat portfolio 2014
vega rochat portfolio 2014
 

Similar to Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction

Stories About Renraku — the new Quality Model of Pharo (esug2016)
Stories About Renraku — the new Quality Model of Pharo (esug2016)Stories About Renraku — the new Quality Model of Pharo (esug2016)
Stories About Renraku — the new Quality Model of Pharo (esug2016)Yuriy Tymchuk
 
Stories About Renraku — the new Quality Model of Pharo
Stories About Renraku — the new Quality Model of PharoStories About Renraku — the new Quality Model of Pharo
Stories About Renraku — the new Quality Model of PharoESUG
 
Troubleshooting Deep Neural Networks - Full Stack Deep Learning
Troubleshooting Deep Neural Networks - Full Stack Deep LearningTroubleshooting Deep Neural Networks - Full Stack Deep Learning
Troubleshooting Deep Neural Networks - Full Stack Deep LearningSergey Karayev
 
Hpai class 16 - learning - 041320
Hpai class 16 - learning - 041320Hpai class 16 - learning - 041320
Hpai class 16 - learning - 041320melendez321
 
Using data science to analyse feedback for complex systems
Using data science to analyse feedback for complex systemsUsing data science to analyse feedback for complex systems
Using data science to analyse feedback for complex systemsKaren Pudner
 
CM UTaipei Kaggle Share
CM UTaipei Kaggle ShareCM UTaipei Kaggle Share
CM UTaipei Kaggle Share志明 陳
 
Machine Learning Experimentation at Sift Science
Machine Learning Experimentation at Sift ScienceMachine Learning Experimentation at Sift Science
Machine Learning Experimentation at Sift ScienceSift Science
 
Deep learning in retail london
Deep learning in retail londonDeep learning in retail london
Deep learning in retail londonDenis Troyanov
 
Machine Learning for (DF)IR with Velociraptor: From Setting Expectations to a...
Machine Learning for (DF)IR with Velociraptor: From Setting Expectations to a...Machine Learning for (DF)IR with Velociraptor: From Setting Expectations to a...
Machine Learning for (DF)IR with Velociraptor: From Setting Expectations to a...Chris Hammerschmidt
 
The deep bootstrap framework review
The deep bootstrap framework reviewThe deep bootstrap framework review
The deep bootstrap framework reviewtaeseon ryu
 
The Art Of Performance Tuning - with presenter notes!
The Art Of Performance Tuning - with presenter notes!The Art Of Performance Tuning - with presenter notes!
The Art Of Performance Tuning - with presenter notes!Jonathan Ross
 
Recommender Systems from A to Z – Model Training
Recommender Systems from A to Z – Model TrainingRecommender Systems from A to Z – Model Training
Recommender Systems from A to Z – Model TrainingCrossing Minds
 
Bridging the Gap: Machine Learning for Ubiquitous Computing -- Applied Machin...
Bridging the Gap: Machine Learning for Ubiquitous Computing -- Applied Machin...Bridging the Gap: Machine Learning for Ubiquitous Computing -- Applied Machin...
Bridging the Gap: Machine Learning for Ubiquitous Computing -- Applied Machin...Thomas Ploetz
 
CM NCCU Class2
CM NCCU Class2CM NCCU Class2
CM NCCU Class2志明 陳
 
The UNE VLE Implementation: an initial overview
The UNE VLE Implementation: an initial overviewThe UNE VLE Implementation: an initial overview
The UNE VLE Implementation: an initial overviewmylescarrick
 
Java andml may17-v1
Java andml may17-v1Java andml may17-v1
Java andml may17-v1Ferosh Jacob
 
Begin with Machine Learning
Begin with Machine LearningBegin with Machine Learning
Begin with Machine LearningNarong Intiruk
 
Scalable Learning Technologies for Big Data Mining
Scalable Learning Technologies for Big Data MiningScalable Learning Technologies for Big Data Mining
Scalable Learning Technologies for Big Data MiningGerard de Melo
 

Similar to Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction (20)

Stories About Renraku — the new Quality Model of Pharo (esug2016)
Stories About Renraku — the new Quality Model of Pharo (esug2016)Stories About Renraku — the new Quality Model of Pharo (esug2016)
Stories About Renraku — the new Quality Model of Pharo (esug2016)
 
Stories About Renraku — the new Quality Model of Pharo
Stories About Renraku — the new Quality Model of PharoStories About Renraku — the new Quality Model of Pharo
Stories About Renraku — the new Quality Model of Pharo
 
Troubleshooting Deep Neural Networks - Full Stack Deep Learning
Troubleshooting Deep Neural Networks - Full Stack Deep LearningTroubleshooting Deep Neural Networks - Full Stack Deep Learning
Troubleshooting Deep Neural Networks - Full Stack Deep Learning
 
Kanban in Action
Kanban in ActionKanban in Action
Kanban in Action
 
Hpai class 16 - learning - 041320
Hpai class 16 - learning - 041320Hpai class 16 - learning - 041320
Hpai class 16 - learning - 041320
 
Using data science to analyse feedback for complex systems
Using data science to analyse feedback for complex systemsUsing data science to analyse feedback for complex systems
Using data science to analyse feedback for complex systems
 
An Introduction to Face Detection
An Introduction to Face DetectionAn Introduction to Face Detection
An Introduction to Face Detection
 
CM UTaipei Kaggle Share
CM UTaipei Kaggle ShareCM UTaipei Kaggle Share
CM UTaipei Kaggle Share
 
Machine Learning Experimentation at Sift Science
Machine Learning Experimentation at Sift ScienceMachine Learning Experimentation at Sift Science
Machine Learning Experimentation at Sift Science
 
Deep learning in retail london
Deep learning in retail londonDeep learning in retail london
Deep learning in retail london
 
Machine Learning for (DF)IR with Velociraptor: From Setting Expectations to a...
Machine Learning for (DF)IR with Velociraptor: From Setting Expectations to a...Machine Learning for (DF)IR with Velociraptor: From Setting Expectations to a...
Machine Learning for (DF)IR with Velociraptor: From Setting Expectations to a...
 
The deep bootstrap framework review
The deep bootstrap framework reviewThe deep bootstrap framework review
The deep bootstrap framework review
 
The Art Of Performance Tuning - with presenter notes!
The Art Of Performance Tuning - with presenter notes!The Art Of Performance Tuning - with presenter notes!
The Art Of Performance Tuning - with presenter notes!
 
Recommender Systems from A to Z – Model Training
Recommender Systems from A to Z – Model TrainingRecommender Systems from A to Z – Model Training
Recommender Systems from A to Z – Model Training
 
Bridging the Gap: Machine Learning for Ubiquitous Computing -- Applied Machin...
Bridging the Gap: Machine Learning for Ubiquitous Computing -- Applied Machin...Bridging the Gap: Machine Learning for Ubiquitous Computing -- Applied Machin...
Bridging the Gap: Machine Learning for Ubiquitous Computing -- Applied Machin...
 
