SlideShare a Scribd company logo
1 of 1
Download to read offline
Poster Design & Printing by Genigraphics®
- 800.790.4001
Ted Pedersen
Department of Computer Science
University of Minnesota, Duluth
tpederse@d.umn.edu
http://www.d.umn.edu/~tpederse
Ted Pedersen
University of Minnesota, Duluth
http://senseclusters.sourceforge.net
Systems use different corpora to build co-
occurrence matrix, otherwise nearly identical.
Sys7 : Uses smallest corpora, just the 64
snippets per term. Resulting matrices range
from 102 X 113 to 221 X 222. No SVD.
Sys1: Use all 6400 snippets to create matrix
of size 771 X 952. SVD reduced to 771 X 90.
Sys9: Use first 10,000 paragraphs of APW data
from English Gigaword. Resulting matrix is
9,853 X 10,995. SVD reduced to 9,853 X 300.
RandX: a random baseline, assign each Web
snippet to one of X random senses. Evaluation
measures should be able to expose random
baselines and give them appropriately low
scores. Both the paired F-score used in
SemEval-2010 and the Jaccard Coefficient
satisfy this requirement.
Given 64 snippets per ambiguous term,
first order features were unlikely to
succeed and achieved F-10 score of 36.10
According to F-10, F-SC, and Jaccard,
smaller amounts of task specific data
(sys7 and sys1) are more effective than
large amounts of out of domain text (sys9)
when using second order methods.
Newspaper text (like APW as used in Sys9)
is not typically what Web search locates.
Results are more often commercial,
Wikipedia, or current celebrity.
Lessons learned? :
●
For second order methods, use Web-like
data, not news text
●
Use more data, increase snippets with
Web, Wikipedia, etc and then discard
additions after clustering (?)
●
Expand snippets by going to site in result
•
Task 11 goal? Cluster Web search results!
Test Data? The top 64 Google results for
each of 100 potentially ambiguous queries.
Each result is a Web snippet about 25
words long.
Challenges? Small amounts of data!
• 64 results/term X 25 words/result =
1,600 words/term X 100 terms =
160,000 words
Solution? Augment the data!
• Use second order co-occurrences to
enrich Web snippets to be clustered, friend
of a friend relation
• The bigrams car motor, car insurance,
car magazine, life sentence, life force, and
life insurance each represent a first order
co-occurrence. Car and life are second
order co-occurrences because both occur
with insurance.
Introduction
Generic Duluth System
Conclusion and Future Directions
DiscussionThe Duluth SystemsAbstract
Contact
Duluth : Word Sense Induction Applied to Web Page Clustering
Task 11 : Word Sense Induction & Disambiguation within an End-User Application
Experimental Results
System F-10
2010
Jaccard F1-13
2014
ARI Clusters
/Size
Sys1 46.53 31.79 56.83 5.75
2.5/
26.5
Sys7 45.89 31.03 58.78 6.78
3.0/
25.2
Sys9 35.56 22.24 57.02 2.59
3.3/
19.8
Rand2 41.49 26.99 54.89 –0.04
2.0/
32.0
Rand5 25.17 14.52 56.73 0.12
5.0/
12.8
Rand10 15.05 8.18 59.67 0.02
10.0/
6.4
Rand25 7.01 3.63 66.89? -0.15
23.2/
2.8
Rand50 4.07 2.00 76.19? 0.10
35.9/
1.8
MFS 54.06 39.90 54.42 0.00?
1.0/
64.0
Gold 100.00 100.00 100.00 99.00
7.7/
11.6
The Duluth systems that
participated in Task 11 of
SemEval–2013 carried out
word sense induction (WSI)
in order to cluster Web
search results. They relied
on an approach that
represented Web snippets
using second-order co-
occurrences. These
systems were all
implemented using
SenseClusters, a freely
available open source
software package.
Web page clustering viewed as a word
sense discrimination / induction problem
where query term is the target word,
snippet provides surrounding context
•Create co-occurrence matrix from some
corpus. Rows and columns made up of first
and second word of bigrams identified by
log-likelihood ratio. Low frequency and low
scoring bigrams are excluded, as are stop
words. Optionally reduce dimensionality
with SVD.
•Replace each word in a web snippet with
a vector made up of its row from the co-
occurrence matrix. Average all vectors for
a snippet together to create a new context
vector. This will capture second order co-
occurrences between context vectors.
•Cluster the resulting 64 vectors for an
ambiguous term, find number of clusters
automatically with PK2.
Average number of senses is 7.7, but
variance is quite high ...
•heron island – 1 sense, 100% MFS
•Shakira – 2 senses, 98% MFS
•apple – 2 senses, 98% MFS
•kawasaki – 7 senses, 47% MFS
•Billy the Kid – 7 senses, 44% MFS
•marble – 8 senses, 39% MFS
•kangaroo – 17 senses, 48% MFS
•ghost – 18 senses, 30% MFS
•dog eat dog – 19 senses, 28% MFS
•magic – 19 senses, 27% MFS
Of 769 senses in the test data ...
•467 (61%) occur less than 5 times!!
• Is 1, 2, 3, 4, … instances enough to
identify a cluster?
• Very small clusters often “pure”, can
trick some evaluation methods
•186 (24%) are defined as “Other”
• Criteria for membership unclear or
different than other senses
• Other different than can't cluster?

