Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Semantic Search tutorial at SemTech 2012


Published on


Published in: Education, Technology

Semantic Search tutorial at SemTech 2012

  1. 1. Semantic Search Tutorial Introduction Peter Mika| Yahoo! Research, Spain Thanh Tran | Institute AIFB, KIT, Germany
  2. 2. About the speakers Peter Mika  Senior Research Scientist  Head of Semantic Search group at Yahoo! Research in Barcelona  Semantic Search, Web Object Retrieval, Natural Language Processing Tran Duc Thanh  Assistent Professor at AIFB, Karlsruhe Institute of Technology  Head of Semantic Search group  Semantic Search, Semantic / Linked Data Management
  3. 3. Agenda Introduction (10 min) Semantic Web data (50 min)  The RDF data model  Publishing RDF  Crawling and indexing RDF data Query processing (40 min) Ranking (30 min) Result presentation (15 min) Semantic Search evaluation (15 min) Questions (5 min)
  4. 4. Why Semantic Search? I. “We are at the beginning of search.“ (Marissa Mayer)  Solved large classes of queries, e.g. navigational  Heavy investment in computational power  Remaining queries are hard, not solvable by brute force, and require a deep understanding of the world and human cognition Background knowledge and metadata can help to address poorly solved queries
  5. 5. Poorly solved information needs Many of these queries Ambiguous searches would not be asked by  paris hilton users, who learned over Long tail queries time what search technology can and can  george bush (and I mean the beer brewer notArizona) in do. Multimedia search  paris hilton sexy Imprecise or overly precise searches  jim hendler  pictures of strong adventures people Precise searches for descriptions  countries in africa  32 year old computer scientist living in barcelona  reliable digital camera under 300 dollars
  6. 6. Example: multiple interpretations
  7. 7. Why Semantic Search? II. The Semantic Web is now a reality  Large amounts of data published in RDF  Heterogeneous data of varying quality  Users who are not skilled in writing complex queries (e.g. SPARQL) and may not be experts in the domain Searching data instead or in addition to searching documents  Direct answers  Novel search tasks
  8. 8. Example: direct answers in search Points of Faceted interest in Information search for Information box with Vienna, from the Shopping content from and Austria Knowledgeresults links to Yahoo! Graph Travel Since Aug, 2010, „regular‟ search results are „Powered by Bing‟
  9. 9. Novel search tasks Aggregation of search results  e.g. price comparison across websites Analysis and prediction  e.g. world temperature by 2020 Semantic profiling  recommendations based on particular interests Semantic log analysis  understanding user behavior in terms of objects Support for complex tasks  e.g. booking a vacation using a combination of services
  10. 10. Contextual (pervasive, ambient) search Yahoo! Connected TV: Widget engine embedded into the TV Yahoo! IntoNow: recognize audio and show related content
  11. 11. Interactive search and task completion
  12. 12. Document retrieval and data retrieval Information Retrieval (IR) support the retrieval of documents (document retrieval)  Representation based on lightweight syntax-centric models  Work well for topical search  Not so well for more complex information needs  Web scale Database (DB) and Knowledge-based Systems (KB) deliver more precise answers (data retrieval)  More expressive models  Allow for complex queries  Retrieve concrete answers that precisely match queries  Not just matching and filtering, but also joins  Limitations in scalability
  13. 13. Combination of document and data retrieval Documents with metadata  Metadata may be embedded inside the document  I’m looking for documents that mention countries in Africa. Data retrieval  Structured data, but searchable text fields  I’m looking for directors, who have directed movies where the synopsis mentions dinosaurs.