CM NCCU Class2
CM NCCU Class2CM NCCU Class2
CM NCCU Class2
 
The UNE VLE Implementation: an initial overview
The UNE VLE Implementation: an initial overviewThe UNE VLE Implementation: an initial overview
The UNE VLE Implementation: an initial overview
 
Java andml may17-v1
Java andml may17-v1Java andml may17-v1
Java andml may17-v1
 
Begin with Machine Learning
Begin with Machine LearningBegin with Machine Learning
Begin with Machine Learning
 
Scalable Learning Technologies for Big Data Mining
Scalable Learning Technologies for Big Data MiningScalable Learning Technologies for Big Data Mining
Scalable Learning Technologies for Big Data Mining
 

Recently uploaded

Al Barsha Escorts $#$ O565212860 $#$ Escort Service In Al Barsha
Al Barsha Escorts $#$ O565212860 $#$ Escort Service In Al BarshaAl Barsha Escorts $#$ O565212860 $#$ Escort Service In Al Barsha
Al Barsha Escorts $#$ O565212860 $#$ Escort Service In Al BarshaAroojKhan71
 
Accredited-Transport-Cooperatives-Jan-2021-Web.pdf
Accredited-Transport-Cooperatives-Jan-2021-Web.pdfAccredited-Transport-Cooperatives-Jan-2021-Web.pdf
Accredited-Transport-Cooperatives-Jan-2021-Web.pdfadriantubila
 
Delhi Call Girls Punjabi Bagh 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
Delhi Call Girls Punjabi Bagh 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip CallDelhi Call Girls Punjabi Bagh 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
Delhi Call Girls Punjabi Bagh 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Callshivangimorya083
 
Vip Model Call Girls (Delhi) Karol Bagh 9711199171✔️Body to body massage wit...
Vip Model Call Girls (Delhi) Karol Bagh 9711199171✔️Body to body massage wit...Vip Model Call Girls (Delhi) Karol Bagh 9711199171✔️Body to body massage wit...
Vip Model Call Girls (Delhi) Karol Bagh 9711199171✔️Body to body massage wit...shivangimorya083
 
Ravak dropshipping via API with DroFx.pptx
Ravak dropshipping via API with DroFx.pptxRavak dropshipping via API with DroFx.pptx
Ravak dropshipping via API with DroFx.pptxolyaivanovalion
 
April 2024 - Crypto Market Report's Analysis
April 2024 - Crypto Market Report's AnalysisApril 2024 - Crypto Market Report's Analysis
April 2024 - Crypto Market Report's Analysismanisha194592
 
100-Concepts-of-AI by Anupama Kate .pptx
100-Concepts-of-AI by Anupama Kate .pptx100-Concepts-of-AI by Anupama Kate .pptx
100-Concepts-of-AI by Anupama Kate .pptxAnupama Kate
 
BDSM⚡Call Girls in Mandawali Delhi >༒8448380779 Escort Service
BDSM⚡Call Girls in Mandawali Delhi >༒8448380779 Escort ServiceBDSM⚡Call Girls in Mandawali Delhi >༒8448380779 Escort Service
BDSM⚡Call Girls in Mandawali Delhi >༒8448380779 Escort ServiceDelhi Call girls
 
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdfMarket Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdfRachmat Ramadhan H
 
CebaBaby dropshipping via API with DroFX.pptx
CebaBaby dropshipping via API with DroFX.pptxCebaBaby dropshipping via API with DroFX.pptx
CebaBaby dropshipping via API with DroFX.pptxolyaivanovalion
 
Introduction-to-Machine-Learning (1).pptx
Introduction-to-Machine-Learning (1).pptxIntroduction-to-Machine-Learning (1).pptx
Introduction-to-Machine-Learning (1).pptxfirstjob4
 
Discover Why Less is More in B2B Research
Discover Why Less is More in B2B ResearchDiscover Why Less is More in B2B Research
Discover Why Less is More in B2B Researchmichael115558
 
BabyOno dropshipping via API with DroFx.pptx
BabyOno dropshipping via API with DroFx.pptxBabyOno dropshipping via API with DroFx.pptx
BabyOno dropshipping via API with DroFx.pptxolyaivanovalion
 
Cheap Rate Call girls Sarita Vihar Delhi 9205541914 shot 1500 night
Cheap Rate Call girls Sarita Vihar Delhi 9205541914 shot 1500 nightCheap Rate Call girls Sarita Vihar Delhi 9205541914 shot 1500 night
Cheap Rate Call girls Sarita Vihar Delhi 9205541914 shot 1500 nightDelhi Call girls
 
Zuja dropshipping via API with DroFx.pptx
Zuja dropshipping via API with DroFx.pptxZuja dropshipping via API with DroFx.pptx
Zuja dropshipping via API with DroFx.pptxolyaivanovalion
 
Determinants of health, dimensions of health, positive health and spectrum of...
Determinants of health, dimensions of health, positive health and spectrum of...Determinants of health, dimensions of health, positive health and spectrum of...
Determinants of health, dimensions of health, positive health and spectrum of...shambhavirathore45
 
Carero dropshipping via API with DroFx.pptx
Carero dropshipping via API with DroFx.pptxCarero dropshipping via API with DroFx.pptx
Carero dropshipping via API with DroFx.pptxolyaivanovalion
 

Recently uploaded (20)

Al Barsha Escorts $#$ O565212860 $#$ Escort Service In Al Barsha
Al Barsha Escorts $#$ O565212860 $#$ Escort Service In Al BarshaAl Barsha Escorts $#$ O565212860 $#$ Escort Service In Al Barsha
Al Barsha Escorts $#$ O565212860 $#$ Escort Service In Al Barsha
 
Accredited-Transport-Cooperatives-Jan-2021-Web.pdf
Accredited-Transport-Cooperatives-Jan-2021-Web.pdfAccredited-Transport-Cooperatives-Jan-2021-Web.pdf
Accredited-Transport-Cooperatives-Jan-2021-Web.pdf
 
Delhi Call Girls Punjabi Bagh 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
Delhi Call Girls Punjabi Bagh 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip CallDelhi Call Girls Punjabi Bagh 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
Delhi Call Girls Punjabi Bagh 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
 
Vip Model Call Girls (Delhi) Karol Bagh 9711199171✔️Body to body massage wit...
Vip Model Call Girls (Delhi) Karol Bagh 9711199171✔️Body to body massage wit...Vip Model Call Girls (Delhi) Karol Bagh 9711199171✔️Body to body massage wit...
Vip Model Call Girls (Delhi) Karol Bagh 9711199171✔️Body to body massage wit...
 