More Related Content

Viewers also liked

Information Privilege: Narratives of Challenge and Change
Information Privilege: Narratives of Challenge and ChangeInformation Privilege: Narratives of Challenge and Change
Information Privilege: Narratives of Challenge and Changechar booth
 
Duluth : Word Sense Discrimination in the Service of Lexicography
Duluth : Word Sense Discrimination in the Service of LexicographyDuluth : Word Sense Discrimination in the Service of Lexicography
Duluth : Word Sense Discrimination in the Service of LexicographyUniversity of Minnesota, Duluth
 
Cultivating Campus Collaborations
Cultivating Campus CollaborationsCultivating Campus Collaborations
Cultivating Campus Collaborationschar booth
 
Strategic Cartography: Identifying IL Intersections Across the Curriculum
Strategic Cartography: Identifying IL Intersections Across the CurriculumStrategic Cartography: Identifying IL Intersections Across the Curriculum
Strategic Cartography: Identifying IL Intersections Across the Curriculumchar booth
 
Love Your Library CCL Button Templates - 2.25'' Multiple Color Pages with Hearts
Love Your Library CCL Button Templates - 2.25'' Multiple Color Pages with HeartsLove Your Library CCL Button Templates - 2.25'' Multiple Color Pages with Hearts
Love Your Library CCL Button Templates - 2.25'' Multiple Color Pages with Heartschar booth
 
Why Reflect? The Holistic Practice of Stepping Back.
Why Reflect? The Holistic Practice of Stepping Back. Why Reflect? The Holistic Practice of Stepping Back.
Why Reflect? The Holistic Practice of Stepping Back. char booth
 
Training the Trainers: Faculty Development Meets Information Literacy
Training the Trainers: Faculty Development Meets Information LiteracyTraining the Trainers: Faculty Development Meets Information Literacy
Training the Trainers: Faculty Development Meets Information Literacysusangar
 
Information literacy through inquiry: using problem-based learning in informa...
Information literacy through inquiry: using problem-based learning in informa...Information literacy through inquiry: using problem-based learning in informa...
Information literacy through inquiry: using problem-based learning in informa...Alan Carbery
 
The horizon isn't found in a dictionary : Identifying emerging word senses a...
The horizon isn't found in a  dictionary : Identifying emerging word senses a...The horizon isn't found in a  dictionary : Identifying emerging word senses a...
The horizon isn't found in a dictionary : Identifying emerging word senses a...University of Minnesota, Duluth
 
PDU 211 Research Methods: Identifying a Research Problem
PDU 211 Research Methods: Identifying a Research ProblemPDU 211 Research Methods: Identifying a Research Problem
PDU 211 Research Methods: Identifying a Research ProblemAgatha N. Ardhiati
 
information privilege: access, advocacy, and the critical role of libraries.
information privilege: access, advocacy, and the critical role of libraries.information privilege: access, advocacy, and the critical role of libraries.
information privilege: access, advocacy, and the critical role of libraries.char booth
 
Activate Your Learners! Active Learning Strategies for Fostering Participant ...
Activate Your Learners! Active Learning Strategies for Fostering Participant ...Activate Your Learners! Active Learning Strategies for Fostering Participant ...
Activate Your Learners! Active Learning Strategies for Fostering Participant ...Lisa S.
 