  14. 14. Semantic Search Target (combination of) document and data retrieval Semantic search is a retrieval paradigm that  Exploits the structure/semantics of the data or explicit background knowledge to understand user intent and the meaning of content  Incorporates the intent of the query and the meaning of content into the search process (semantic models) Wide range of semantic search systems  Employ different semantic models, possibly at different steps of the search process and in order to support different tasks
  15. 15. Semantic models Semantics is concerned with the meaning of the resources made available for search  Various representations of meaning Linguistic models: models of relationships among words  Taxonomies, thesauri, dictionaries of entity names  Inference along linguistic relations, e.g. broader/narrower terms Conceptual models: models of relationships among objects  Ontologies capture entities in the world and their relationships  Inference along domain-specific relations We will focus on conceptual models in this tutorial  In particular, the RDF/OWL conceptual model for
  16. 16. Semantic Search – a process view Knowledge Representation Semantic Models • Keywords Resources • Forms Query • NLConstruction • Formal language •IR-style matching & ranking •DB-style precise matching Query •KB-style matching &Processing inferences •Query visualization •Document and data Documents Result presentationPresentation •Summarization •Implicit feedback •Explicit feedback Query •IncentivesRefinement Document Representation
  17. 17. Semantic Search systemsFor data / document retrieval, semantic searchsystems might combine a range of techniques, rangingfrom statistics-based IR methods for ranking,database methods for efficient indexing and queryprocessing, up to complex reasoning techniques formaking inferences!
  18. 18. Example: Information Workbench Addressing the lifecycle of interacting with the Web of Data  Integration of data sources  Content generation by the end user  Search and Exploration  Visualization User- Wikipedia  Publishing generated DBpedia, Yago Integrated management of Earthquake heterogeneous data ( sources  Structured and unstructured Structure  Published and user-generated Dynamic d  Static and dynamic  Open domain
  19. 19. Data Sources in the Application Entire English Wikipedia Data from Linked Open Data  DBpedia  YAGO … Data from (US Government)  E.g. live data about earthquakes Many more
  20. 20. Semantic Search Hybrid Search: Structured queries combined with keywords across structured and unstructured data sources Query interpretation: Translation of keywords into hybrid queries Keyword search/query interpretation combined with faceted search: iterative refinement process based on keywords and operations on facets
  21. 21. Search, Refinement and Navigation Keywords Query Translations Term CompletionsVorlesung Knowledge Discovery - Institut Facets AIFB21
  22. 22. Result Inspection, Analysis and Browsing
  23. 23. Semantic Web data
  24. 24. Data on the Web Data on the Web is not directly accessible  Most web pages are generated from databases, but formatted for human consumption  APIs offer limited views over data Two solutions  Extraction using Information Extraction (IE) techniques  Out of scope for this tutorial  Relying on publishers to expose structured data using standard Semantic Web formats
  25. 25. Information extraction Natural Language Processing  Named entity recognition and disambiguation, sentiment analysis etc. Extraction of information about entities  Suchanek et al. YAGO: A Core of Semantic Knowledge Unifying WordNet and Wikipedia, WWW, 2007.  Wu and Weld. Autonomously Semantifying Wikipedia, CIKM 2007. Extraction from HTML tables  Cafarella et al. WebTables: Exploring the Power of Tables on the Web. VLDB 2008 Wrapper induction  Kushmerick et al. Wrapper Induction for Information ExtractionText extraction. IJCAI 2007 Filling web forms automatically (form-filling)  Madhavan et al. Googles Deep-Web Crawl. VLDB 2008
  26. 26. Semantic Web Sharing data across the Web  Standard data model  RDF  A number of syntaxes (file formats)  RDF/XML, RDFa  Powerful, logic-based languages for schemas  OWL, RIF  Query languages and protocols  HTTP, SPARQL
  27. 27. Resource Description Framework (RDF) Each resource (thing, entity) is identified by a URI  Globally unique identifiers  URLs a subset of URIs  Often abbreviated using namespaces  e.g. example:roi = RDF represents knowledge as a set of triples  Each triple is a single fact about the entity (an attribute or a relationship) A set of triples forms an RDF graph RDF document type foaf:Person example:roi name “Roi Blanco”