Ravak dropshipping via API with DroFx.pptx
Ravak dropshipping via API with DroFx.pptxRavak dropshipping via API with DroFx.pptx
Ravak dropshipping via API with DroFx.pptx
 
April 2024 - Crypto Market Report's Analysis
April 2024 - Crypto Market Report's AnalysisApril 2024 - Crypto Market Report's Analysis
April 2024 - Crypto Market Report's Analysis
 
CHEAP Call Girls in Saket (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICE
CHEAP Call Girls in Saket (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICECHEAP Call Girls in Saket (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICE
CHEAP Call Girls in Saket (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICE
 
100-Concepts-of-AI by Anupama Kate .pptx
100-Concepts-of-AI by Anupama Kate .pptx100-Concepts-of-AI by Anupama Kate .pptx
100-Concepts-of-AI by Anupama Kate .pptx
 
BDSM⚡Call Girls in Mandawali Delhi >༒8448380779 Escort Service
BDSM⚡Call Girls in Mandawali Delhi >༒8448380779 Escort ServiceBDSM⚡Call Girls in Mandawali Delhi >༒8448380779 Escort Service
BDSM⚡Call Girls in Mandawali Delhi >༒8448380779 Escort Service
 
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdfMarket Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
 
CebaBaby dropshipping via API with DroFX.pptx
CebaBaby dropshipping via API with DroFX.pptxCebaBaby dropshipping via API with DroFX.pptx
CebaBaby dropshipping via API with DroFX.pptx
 
(NEHA) Call Girls Katra Call Now 8617697112 Katra Escorts 24x7
(NEHA) Call Girls Katra Call Now 8617697112 Katra Escorts 24x7(NEHA) Call Girls Katra Call Now 8617697112 Katra Escorts 24x7
(NEHA) Call Girls Katra Call Now 8617697112 Katra Escorts 24x7
 
Introduction-to-Machine-Learning (1).pptx
Introduction-to-Machine-Learning (1).pptxIntroduction-to-Machine-Learning (1).pptx
Introduction-to-Machine-Learning (1).pptx
 
Discover Why Less is More in B2B Research
Discover Why Less is More in B2B ResearchDiscover Why Less is More in B2B Research
Discover Why Less is More in B2B Research
 
BabyOno dropshipping via API with DroFx.pptx
BabyOno dropshipping via API with DroFx.pptxBabyOno dropshipping via API with DroFx.pptx
BabyOno dropshipping via API with DroFx.pptx
 
Cheap Rate Call girls Sarita Vihar Delhi 9205541914 shot 1500 night
Cheap Rate Call girls Sarita Vihar Delhi 9205541914 shot 1500 nightCheap Rate Call girls Sarita Vihar Delhi 9205541914 shot 1500 night
Cheap Rate Call girls Sarita Vihar Delhi 9205541914 shot 1500 night
 
Zuja dropshipping via API with DroFx.pptx
Zuja dropshipping via API with DroFx.pptxZuja dropshipping via API with DroFx.pptx
Zuja dropshipping via API with DroFx.pptx
 
Determinants of health, dimensions of health, positive health and spectrum of...
Determinants of health, dimensions of health, positive health and spectrum of...Determinants of health, dimensions of health, positive health and spectrum of...
Determinants of health, dimensions of health, positive health and spectrum of...
 
Carero dropshipping via API with DroFx.pptx
Carero dropshipping via API with DroFx.pptxCarero dropshipping via API with DroFx.pptx
Carero dropshipping via API with DroFx.pptx
 
Abortion pills in Doha Qatar (+966572737505 ! Get Cytotec
Abortion pills in Doha Qatar (+966572737505 ! Get CytotecAbortion pills in Doha Qatar (+966572737505 ! Get Cytotec
Abortion pills in Doha Qatar (+966572737505 ! Get Cytotec
 