Viewers also liked (20)

Information Privilege: Narratives of Challenge and Change
Information Privilege: Narratives of Challenge and ChangeInformation Privilege: Narratives of Challenge and Change
Information Privilege: Narratives of Challenge and Change
 
Eacl 2006 Pedersen
Eacl 2006 PedersenEacl 2006 Pedersen
Eacl 2006 Pedersen
 
Pedersen ACL Disco-2011 workshop
Pedersen ACL Disco-2011 workshopPedersen ACL Disco-2011 workshop
Pedersen ACL Disco-2011 workshop
 
Cicling2005
Cicling2005Cicling2005
Cicling2005
 
Duluth : Word Sense Discrimination in the Service of Lexicography
Duluth : Word Sense Discrimination in the Service of LexicographyDuluth : Word Sense Discrimination in the Service of Lexicography
Duluth : Word Sense Discrimination in the Service of Lexicography
 
Cultivating Campus Collaborations
Cultivating Campus CollaborationsCultivating Campus Collaborations
Cultivating Campus Collaborations
 
Teaching with Technology
Teaching with TechnologyTeaching with Technology
Teaching with Technology
 
Pedersen masters-thesis-oct-10-2014
Pedersen masters-thesis-oct-10-2014Pedersen masters-thesis-oct-10-2014
Pedersen masters-thesis-oct-10-2014
 
Strategic Cartography: Identifying IL Intersections Across the Curriculum
Strategic Cartography: Identifying IL Intersections Across the CurriculumStrategic Cartography: Identifying IL Intersections Across the Curriculum
Strategic Cartography: Identifying IL Intersections Across the Curriculum
 
Love Your Library CCL Button Templates - 2.25'' Multiple Color Pages with Hearts
Love Your Library CCL Button Templates - 2.25'' Multiple Color Pages with HeartsLove Your Library CCL Button Templates - 2.25'' Multiple Color Pages with Hearts
Love Your Library CCL Button Templates - 2.25'' Multiple Color Pages with Hearts
 
Why Reflect? The Holistic Practice of Stepping Back.
Why Reflect? The Holistic Practice of Stepping Back. Why Reflect? The Holistic Practice of Stepping Back.
Why Reflect? The Holistic Practice of Stepping Back.
 
Training the Trainers: Faculty Development Meets Information Literacy
Training the Trainers: Faculty Development Meets Information LiteracyTraining the Trainers: Faculty Development Meets Information Literacy
Training the Trainers: Faculty Development Meets Information Literacy
 
Screening Twitter Users for Depression and PTSD
Screening Twitter Users for Depression and PTSDScreening Twitter Users for Depression and PTSD
Screening Twitter Users for Depression and PTSD
 
A Gentle Introduction to the EM Algorithm
A Gentle Introduction to the EM AlgorithmA Gentle Introduction to the EM Algorithm
A Gentle Introduction to the EM Algorithm
 
Research problem
Research problem Research problem
Research problem
 
Information literacy through inquiry: using problem-based learning in informa...
Information literacy through inquiry: using problem-based learning in informa...Information literacy through inquiry: using problem-based learning in informa...
Information literacy through inquiry: using problem-based learning in informa...
 
The horizon isn't found in a dictionary : Identifying emerging word senses a...
The horizon isn't found in a  dictionary : Identifying emerging word senses a...The horizon isn't found in a  dictionary : Identifying emerging word senses a...
The horizon isn't found in a dictionary : Identifying emerging word senses a...
 
PDU 211 Research Methods: Identifying a Research Problem
PDU 211 Research Methods: Identifying a Research ProblemPDU 211 Research Methods: Identifying a Research Problem
PDU 211 Research Methods: Identifying a Research Problem
 
information privilege: access, advocacy, and the critical role of libraries.
information privilege: access, advocacy, and the critical role of libraries.information privilege: access, advocacy, and the critical role of libraries.
information privilege: access, advocacy, and the critical role of libraries.
 
Activate Your Learners! Active Learning Strategies for Fostering Participant ...
Activate Your Learners! Active Learning Strategies for Fostering Participant ...Activate Your Learners! Active Learning Strategies for Fostering Participant ...
Activate Your Learners! Active Learning Strategies for Fostering Participant ...
 

Similar to Pedersen semeval-2013-poster-may24

IA3_presentation.pptx
IA3_presentation.pptxIA3_presentation.pptx
IA3_presentation.pptxKtonNguyn2
 
Using networks to explore, quantify, and summarize phylogenetic tree space
Using networks to explore, quantify, and summarize phylogenetic tree spaceUsing networks to explore, quantify, and summarize phylogenetic tree space
Using networks to explore, quantify, and summarize phylogenetic tree spacejembrown
 
Improving Hardware Efficiency for DNN Applications
Improving Hardware Efficiency for DNN ApplicationsImproving Hardware Efficiency for DNN Applications
Improving Hardware Efficiency for DNN ApplicationsChester Chen
 
Task Adaptive Neural Network Search with Meta-Contrastive Learning
Task Adaptive Neural Network Search with Meta-Contrastive LearningTask Adaptive Neural Network Search with Meta-Contrastive Learning
Task Adaptive Neural Network Search with Meta-Contrastive LearningMLAI2
 
SVD and the Netflix Dataset
SVD and the Netflix DatasetSVD and the Netflix Dataset
SVD and the Netflix DatasetBen Mabey
 
powerpoint feb
powerpoint febpowerpoint feb
powerpoint febimu409
 
Junhua wang ai_next_con
Junhua wang ai_next_conJunhua wang ai_next_con
Junhua wang ai_next_conJunhua Wang
 
A general framework for predicting the optimal computing configuration for cl...
A general framework for predicting the optimal computing configuration for cl...A general framework for predicting the optimal computing configuration for cl...
A general framework for predicting the optimal computing configuration for cl...Scott Farley
 
ASE2010
ASE2010ASE2010
ASE2010swy351
 
"Quantum Hierarchical Risk Parity - A Quantum-Inspired Approach to Portfolio ...
"Quantum Hierarchical Risk Parity - A Quantum-Inspired Approach to Portfolio ..."Quantum Hierarchical Risk Parity - A Quantum-Inspired Approach to Portfolio ...
"Quantum Hierarchical Risk Parity - A Quantum-Inspired Approach to Portfolio ...Quantopian
 
DSUS_MAO_2012_Jie
DSUS_MAO_2012_JieDSUS_MAO_2012_Jie
DSUS_MAO_2012_JieMDO_Lab
 
Two methods for optimising cognitive model parameters
Two methods for optimising cognitive model parametersTwo methods for optimising cognitive model parameters
Two methods for optimising cognitive model parametersUniversity of Huddersfield
 
Keynote: Machine Learning for Design Automation at DAC 2018
Keynote:  Machine Learning for Design Automation at DAC 2018Keynote:  Machine Learning for Design Automation at DAC 2018
Keynote: Machine Learning for Design Automation at DAC 2018Manish Pandey
 
Deep Learning Inference at speed and scale
Deep Learning Inference at speed and scaleDeep Learning Inference at speed and scale
Deep Learning Inference at speed and scaleBill Liu
 
Enhanced Deep Residual Networks for Single Image Super-Resolution
Enhanced Deep Residual Networks for Single Image Super-ResolutionEnhanced Deep Residual Networks for Single Image Super-Resolution
Enhanced Deep Residual Networks for Single Image Super-ResolutionNAVER Engineering
 
Presentation vision transformersppt.pptx
Presentation vision transformersppt.pptxPresentation vision transformersppt.pptx
Presentation vision transformersppt.pptxhtn540
 
Inerview Quesion on Data Mining and Machine Learning
Inerview Quesion on Data Mining and Machine LearningInerview Quesion on Data Mining and Machine Learning
Inerview Quesion on Data Mining and Machine LearningYash Diwakar
 
Modeling Catastrophic Events in Spark: Spark Summit East Talk by Georg Hofman...
Modeling Catastrophic Events in Spark: Spark Summit East Talk by Georg Hofman...Modeling Catastrophic Events in Spark: Spark Summit East Talk by Georg Hofman...
Modeling Catastrophic Events in Spark: Spark Summit East Talk by Georg Hofman...Spark Summit
 
Memory efficient java tutorial practices and challenges
Memory efficient java tutorial practices and challengesMemory efficient java tutorial practices and challenges
Memory efficient java tutorial practices and challengesmustafa sarac
 

Similar to Pedersen semeval-2013-poster-may24 (20)