Machine Reading the Web: beyond Named Entity Recognition and Relation Extraction

  • 1. Estevam  R.  Hruschka  Jr.   Federal  University  of  São  Carlos   Machine Reading the Web: Beyond Named Entity Recognition and Relation Extraction  
  • 2. Disclaimers   •  Previous  versions  of  this  tutorial  were  presented  at   IBERAMIA2012  (h@p://iberamia2012.dsic.upv.es/tutorials/)   and  WWW2013  (h@p://www2013.org/program/machine-­‐ reading-­‐the-­‐web/).  Also,  a  short  version  was  presented  at   ECMLPKDD2015  Summer  School(h@p:// www.ecmlpkdd2015.org/summer-­‐school/ss-­‐schedule).   •  Feel  free  to  e-­‐mail  me  (estevam.hruschka@gmail.com)  with   quesTons  about  this  tutorial  or  any  feedback/suggesTons/ criTcisms.  Your  feedback  can  help  improving  the  quality  of   these  slides,  thus,  they  are  very  welcome.   •  As  in  many  tutorials’  slides,  these  slides  were  prepared  to  be   presented,  and  la@er  studied.  Thus,  they  are  meant  to  be   more  self-­‐contained  than  slides  from  a  paper  presentaTon.  
  • 3. Disclaimers   •  Due  to  Tme  constraints,  I  do  not  intend  to  cover  all  the   algorithms  and  publicaTons  related  to  YAGO,  KnowItAll,  NELL   and  DBPedia.  What  I  do  intend,  instead,  is  to  give  an  overview   of  all  four  projects  and  what  is  the  main  approach  to  “Read   the  Web”,  used  in  each  project.     •  YAGO,  KnowItAll,  NELL  and  DBPedia  are  not  the  only  research   efforts  focusing  on  “Reading  the  Web”.  They  were  selected,   to  be  presented  in  this  tutorial,  because  they  represent  four   different  and  very  relevant  approaches  to  this  problem,  but  it   does  not  mean  they  are  the  best  (or  the  only  relevant)  ones   at  all.  
  • 4. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – YAGO   – KnowItAll   – NELL   – DBPedia  
  • 5. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – YAGO   – KnowItAll   – NELL   – DBPedia  
  • 6. Picture  taken  from  [Fern,  2008]    
  • 7. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – YAGO   – KnowItAll   – NELL   – DBPedia  
  • 8. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – YAGO   – KnowItAll   – NELL   – DBPedia  
  • 9. Picture  taken  from  [DARPA,  2012]    
  • 10. Picture  taken  from  [DARPA,  2012]    
  • 11. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 12. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 13. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 15. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 16. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 17. The  YAGO-­‐NAGA  Project:   HarvesAng,  Searching,  and  Ranking   Knowledge  from  the  Web    
  • 18. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 19. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 22. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 23. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 25. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 26. Machine  Learning   •  What  is  Machine  Learning?   The  field  of  Machine  Learning  seeks  to  answer   the  quesTon   “How  can  we  build  computer  systems  that   automaTcally  improve  with  experience,  and   what  are  the  fundamental  laws  that  govern  all   learning  processes?”  [Mitchell,  2006]  
  • 27. Machine  Learning   •  What  is  Machine  Learning?   a  machine  learns  with  respect  to  a  parTcular:   -­‐  task  T     -­‐  performance  metric  P   -­‐  type  of  experience  E       if  the  system  reliably  improves  its  performance  P  at   task  T,  following  experience  E.  [Mitchell,  1997]  
  • 28. Machine  Learning   •  Examples  of  Machine  Learning  approaches   for  different  tasks  (T),  performance  metrics   (P)  an  experiences  (E)   -­‐  data  mining   -­‐  autonomous  discovery   -­‐  database  updaTng   -­‐  programming  by  example   -­‐  Pa@ern  recogniTon    
  • 29. Machine  Learning   •  Supervised  Learning;   •  Unsupervised  Learning   •  Semi-­‐Supervised  Learning  
  • 31. Supervised  Learning   (one  simple  anecdotal  approach)   0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2  
  • 32. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 33. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 34. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 35. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 36. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 37. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 38. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 39. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 40. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 41. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 42. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)  
  • 43. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Supervised  Learning   (one  simple  anecdotal  approach)   ?????????   What  model   should  be   chosen?   ?????????  
  • 44. 0   5   10   15   20   25   0   5   10   15   20   25   Unsupervised  Learning   (one  simple  anecdotal  approach)  
  • 45. 0   5   10   15   20   25   0   5   10   15   20   25   Unsupervised  Learning   (one  simple  anecdotal  approach)  
  • 46. 0   5   10   15   20   25   0   5   10   15   20   25   Unsupervised  Learning   (one  simple  anecdotal  approach)  
  • 47. 0   5   10   15   20   25   0   5   10   15   20   25   Unsupervised  Learning   (one  simple  anecdotal  approach)  
  • 48. 0   5   10   15   20   25   0   5   10   15   20   25   ?????????   What  model   should  be   chosen?   ?????????   Unsupervised  Learning   (one  simple  anecdotal  approach)  
  • 49. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 50. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 51. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 52. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 53. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 54. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 55. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 56. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 57. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 58. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 59. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 60. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 61. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)  
  • 62. 0   5   10   15   20   25   0   5   10   15   20   25   Series1   Series2   Unlabeled   Semi-­‐supervised  Learning   (one  simple  anecdotal  approach)   ?????????   What  model   should  be   chosen?   ?????????  
  • 63. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 64. Machine  Reading   •  “The  autonomous  understanding  of   text”  [Etzioni  et  al.,  2007]   •  “One  of  the  most  important  methods  by  which   human  beings  learn  is  by  reading”  [Clark  et  al.,   2007],  thus  why  not  building  machines  capable   of  learning  by  reading?  
  • 65. Machine  Reading   •  “The  problem  of  deciding  what  was  implied  by  a   wri@en  text,  of  reading  between  the  lines  is  the   problem  of  inference.”  [Norvig,  2007]     •  Typically,  Machine  Reading  is  different  from   Natural  Language  Processing  alone  
  • 66. It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                                                                   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
  • 67. Machine  Reading                                                       same                                               This  slide  was  adapted  from  [Hady  et  al.,  2011]     It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill."
  • 68. It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                                               This  slide  was  adapted  from  [Hady  et  al.,  2011]    
  • 69. It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                               uncleOf           owns   hires           headOf   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