IA3_presentation.pptx
IA3_presentation.pptxIA3_presentation.pptx
IA3_presentation.pptx
 
Using networks to explore, quantify, and summarize phylogenetic tree space
Using networks to explore, quantify, and summarize phylogenetic tree spaceUsing networks to explore, quantify, and summarize phylogenetic tree space
Using networks to explore, quantify, and summarize phylogenetic tree space
 
Improving Hardware Efficiency for DNN Applications
Improving Hardware Efficiency for DNN ApplicationsImproving Hardware Efficiency for DNN Applications
Improving Hardware Efficiency for DNN Applications
 
Task Adaptive Neural Network Search with Meta-Contrastive Learning
Task Adaptive Neural Network Search with Meta-Contrastive LearningTask Adaptive Neural Network Search with Meta-Contrastive Learning
Task Adaptive Neural Network Search with Meta-Contrastive Learning
 
SVD and the Netflix Dataset
SVD and the Netflix DatasetSVD and the Netflix Dataset
SVD and the Netflix Dataset
 
powerpoint feb
powerpoint febpowerpoint feb
powerpoint feb
 
Junhua wang ai_next_con
Junhua wang ai_next_conJunhua wang ai_next_con
Junhua wang ai_next_con
 
A general framework for predicting the optimal computing configuration for cl...
A general framework for predicting the optimal computing configuration for cl...A general framework for predicting the optimal computing configuration for cl...
A general framework for predicting the optimal computing configuration for cl...
 
Blinkdb
BlinkdbBlinkdb
Blinkdb
 
ASE2010
ASE2010ASE2010
ASE2010
 
"Quantum Hierarchical Risk Parity - A Quantum-Inspired Approach to Portfolio ...
"Quantum Hierarchical Risk Parity - A Quantum-Inspired Approach to Portfolio ..."Quantum Hierarchical Risk Parity - A Quantum-Inspired Approach to Portfolio ...
"Quantum Hierarchical Risk Parity - A Quantum-Inspired Approach to Portfolio ...
 
DSUS_MAO_2012_Jie
DSUS_MAO_2012_JieDSUS_MAO_2012_Jie
DSUS_MAO_2012_Jie
 
Two methods for optimising cognitive model parameters
Two methods for optimising cognitive model parametersTwo methods for optimising cognitive model parameters
Two methods for optimising cognitive model parameters
 
Keynote: Machine Learning for Design Automation at DAC 2018
Keynote:  Machine Learning for Design Automation at DAC 2018Keynote:  Machine Learning for Design Automation at DAC 2018
Keynote: Machine Learning for Design Automation at DAC 2018
 
Deep Learning Inference at speed and scale
Deep Learning Inference at speed and scaleDeep Learning Inference at speed and scale
Deep Learning Inference at speed and scale
 
Enhanced Deep Residual Networks for Single Image Super-Resolution
Enhanced Deep Residual Networks for Single Image Super-ResolutionEnhanced Deep Residual Networks for Single Image Super-Resolution
Enhanced Deep Residual Networks for Single Image Super-Resolution
 
Presentation vision transformersppt.pptx
Presentation vision transformersppt.pptxPresentation vision transformersppt.pptx
Presentation vision transformersppt.pptx
 
Inerview Quesion on Data Mining and Machine Learning
Inerview Quesion on Data Mining and Machine LearningInerview Quesion on Data Mining and Machine Learning
Inerview Quesion on Data Mining and Machine Learning
 
Modeling Catastrophic Events in Spark: Spark Summit East Talk by Georg Hofman...
Modeling Catastrophic Events in Spark: Spark Summit East Talk by Georg Hofman...Modeling Catastrophic Events in Spark: Spark Summit East Talk by Georg Hofman...
Modeling Catastrophic Events in Spark: Spark Summit East Talk by Georg Hofman...
 