  • 70. It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                               uncleOf           owns   hires           headOf   affairWith   affairWith   enemyOf   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
  • 71. Machine  Reading   •  One  important  (ini6al)  approach  to  machine   reading  is  to  extract  facts  from  text  and  store   them  in  a  structured  form.   •  Facts  can  be  seen  as  enTTes  and  their   relaTons   •  Ontology  is  one  of  the  most  common   representaTon  for  the  extracted  facts    
  • 80. Machine  Reading   •  Named  EnTty  ResoluTon/RecogniTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Polysemy  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
  • 81. Machine  Reading   •  Named  EnTty  ResoluTon/RecogniTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Polysemy  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
  • 82. Machine  Reading   •  Named  EnTty  ResoluTon/RecogniTon   –  Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.     –  Free  text   •  Hearst-­‐pa@erns;  clustering  by  verbal  phrases   •  Natural-­‐language  processing   •  Advanced  pa@erns  &  iteraTve  bootstrapping    (“Dual  IteraTve  Pa@ern  RelaTon  ExtracTon”)  
  • 83. Named  EnTty  RecogniTon   •  Named  EnTty  RecogniTon  [Nadeau  &  Sekine,   2007]   – term  “Named  EnTty”  coined  for  the  Sixth  Message   Understanding  Conference  (MUC-­‐6)  (R.  Grishman   &  Sundheim  1996).     – important  sub-­‐tasks  of  IE  called  “Named  EnTty   RecogniTon  and  ClassificaTon  (NERC)”.  
  • 84. •  recognize  informaTon  units  like  names,   including  person,  organizaAon  and  locaAon   names,  and  numeric  expressions  including   Ame,  date,  money  and  percent  expressions.   •  In  Machine  Reading,  many  other  enTTes:   product,  kitchen  item,  sport,  etc.   Named  EnTty  RecogniTon   [Nadeau  &  Sekine,  2007]  
  • 85. •  Named  EnTty  ResoluTon  [Theobald  &  Weikum,   2012]   – Which  individual  enTTes  belong  to  which  classes?   •  instanceOf  (Surajit  Chaudhuri,  computer  scien6sts),   •  instanceOf  (BarbaraLiskov,  computer  scien6sts),   •  instanceOf  (Barbara  Liskov,  female  humans),  …   Named  EnTty  ResoluTon  
  • 86. •  Named  EnTTes  RecogniTon  as  a  machine   learning  task.   – Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   Named  EnTty  RecogniTon  
  • 87. •  Named  EnTty  RecogniTon  as  a  Machine  Learning  task.   –  Supervised  Learning   –  Possible  features  [RaTnov  &  Roth,  2009],  [Khambhatla,   2004],  [Zhou  et.  al.  2005]   •   Words  “around”  and  including  enTTes     •  POS  (Part-­‐Of-­‐Speech)   •  Prefixes  and  suffixes   •  CapitalizaTon   •  Number  of  words   •  Number  of  characters   •  First  word,  last  word   •  gaze@eer  matches       Named  EnTty  RecogniTon  
  • 88. •  Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   Named  EnTty  RecogniTon  
  • 89. •  Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   Kernels   Named  EnTty  RecogniTon  
  • 90. •  Supervised  Learning  using  Kernels   – A  Kernel  defines  similarity  implicitly  in  a  higher   dimensional  space     – Can  be  based  on  Strings,  Word  Sequences,  Parse   Trees,  etc.   •  For  strings  similarity∝  number  of  common  substrings   (or  subsequences)     •  Recommended  reading  on  string  kernels  [Lodhi  et.  al.,   2002]       Named  EnTty  RecogniTon  [Bach  &  Badaskar,  2007]  
  • 91. [Bach  &  Badaskar,  2007]   Named  EnTty  RecogniTon  
  • 92. •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     Set  of   labeled  Pa@ern   Examples   Named  EnTty  RecogniTon  
  • 93. •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon  
  • 94. •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon  
  • 95. Set  of  labeled  Instances     •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon  
  • 96. Set  of  labeled  Instances     •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon   Google   Apple  
  • 97.   Set  of  labeled  Instances       •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon   Google   Apple   NE   Pa@ern   Classifier  
  • 98.   Set  of  labeled  Instances       •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in    is  the  CEO  of  X     Named  EnTty  RecogniTon   Google   Apple   NE   Pa@ern   Classifier   What  about   unsupervised?  
  • 99.   Set  of  labeled  Instances       •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   Named  EnTty  RecogniTon   NE   Pa@ern   Classifier   What  about   unsupervised?  
  • 100.   Set  of  labeled  Instances       •  Unsupervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  NE  instances.     NE   Instances   Classifier   Set  of   labeled  Pa@ern   Examples   Named  EnTty  RecogniTon   NE   Pa@ern   Classifier  
  • 101. •  [RaTnov  &  Roth,  2009]   Named  EnTty  RecogniTon  
  • 102. •  [Pennington  &  Socher  &  Manning,  2014]   Named  EnTty  RecogniTon  
  • 103. Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Polysemy  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
  • 104. Machine  Reading   •  RelaTon  ExtracTon   –  Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.     –  Free  text   •  Hearst-­‐pa@erns;  clustering  by  verbal  phrases   •  Natural-­‐language  processing   •  Advanced  pa@erns  &  iteraTve  bootstrapping    (“Dual  IteraTve  Pa@ern  RelaTon  ExtracTon”)  
  • 105. Machine  Reading   •  RelaTon  ExtracTon  [Theobald  &  Weikum,  2012]   –  Which  instances  (pairs  of  individual  enTTes)  are  there   for  given  binary  relaTons  with  specific  type   signatures?   •  hasAdvisor  (JimGray,  MikeHarrison)   •  hasAdvisor  (HectorGarcia-­‐Molina,  Gio  Wiederhold)   •  hasAdvisor  (Susan  Davidson,  Hector  Garcia-­‐Molina)   •  graduatedAt  (JimGray,  Berkeley)   •  graduatedAt  (HectorGarcia-­‐Molina,  Stanford)   •  hasWonPrize  (JimGray,  TuringAward)   •  bornOn  (JohnLennon,  9Oct1940)   •  diedOn  (JohnLennon,  8Dec1980)   •  marriedTo  (JohnLennon,  YokoOno)  
  • 106. RelaTon  ExtracTon   •  ExtracTng  semanTc  relaTons  between  enTTes   in  text   •  RelaTon  extracTon  as  a  Machine  Learning  task.   – Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   [Bach  &  Badaskar,  2007]  
  • 107. RelaTon  ExtracTon   •  RelaTon  extracTon  as  a  Machine  Learning  task.   –  Supervised  Learning   –  Possible  features  [Khambhatla,  2004],  [Zhou  et.  al.   2005]   •   Words  between  and  including  enTTes     •  Types  of  enTTes  (person,  locaTon,  etc)     •  Number  of  enTTes  between  the  two  enTTes,  whether  both   enTTes  belong  to  same  chunk     •  #  words  separaTng  the  two  enTTes     •  Path  between  the  two  enTTes  in  a  parse  tree       [Bach  &  Badaskar,  2007]  
  • 108. RelaTon  ExtracTon   •  ExtracTng  semanTc  relaTons  between  enTTes   in  text   •  RelaTon  extracTon  as  a  classificaTon  task.   – Supervised  Learning     NLP  tools   (POS,  Parse   Trees,  NER)   text   Features   ExtracTon   Classifier   [Bach  &  Badaskar,  2007]  
  • 109. RelaTon  ExtracTon   •  ExtracTng  semanTc  relaTons  between  enTTes   in  text   •  RelaTon  extracTon  as  a  classificaTon  task.   – Supervised  Learning     NLP  tools   (POS,  Parse   Trees,  NER)   text   Features   ExtracTon   Classifier   Kernels   [Bach  &  Badaskar,  2007]  
  • 110. RelaTon  ExtracTon   •  Supervised  Learning  using  Kernels   – A  Kernel  defines  similarity  implicitly  in  a  higher   dimensional  space     – Can  be  based  on  Strings,  Word  Sequences,  Parse   Trees,  etc.   •  For  strings,  similarity∝  number  of  common  substrings   (or  subsequences)     •  Recommended  reading  on  string  kernels  [Lodhi  et.  al.,   2002]       [Bach  &  Badaskar,  2007]  
  • 111. RelaTon  ExtracTon  [Bach  &  Badaskar,  2007]  
  • 112. RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples  
  • 113. RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X    
  • 114. RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Pair  of   Instances   Classifier  