Memory efficient java tutorial practices and challenges
Memory efficient java tutorial practices and challengesMemory efficient java tutorial practices and challenges
Memory efficient java tutorial practices and challenges
 

More from University of Minnesota, Duluth

Muslims in Machine Learning workshop (NeurlPS 2021) - Automatically Identifyi...
Muslims in Machine Learning workshop (NeurlPS 2021) - Automatically Identifyi...Muslims in Machine Learning workshop (NeurlPS 2021) - Automatically Identifyi...
Muslims in Machine Learning workshop (NeurlPS 2021) - Automatically Identifyi...University of Minnesota, Duluth
 
Algorithmic Bias - What is it? Why should we care? What can we do about it?
Algorithmic Bias - What is it? Why should we care? What can we do about it? Algorithmic Bias - What is it? Why should we care? What can we do about it?
Algorithmic Bias - What is it? Why should we care? What can we do about it? University of Minnesota, Duluth
 
Algorithmic Bias : What is it? Why should we care? What can we do about it?
Algorithmic Bias : What is it? Why should we care? What can we do about it?Algorithmic Bias : What is it? Why should we care? What can we do about it?
Algorithmic Bias : What is it? Why should we care? What can we do about it?University of Minnesota, Duluth
 
Duluth at Semeval 2017 Task 6 - Language Models in Humor Detection
Duluth at Semeval 2017 Task 6 - Language Models in Humor Detection Duluth at Semeval 2017 Task 6 - Language Models in Humor Detection
Duluth at Semeval 2017 Task 6 - Language Models in Humor Detection University of Minnesota, Duluth
 
Who's to say what's funny? A computer using Language Models and Deep Learning...
Who's to say what's funny? A computer using Language Models and Deep Learning...Who's to say what's funny? A computer using Language Models and Deep Learning...
Who's to say what's funny? A computer using Language Models and Deep Learning...University of Minnesota, Duluth
 
Duluth at Semeval 2017 Task 7 - Puns upon a Midnight Dreary, Lexical Semantic...
Duluth at Semeval 2017 Task 7 - Puns upon a Midnight Dreary, Lexical Semantic...Duluth at Semeval 2017 Task 7 - Puns upon a Midnight Dreary, Lexical Semantic...
Duluth at Semeval 2017 Task 7 - Puns upon a Midnight Dreary, Lexical Semantic...University of Minnesota, Duluth
 
Puns upon a midnight dreary, lexical semantics for the weak and weary
Puns upon a midnight dreary, lexical semantics for the weak and wearyPuns upon a midnight dreary, lexical semantics for the weak and weary
Puns upon a midnight dreary, lexical semantics for the weak and wearyUniversity of Minnesota, Duluth
 
MICAI 2013 Tutorial Slides - Measuring the Similarity and Relatedness of Conc...
MICAI 2013 Tutorial Slides - Measuring the Similarity and Relatedness of Conc...MICAI 2013 Tutorial Slides - Measuring the Similarity and Relatedness of Conc...
MICAI 2013 Tutorial Slides - Measuring the Similarity and Relatedness of Conc...University of Minnesota, Duluth
 
What it's like to do a Master's thesis with me (Ted Pedersen)
What it's like to do a Master's thesis with me (Ted Pedersen)What it's like to do a Master's thesis with me (Ted Pedersen)
What it's like to do a Master's thesis with me (Ted Pedersen)University of Minnesota, Duluth
 

More from University of Minnesota, Duluth (20)

Muslims in Machine Learning workshop (NeurlPS 2021) - Automatically Identifyi...
Muslims in Machine Learning workshop (NeurlPS 2021) - Automatically Identifyi...Muslims in Machine Learning workshop (NeurlPS 2021) - Automatically Identifyi...
Muslims in Machine Learning workshop (NeurlPS 2021) - Automatically Identifyi...
 
Automatically Identifying Islamophobia in Social Media
Automatically Identifying Islamophobia in Social MediaAutomatically Identifying Islamophobia in Social Media
Automatically Identifying Islamophobia in Social Media
 
What Makes Hate Speech : an interactive workshop
What Makes Hate Speech : an interactive workshopWhat Makes Hate Speech : an interactive workshop
What Makes Hate Speech : an interactive workshop
 
Algorithmic Bias - What is it? Why should we care? What can we do about it?
Algorithmic Bias - What is it? Why should we care? What can we do about it? Algorithmic Bias - What is it? Why should we care? What can we do about it?
Algorithmic Bias - What is it? Why should we care? What can we do about it?
 
Algorithmic Bias : What is it? Why should we care? What can we do about it?
Algorithmic Bias : What is it? Why should we care? What can we do about it?Algorithmic Bias : What is it? Why should we care? What can we do about it?
Algorithmic Bias : What is it? Why should we care? What can we do about it?
 
Duluth at Semeval 2017 Task 6 - Language Models in Humor Detection
Duluth at Semeval 2017 Task 6 - Language Models in Humor Detection Duluth at Semeval 2017 Task 6 - Language Models in Humor Detection
Duluth at Semeval 2017 Task 6 - Language Models in Humor Detection
 
Who's to say what's funny? A computer using Language Models and Deep Learning...
Who's to say what's funny? A computer using Language Models and Deep Learning...Who's to say what's funny? A computer using Language Models and Deep Learning...
Who's to say what's funny? A computer using Language Models and Deep Learning...
 
Duluth at Semeval 2017 Task 7 - Puns upon a Midnight Dreary, Lexical Semantic...
Duluth at Semeval 2017 Task 7 - Puns upon a Midnight Dreary, Lexical Semantic...Duluth at Semeval 2017 Task 7 - Puns upon a Midnight Dreary, Lexical Semantic...
Duluth at Semeval 2017 Task 7 - Puns upon a Midnight Dreary, Lexical Semantic...
 
Puns upon a midnight dreary, lexical semantics for the weak and weary
Puns upon a midnight dreary, lexical semantics for the weak and wearyPuns upon a midnight dreary, lexical semantics for the weak and weary
Puns upon a midnight dreary, lexical semantics for the weak and weary
 
MICAI 2013 Tutorial Slides - Measuring the Similarity and Relatedness of Conc...
MICAI 2013 Tutorial Slides - Measuring the Similarity and Relatedness of Conc...MICAI 2013 Tutorial Slides - Measuring the Similarity and Relatedness of Conc...
MICAI 2013 Tutorial Slides - Measuring the Similarity and Relatedness of Conc...
 
What it's like to do a Master's thesis with me (Ted Pedersen)
What it's like to do a Master's thesis with me (Ted Pedersen)What it's like to do a Master's thesis with me (Ted Pedersen)
What it's like to do a Master's thesis with me (Ted Pedersen)
 
Talk at UAB, April 12, 2013
Talk at UAB, April 12, 2013Talk at UAB, April 12, 2013
Talk at UAB, April 12, 2013
 
Feb20 mayo-webinar-21feb2012
Feb20 mayo-webinar-21feb2012Feb20 mayo-webinar-21feb2012
Feb20 mayo-webinar-21feb2012
 
Ihi2012 semantic-similarity-tutorial-part1
Ihi2012 semantic-similarity-tutorial-part1Ihi2012 semantic-similarity-tutorial-part1
Ihi2012 semantic-similarity-tutorial-part1
 
Pedersen acl2011-business-meeting
Pedersen acl2011-business-meetingPedersen acl2011-business-meeting
Pedersen acl2011-business-meeting
 
Acm ihi-2010-pedersen-final
Acm ihi-2010-pedersen-finalAcm ihi-2010-pedersen-final
Acm ihi-2010-pedersen-final
 
Pedersen naacl-2010-poster
Pedersen naacl-2010-posterPedersen naacl-2010-poster
Pedersen naacl-2010-poster
 
Aaai 2006 Pedersen
Aaai 2006 PedersenAaai 2006 Pedersen
Aaai 2006 Pedersen
 
Advances In Wsd Aaai 2005
Advances In Wsd Aaai 2005Advances In Wsd Aaai 2005
Advances In Wsd Aaai 2005
 