  • 115. RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Pair  of   Instances   Classifier  
  • 116. RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Pair  of   Instances   Classifier     Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Google-­‐Mountain  View   Apple-­‐CuperAno  
  • 117. RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Google-­‐Mountain  View   Apple-­‐CuperAno   Pair  of   Instances   Classifier  
  • 118. RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   X  is  headquartered  in  Y   Y  is  the  headquarter  of  X     Google-­‐Mountain  View   Apple-­‐CuperAno   Pair  of   Instances   Classifier   What  about   unsupervised?  
  • 119. RelaTon  ExtracTon   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   What  about   unsupervised?  
  • 120. RelaTon  ExtracTon   •  Unsupervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.       Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier  
  • 121. •  Supervised  learning  [Bunescu  &  Mooney,  2005]   •  Distant  and  ParTal  Supervised  [Angeli  &   Tibshirani  &  Wu  &  Manning,  2014]   RelaTon  ExtracTon  
  • 122. Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Polysemy  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
  • 123. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  taken  from:  h@p://nlp.stanford.edu/projects/coref.shtml    
  • 124. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  taken  from:  h@p://nlp.stanford.edu/projects/coref.shtml     within-document co-reference
  • 125. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  taken  from:  h@p://nlp.stanford.edu/projects/coref.shtml     within-document co-reference
  • 126. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   computer     Apple   Computer    
  • 127. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   apple   computer     Apple   Computer    
  • 128. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the   same  enTty   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   apple   computer     Apple   Computer     cross-document co-reference
  • 129. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference:  expressions  that  refer  to  the  same   enTty   •  Which  names  denote  which  enTTes?  [Theobald   &  Weikum,  2012]   –  means  (“Lady  Di“,  Diana  Spencer),   –  means  (“Diana  Frances  Mountba@en-­‐Windsor”,  Diana   Spencer),  …   –  means  (“Madonna“,  Madonna  Louise  Ciccone),   –  means  (“Madonna“,  Madonna(painTng  by  Edward   Munch)),  …   cross-document co-reference
  • 130. Co-­‐Reference  and  Polysemy  ResoluTon   •  Polysemy:  is  the  capacity  for  a  sign  (such  as  a   word,  phrase,  or  symbol)  to  have  mulTple   meanings  [Wikipedia]  
  • 131. Co-­‐Reference  and  Polysemy  ResoluTon   •  Polysemy:  is  the  capacity  for  a  sign  (such  as  a   word,  phrase,  or  symbol)  to  have  mulTple   meanings  [Wikipedia]   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   apple     (the  fruit)   Apple   Computer    
  • 132. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐Reference  and  Polysemy   Example  (figure)  adapted  from  [Krishnamurthy  &  Mitchell,  2011]   apple   apple   computer     apple     (the  fruit)   Apple   Computer    
  • 133. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference  and  Polysemy:     – Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier  
  • 134. •  Co-­‐Reference   ResoluTon.   – Supervised   Learning   – Possible   features   [Bengtson  &   Roth,  2008]     Co-­‐Reference  and  Polysemy  ResoluTon  
  • 135. •  Co-­‐Reference   ResoluTon.   – Supervised   Learning   – Possible   features   [Bengtson  &   Roth,  2008]     Co-­‐Reference  and  Polysemy  ResoluTon  
  • 136. Co-­‐Reference  and  Polysemy  ResoluTon   •  Co-­‐reference  and  Polysemy:     – Supervised  Learning     NLP  tools   (POS,  Parse   Trees)   text   Features   ExtracTon   Classifier   Kernels  
  • 137. •  Supervised  Learning  using  Kernels   – A  Kernel  defines  similarity  implicitly  in  a  higher   dimensional  space     – Can  be  based  on  Strings,  Word  Sequences,  Parse   Trees,  etc.   •  For  strings  similarity∝  number  of  common  substrings   (or  subsequences)     •  Recommended  reading  on  string  kernels  [Lodhi  et.  al.,   2002]       Co-­‐Reference  and  Polysemy  ResoluTon  
  • 138. •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   Co-­‐Reference  and  Polysemy  ResoluTon  
  • 139. •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   X  also  know  as  Y     Co-­‐Reference  and  Polysemy  ResoluTon  
  • 140. •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Set  of   labeled  Pa@ern   Examples   Pair  of   Instances   Classifier   Co-­‐Reference  and  Polysemy  ResoluTon   X  also  know  as  Y    
  • 141.   Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   X  also  know  as  Y    
  • 142. Pair  of   Instances   Classifier     Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Apple  Computer  -­‐   Apple   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   X  also  know  as  Y    
  • 143.   Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   Apple  Computer  -­‐   Apple   X  also  know  as  Y    
  • 144.   Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   Apple  Computer  -­‐   Apple   X  also  know  as  Y     What  about   unsupervised?  
  • 145.   Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon   What  about   unsupervised?  
  • 146.   Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   Pa@ern   Classifier   Pair  of   Instances   Classifier   •  Semi-­‐supervised   Approaches     – Bootstrap  can   generate  a  large   number  of  pa@erns   and  relaTon   instances.     Co-­‐Reference  and  Polysemy  ResoluTon  
  • 147. •  Co-­‐Reference  ResoluTon:     [Singh  et  al.,  2011],  [Krishnamurthy  &  Mitchell,   2011],[Du@a  &  Weikum,  2015]   •  Polysemy  ResoluTon:   [Krishnamurthy  &  Mitchell,  2011],  [Galárraga  et   al.,    2014]   Co-­‐Reference  and  Polysemy  ResoluTon  
  • 148. Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Synonym  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
  • 149. Machine  Reading   •  RelaTon  Discovery   – Which  new  relaTons  are  there  for  given  pair  of   enTTes?   •  hasAdvisor  (JimGray,  MikeHarrison)  
  • 150. Machine  Reading   •  RelaTon  Discovery   – Which  new  relaTons  are  there  for  given  pair  of   enTTes?   •  hasAdvisor  (JimGray,  MikeHarrison)   •  hasCoAuthor(HectorGarcia-­‐Molina,  Gio  Wiederhold)  
  • 151. Machine  Reading   •  RelaTon  Discovery   – Which  new  relaTons  are  there  for  given  pair  of   enTTes?   •  hasAdvisor  (JimGray,  MikeHarrison)   •  hasCoAuthor(HectorGarcia-­‐Molina,  Gio  Wiederhold)   •  graduatedAt  (JimGray,  Berkeley)  
  • 152. Machine  Reading   •  RelaTon  Discovery   – Which  new  relaTons  are  there  for  given  pair  of   enTTes?   •  hasAdvisor  (JimGray,  MikeHarrison)   •  hasCoAuthor(HectorGarcia-­‐Molina,  Gio  Wiederhold)   •  graduatedAt  (JimGray,  Berkeley)   •  studiedAt  (HectorGarcia-­‐Molina,  Stanford)   •  bornOn  (JohnLennon,  9Oct1940)   •  releasedAlbum  (JohnLennon,  10Dec1965)  
  • 153.   Set  of  labeled  pairs  of   Instances  Examples     Set  of   labeled  Pa@ern   Examples   RelaTon  Discovery   Clustering   Algorithm  
  • 154. Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Synonym  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
  • 155. Inference   •  Inference  is  the  act  or  process  of  deriving   logical  conclusions  from  premises  known  or   assumed  to  be  true  [Wikipedia]  
  • 156. Inference   •  Manually  craved  inference  rules   •  AutomaTcally  learned  inference  rules   •  Data  mining  the  Knowledge  Base  
  • 157. Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Synonym  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