Ijcai 2007 Pedersen
Ijcai 2007 PedersenIjcai 2007 Pedersen
Ijcai 2007 Pedersen
 

Pedersen semeval-2013-poster-may24

  • 1. Poster Design & Printing by Genigraphics® - 800.790.4001 Ted Pedersen Department of Computer Science University of Minnesota, Duluth tpederse@d.umn.edu http://www.d.umn.edu/~tpederse Ted Pedersen University of Minnesota, Duluth http://senseclusters.sourceforge.net Systems use different corpora to build co- occurrence matrix, otherwise nearly identical. Sys7 : Uses smallest corpora, just the 64 snippets per term. Resulting matrices range from 102 X 113 to 221 X 222. No SVD. Sys1: Use all 6400 snippets to create matrix of size 771 X 952. SVD reduced to 771 X 90. Sys9: Use first 10,000 paragraphs of APW data from English Gigaword. Resulting matrix is 9,853 X 10,995. SVD reduced to 9,853 X 300. RandX: a random baseline, assign each Web snippet to one of X random senses. Evaluation measures should be able to expose random baselines and give them appropriately low scores. Both the paired F-score used in SemEval-2010 and the Jaccard Coefficient satisfy this requirement. Given 64 snippets per ambiguous term, first order features were unlikely to succeed and achieved F-10 score of 36.10 According to F-10, F-SC, and Jaccard, smaller amounts of task specific data (sys7 and sys1) are more effective than large amounts of out of domain text (sys9) when using second order methods. Newspaper text (like APW as used in Sys9) is not typically what Web search locates. Results are more often commercial, Wikipedia, or current celebrity. Lessons learned? : ● For second order methods, use Web-like data, not news text ● Use more data, increase snippets with Web, Wikipedia, etc and then discard additions after clustering (?) ● Expand snippets by going to site in result • Task 11 goal? Cluster Web search results! Test Data? The top 64 Google results for each of 100 potentially ambiguous queries. Each result is a Web snippet about 25 words long. Challenges? Small amounts of data! • 64 results/term X 25 words/result = 1,600 words/term X 100 terms = 160,000 words Solution? Augment the data! • Use second order co-occurrences to enrich Web snippets to be clustered, friend of a friend relation • The bigrams car motor, car insurance, car magazine, life sentence, life force, and life insurance each represent a first order co-occurrence. Car and life are second order co-occurrences because both occur with insurance. Introduction Generic Duluth System Conclusion and Future Directions DiscussionThe Duluth SystemsAbstract Contact Duluth : Word Sense Induction Applied to Web Page Clustering Task 11 : Word Sense Induction & Disambiguation within an End-User Application Experimental Results System F-10 2010 Jaccard F1-13 2014 ARI Clusters /Size Sys1 46.53 31.79 56.83 5.75 2.5/ 26.5 Sys7 45.89 31.03 58.78 6.78 3.0/ 25.2 Sys9 35.56 22.24 57.02 2.59 3.3/ 19.8 Rand2 41.49 26.99 54.89 –0.04 2.0/ 32.0 Rand5 25.17 14.52 56.73 0.12 5.0/ 12.8 Rand10 15.05 8.18 59.67 0.02 10.0/ 6.4 Rand25 7.01 3.63 66.89? -0.15 23.2/ 2.8 Rand50 4.07 2.00 76.19? 0.10 35.9/ 1.8 MFS 54.06 39.90 54.42 0.00? 1.0/ 64.0 Gold 100.00 100.00 100.00 99.00 7.7/ 11.6 The Duluth systems that participated in Task 11 of SemEval–2013 carried out word sense induction (WSI) in order to cluster Web search results. They relied on an approach that represented Web snippets using second-order co- occurrences. These systems were all implemented using SenseClusters, a freely available open source software package. Web page clustering viewed as a word sense discrimination / induction problem where query term is the target word, snippet provides surrounding context •Create co-occurrence matrix from some corpus. Rows and columns made up of first and second word of bigrams identified by log-likelihood ratio. Low frequency and low scoring bigrams are excluded, as are stop words. Optionally reduce dimensionality with SVD. •Replace each word in a web snippet with a vector made up of its row from the co- occurrence matrix. Average all vectors for a snippet together to create a new context vector. This will capture second order co- occurrences between context vectors. •Cluster the resulting 64 vectors for an ambiguous term, find number of clusters automatically with PK2. Average number of senses is 7.7, but variance is quite high ... •heron island – 1 sense, 100% MFS •Shakira – 2 senses, 98% MFS •apple – 2 senses, 98% MFS •kawasaki – 7 senses, 47% MFS •Billy the Kid – 7 senses, 44% MFS •marble – 8 senses, 39% MFS •kangaroo – 17 senses, 48% MFS •ghost – 18 senses, 30% MFS •dog eat dog – 19 senses, 28% MFS •magic – 19 senses, 27% MFS Of 769 senses in the test data ... •467 (61%) occur less than 5 times!! • Is 1, 2, 3, 4, … instances enough to identify a cluster? • Very small clusters often “pure”, can trick some evaluation methods •186 (24%) are defined as “Other” • Criteria for membership unclear or different than other senses • Other different than can't cluster?