  • 158. Machine  Reading   •  Ontology  RepresentaTon    Facts  (RDF  triples)   1:  (Jim,  hasAdvisor,  Mike)   2:  (Surajit,  hasAdvisor,  Jeff)   3:  (Madonna,  marriedTo,  GuyRitchie)   4:  (Nicolas,  marriedTo,  Carla)   5:  (ManchesterU,  wonCup,  ChampionsLeague)   ReificaTon:   “Facts  about  Facts”:   6:      (1,  inYear,  1968)   7:      (2,  inYear,  2006)   8:      (3,  validFrom,  22-­‐Dec-­‐2000)     9:      (3,  validUnTl,  Nov-­‐2008)   10:  (4,  validFrom,  2-­‐Feb-­‐2008)   11:  (2,  source,  SigmodRecord)   12:  (5,  inYear,  1999)   13:  (5,  locaTon,  CampNou)   14:  (5,  source,  Wikipedia)  
  • 159. Machine  Reading   •  Named  EnTty  ResoluTon/ExtracTon   •  RelaTon  ExtracTon   •  Co-­‐reference  and  Synonym  ResoluTon   •  RelaTon  Discovery   •  Inference   •  Knowledge  Base  RepresentaTon   •  Document/Sentence  Understanding  (Micro-­‐ Reading)  
  • 160. Document/Sentence  UnderstanTng     (MicroRead)   •  “The  scienTst  observed  the  bu@erfly  with  the   blue  circle”      
  • 161. Document/Sentence  UnderstanTng     (MicroRead)   •  “The  scienTst  observed  the  bu[erfly  with  the   blue  circle”      
  • 162. Document/Sentence  UnderstanTng     (MicroRead)   •  “The  scienTst  observed  the  bu[erfly  with  the   blue  circle”       •  “The  scienTst  observed  the  bu@erfly  with  the   blue  microscope”  
  • 163. Document/Sentence  UnderstanTng     (MicroRead)   •  “The  scienTst  observed  the  bu[erfly  with  the   blue  circle”       •  “The  scienAst  observed  the  bu@erfly  with  the   blue  microscope”  
  • 164. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 165. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 167. DBPedia   Mapping  Wikipedia  semi-­‐structured  data  into  RDF  triples  
  • 168. DBPedia   Mapping  Wikipedia  semi-­‐structured  data  into  RDF  triples   Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”  
  • 169. DBPedia   •  How  to  Read  Wikipedia  Semi-­‐structured  data?   [Lehmann  et  al.,  2014]   – Parse  Wikipedia  Markup  language   – Overcome  the  lack  of  standard  problem   •  Same  properTes  might  have  different  names   •  “Datebirth”  and  “Birth_date”   •  “Birthplace”  and  “Birth_place”   – Instead  of  “Modeling  the  World”,  try  to  structure   the  available  informaTon  
  • 171. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 172. The  YAGO-­‐NAGA  Project:   HarvesAng,  Searching,  and  Ranking  Knowledge   from  the  Web    
  • 173. The  YAGO-­‐NAGA  Project:   HarvesAng,  Searching,  and  Ranking   Knowledge  from  the  Web    
  • 174. YAGO   •  Yet  Another  Great  Ontology  -­‐  YAGO   •  Main  Goal:  building  a  conveniently  searchable,   large-­‐scale,  highly  accurate  knowledge  base  of   common  facts  in  a  machine-­‐processable   representaTon  
  • 176. YAGO   •  Turn  Web  into  Knowledge  Base  [Weikum  et   al.,  2009]   – Building  a  comprehensive  Knowledge  Base  of   human  knowledge   – knowledge  from  Wikipedia  and  WordNet   – the  ontology  check  itself  for  precision    
  • 177. YAGO   •  The  knowledge  base  is  automaTcally   constructed  from  Wikipedia   •  Each  arTcle  in  Wikipedia  becomes  an  enTty  in   the  kb  (e.g.,  since  Leonard  Cohen  has  an   arTcle  in  Wikipedia,  LeonardCohen  becomes   an  enTty  in  YAGO).    
  • 181. YAGO   Free   Text   InfoBox  
  • 183. YAGO   Wikipedia  InfoBox   Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”  
  • 184. YAGO   Wikipedia  InfoBox   Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”  
  • 185. YAGO   •  Certain  categories  are  exploited  to  deliver   type  informaTon  (e.g.,  the  arTcle  about   Leonard  Cohen  is  in  the  category  Canadian   male  poets,  so  he  becomes  a  Canadian  poet).    
  • 188. YAGO   •  For  each  category  of  a  page  [Hoffart  et  al.,  2012]   –  Using  shallow  parsing,  determine  the  head  word  of  the   category  name.  In  the  example  of  Canadian  poets,  the   head  word  is  poets.     –  If  the  head  word  is  in  plural,  then  proposes  the  category  as   a  class  and  the  arTcle  enTty  as  an  instance     –  Link  the  class  to  the  WordNet  taxonomy  (most  frequent   sense  of  the  head  word  in  WordNet)   •  only  countable  nouns  can  appear  in  plural  form   •  only  countable  nouns  can  be  ontological  classes   •  themaTc  categories  (such  as  Canadian  poetry)  are   different  from  conceptual  Categories  
  • 189. YAGO   •  head  words  that  are  not  conceptual  even  though   they  appear  in  plural  (such  as  stubs  in  Canadian   poetry  stubs)  are  in  the  first  list  of  excepTons.     •  words  that  do  not  map  to  their  most  frequent   sense,  but  to  a  different  sense  are  in  the  second   excepTon  list   –  The  word  capital,  e.g.,  refers  to  the  main  city  of  a   country  in  the  majority  of  cases  and  not  to  the   financial  amount,  which  is  the  most  frequent  sense  in   WordNet.  
  • 190. YAGO   •  About  100  manually  defined  relaTons   –  wasBornOnDate     –  locatedIn     –  hasPopulaTon     •  Categories  and  infoboxes  are  exploited  to  deliver  facts   (instances  of  relaTons).     •  Manually  defined  pa@erns  that  map  categories  and   infobox  a@ributes  to  fact  templates   –  infobox  a@ribute  born=Montreal,  thus   wasBornIn(LeonardCohen,  Montreal)     •  Pa@ern-­‐based  extracTons  resulted  in  2  million   extracted  enTTes  and  20  million  facts  
  • 191. YAGO   •  Based  on  declaraTve  rules  (stored  in  text  files)   •  The  rules  take  the  form  of  subject-­‐  predicate-­‐ object  triples,  so  that  they  are  basically   addiTonal  facts   •  There  are  different  types  of  rules  
  • 192. YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
  • 193. YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
  • 194. YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.   Knowledge   RepresentaTon  
  • 195. YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
  • 196. YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.     Inference    
  • 197. YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
  • 198. YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  Knowledge   RepresentaTon  
  • 199. YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.  
  • 200. YAGO   •  Factual  rules:    definiTon  of  all  relaTons,  their  domains  and   ranges,  and  the  definiTon  of  the  classes  that  make  up  the   YAGO  hierarchy  of  literal  types.   •  ImplicaAon  rules:  express  that  if  certain  facts  appear  in  the   knowledge  base,  then  another  fact  shall  be  added.  Horn   clause  rules.   •  Replacement  rules:  for  interpreTng  micro-­‐formats,   cleaning  up  HTML  tags,  and  normalizing  numbers.   •  ExtracAon  rules:  apply  primarily  to  pa@erns  found  in  the   Wikipedia  infoboxes,  but  also  to  Wikipedia  categories,   arTcle  Ttles,  and  even  other  regular  elements  in  the  source   such  as  headings,  links,  or  references.   InformaTon   ExtracTon  
  • 201. YAGO   •  AutomaTcally  verifies  consistency   – Check  uniqueness  of  funcTonal  arguments   •  spouse(x,y)  ∧  diff(y,z)  ⇒  ¬spouse(x,z)   – Check  domains  and  ranges  of  relaTons   •  spouse(x,y)  ⇒  female(x)   •  spouse(x,y)  ⇒  male(y)   •  spouse(x,y)  ⇒  (f(x)∧m(y))  ∨  (m(x)∧f(y))      
  • 202. YAGO   •  AutomaTcally  verifies  consistency   – Hard  Constraint   •  hasAdvisor(x,y)  ∧  graduatedInYear(x,t)  ∧  graduatedInYear(y,s)  ⇒  s  <  t   – Sov  Constraint     •  firstPaper(x,p)  ∧  firstPaper(y,q)  ∧  author(p,x)  ∧  author(p,y)  )  ∧      inYear(p)  >  inYear(q)  +  5years  ⇒  hasAdvisor(x,y)  [0.6]    
  • 203. YAGO   •  AutomaTcally  verifies  consistency   – Hard  Constraint   •  hasAdvisor(x,y)  ∧  graduatedInYear(x,t)  ∧  graduatedInYear(y,s)  ⇒  s  <  t   – Sov  Constraint     •  firstPaper(x,p)  ∧  firstPaper(y,q)  ∧  author(p,x)  ∧  author(p,y)  )  ∧      inYear(p)  >  inYear(q)  +  5years  ⇒  hasAdvisor(x,y)  [0.6]     Inference  
  • 204. YAGO   •  Ontology  RepresentaTon   – EnTTes  and  RelaTons  of  public  interest   – Format:  TSV,  RDF,  XML,  N3,  Web  Interface   – Learns   •  Instances  and  pa@erns  from  Wikipedia;   •  Taxonomy  from  WordNet;   •  Geotags  informaTon  from  Geonames.  
  • 205. YAGO   •  Named  EnTty  ResoluTon/ExtracTon  [Theobald  &   Weikum,  2012]   – Based  on  rules  and  pa@erns  extracted  from   Wikipedia   – DisambiguaTon  is  a  relevant  issue   – Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.  
  • 206. YAGO   •  Named  EnTty  ResoluTon/ExtracTon  [Theobald  &   Weikum,  2012]   – Based  on  rules  and  pa@erns  extracted  from   Wikipedia   – DisambiguaTon  is  a  relevant  issue   – Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.   Natural  Language   Processing   Machine   Learning  
  • 207. It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                                                       This  slide  was  adapted  from  [Hady  et  al.,  2011]    
  • 208. It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                                                                   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
  • 209. YAGO   •  RelaTon  ExtracTon  [Theobald  &  Weikum,  2012]   – Based  on  rules  and  pa@erns  extracted  from   Wikipedia   – Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.  
  • 210. YAGO   •  RelaTon  ExtracTon  [Theobald  &  Weikum,  2012]   – Based  on  rules  and  pa@erns  extracted  from   Wikipedia   – Semi-­‐structured  data   The  “Low-­‐Hanging  Fruit”   •  Wikipedia  infoboxes  &  categories   •  HMTL  lists  &  tables,  etc.   Natural  Language   Processing   Machine   Learning  
  • 211. It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                                                                   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
  • 212. Machine  Reading                                                       same                                               This  slide  was  adapted  from  [Hady  et  al.,  2011]     It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill."
  • 213. It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                                               This  slide  was  adapted  from  [Hady  et  al.,  2011]    
  • 214. It’s about the disappearance forty years ago of Harriet Vanger, a young scion of one of the wealthiest families in Sweden, and about her uncle, determined to know the truth about what he believes was her murder. Blomkvist visits Henrik Vanger at his estate on the tiny island of Hedeby. The old man draws Blomkvist in by promising solid evidence against Wennerström. Blomkvist agrees to spend a year writing the Vanger family history as a cover for the real assignment: the disappearance of Vanger's niece Harriet some 40 years earlier. Hedeby is home to several generations of Vangers, all part owners in Vanger Enterprises. Blomkvist becomes acquainted with the members of the extended Vanger family, most of whom resent his presence. He does, however, start a short lived affair with Cecilia, the niece of Henrik. After discovering that Salander has hacked into his computer, he persuades her to assist him with research. They eventually become lovers, but Blomkvist has trouble getting close to Lisbeth who treats virtually everyone she meets with hostility. Ultimately the two discover that Harriet's brother Martin, CEO of Vanger Industries, is secretly a serial killer. A 24-year-old computer hacker sporting an assortment of tattoos and body piercings supports herself by doing deep background investigations for Dragan Armansky, who, in turn, worries that Lisbeth Salander is “the perfect victim for anyone who wished her ill." Machine  Reading                                                       same   same   same   same   same   same                               uncleOf           owns   hires           headOf   This  slide  was  adapted  from  [Hady  et  al.,  2011]    
  • 215. Machine  Reading   This  slide  was  taken  from  [Hoffart  et  al.,  2015]    
  • 216. YAGO   •  YAGO2:  Exploring  and  Querying  World   Knowledge  in  Time,  Space,  Context,  and  Many   Languages   – New  relaTons  specifically  designed  to  cover  Tme,   space  and  context   – Wikipedia  translated  pages  as  sources  for  other   languages  
  • 217. YAGO   •  YAGO3  [Mahdisoltani  &  Biega  &  Suchanek,  2015]   –  an  extension  of  the  YAGO  knowledge  base;   –  built  from  the  Wikipedias  in  mulTple  languages.     –  fuses  the  mulTlingual  informaTon  with  the  English  WordNet   –  categories,  infoboxes,  and  Wikidata,  to  learn  the  meaning  of   infobox  a@ributes  across  languages   –  10  different  languages   –  precision  of  95%-­‐100%  in  the  a@ribute  mapping   –  enlarges  YAGO  by  1m  new  enTTes  and  7m  new  facts.    
  • 218. YAGO   •  More  on  YAGO:   –  Very  nice  tutorials:   •  “Knowledge  Bases  for  Web  Content  AnalyTcs”  at  WWW   2015,  Florence,  May  2015.   •  "SemanTc  Knowledge  Bases  from  Web  Sources"  at  IJCAI   2011,  Barcelona,  July  2011   "HarvesTng  Knowledge  from  Web  Data  and  Text"  at  CIKM   2010,  Toronto,  October  2010   "From  InformaTon  to  Knowledge:  HarvesTng  EnTTes  and   RelaTonships  from  Web  Sources"  at  PODS  2010,   Indianapolis,  June  2010   –  Project  Website:   •  h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/  
  • 219. YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)  
  • 220. YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)  
  • 221. YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)  
  • 222. YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)   ?X  <hasChild>  ?C  ?Y  <hasChild>  ?C  =>  ?X  <isMarriedTo>  ?Y  
  • 223. YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)   ?X  <hasChild>  ?C  ?Y  <hasChild>  ?C  =>  ?X  <isMarriedTo>  ?Y   Machine   Learning  
  • 224. YAGO   •  More  on  YAGO  (h[p://www.mpi-­‐inf.mpg.de/yago-­‐naga/)   ?X  <hasChild>  ?C  ?Y  <hasChild>  ?C  =>  ?X  <isMarriedTo>  ?Y   Machine   Learning   Inference  
  • 225. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 226. Outline   •  Machine  Learning   •  Machine  Reading   •  Reading  the  Web   – DBPedia   – YAGO   – KnowItAll   – NELL  
  • 230. KnowItAll   •  MoTvaTon:  New  Paradigm  for  Search  [Etzioni,  2008]   –  The  future  of  Web  Search   –  Read  the  Web  instead  of  retrieving  Web  pages  to   perform  Web  Search  
  • 231. KnowItAll   •  InformaTon  ExtracTon  (IE)  +  tractable   inference     –  IE(sentence)  =  who  did  what?   •  speaker(P.  Smith,  ECMLPKDD2012)   –  Inference  =  uncover  implicit  informaTon   •  Will  Pi@sburgh  Steelers  be  champions  again?     •  Open  InformaTon  ExtracTon  [Banko  et  al.,  2007]  
  • 232. Open  InformaTon  ExtracTon     [Banko  et  al.,  2007]   •  Open  IE  systems  avoid  specific  nouns  and   verbs     •  Extractors  are  unlexicalized—formulated  only   in  terms  of:   –   syntacTc  tokens  (e.g.,  part-­‐of-­‐speech  tags)     –  closed-­‐word  classes  (e.g.,  of,  in,  such  as).     •  Open  IE  extractors  focus  on  generic  ways  in   which  relaTonships  are  expressed  in  English   –  naturally  generalizing  across  domains.  
  • 233. Open  InformaTon  ExtracTon     [Banko  et  al.,  2007]   •  Open  IE  extractors  focus  on  generic  ways  in   which  relaTonships  are  expressed  in  English   –  naturally  generalizing  across  domains.   RelaTon   Discovery  
  • 234. Open  InformaTon  ExtracTon     •  Open  IE  systems  are  tradiTonally  based  on     three  steps  [Etzioni  et  al.,  2011]:   –  1.  Label:  Sentences  are  automaTcally  labeled  with   extracTons  using  heurisTcs  or  distant  supervision.   Unsupervised   Learning  
  • 235. Open  InformaTon  ExtracTon     •  Open  IE  systems  are  tradiTonally  based  on     three  steps  [Etzioni  et  al.,  2011]:   –  1.  Label:  Sentences  are  automaTcally  labeled  with   extracTons  using  heurisTcs  or  distant  supervision.   –  2.  Learn:  A  relaTon  phrase  extractor  is  learned  using  a   sequence-­‐labeling  graphical  model  (e.g.,  CRF).   Supervised   Learning