  28. 28. Linking resourcesRoi‟s homepage Friend-of-a-Friend ontology type example:roi foaf:Person name “Roi Blanco” knows sameAsYahoo!‟s website type worksWith #roi2 #peter email “”
  29. 29. Ontologies Ontologies are the schemas for RDF graphs  Define the intended meaning of certain classes and relationships in a domain  e.g. the FOAF ontology defines the foaf:Person class and relationships such as foaf:knows Ontologies are published in standard languages such as OWL  Language defines the constructs for modeling and their meaning  e.g. subClassOf, sameAs  Tools for editing ontologies such as Protégé, TopBraid Composer Reusing and extending well-known ontologies helps to interpret unknown data  e.g. if X is subClassOf foaf:Person, and A is of type X, then we know that A is a person
  30. 30. Example: Agreement on a shared set of schemas for common types of web content  Bing, Google, and Yahoo! as initial supporters  Similar in intent to (2006)  Use a single format to communicate the same information to all three search engines Support for microdata covers areas of interest to all search engines  Business listings (local), creative works (video), recipes, reviews  User defined extensions Each search engine continues to develop its products
  31. 31. Documentation and OWL ontology
  32. 32. Publishing RDF Interlinked RDF documents (Linked Data)  Each document describes a single resource with URIs pointing to related resources  Common RDF file formats are RDF/XML and Turtle  Mostly implemented as a wrapper around a database or Web service Embedding RDF inside HTML  RDFa, microdata SPARQL endpoints  Triple stores are databases for managing RDF data  SPARQL is a standard protocol and query language for accessing triple stores using HTTP
  33. 33. Linked Data Grass-roots community effort to (re)publish open datasets in RDF  Centered around the Dbpedia project  Directory of datasets at,
  34. 34. Metadata in HTML anno 1995 What does this term<HTML> mean?<HEAD profile=" html/"><META name="" content="Peter Mika"><LINK rel="DC.rights copyright" href="" /><LINK rel="meta" type="application/rdf+xml" title="FOAF" href= ""></HEAD>…</HTML> How is this data related?
  35. 35. Microformats (anno 2003) Mark-up for data in HTML pages  Reuse existing HTML elements (class, rel) Microformats exist for a limited set of objects  Persons, organizations, reviews, events etc., see No formal schemas  Limited reuse, extensibility of schemas  Unclear which combinations are allowed No identifiers for entities  No interlinking between entities<div class="vcard"> <a class="email fn" href="">Joe Friday</a> <div class="tel">+1-919-555-7878</div> <div class="title">Area Administrator, Assistant</div></div>
  36. 36. RDFa and RDFa Lite (2008-2012) W3C standards for embedding RDF data in HTML documents  A set of new HTML attributes to be used in head or body  A specification of how to extract the data from these attributes RDFa is just a syntax, you have to choose a vocabulary separately RDFa family  RDFa 1.0  Recommendation since October, 2008  RDFa 1.1 is a small update on RDFa to make it easier to use  RDFa Primer (Working Draft, May 8, 2012)  RDFa 1.1 Lite is a subset of RDFa 1.1 with the most
  37. 37. Example: Facebook‟s Open Graph Protocol The „Like‟ button provides publishers with a way to promote their content on Facebook and build communities  Shows up in profiles and news feed  Site owners can later reach users who have liked an object  Facebook Graph API allows 3rd party developers to access the data Open Graph Protocol is an RDFa-based format that allows to describe the object that the user „Likes‟
  38. 38. Example: Facebook‟s Open Graph Protocol RDF vocabulary to be used in conjunction with RDFa  Simplify the work of developers by restricting the freedom in RDFa Activities, Businesses, Groups, Organizations, People, Places, Product s and Entertainment Only HTML <head> accepted<html xmlns:og=""><head> <title>The Rock (1996)</title> <meta property="og:title" content="The Rock" /> <meta property="og:type" content="movie" /> <meta property="og:url" content="" /> <meta property="og:image" content="" /> …</head> ...
  39. 39. Microdata HTML5  Currently under standardization at the W3C  Originally part of the HTML5 spec, but now a separate document Comparable to RDFa 1.1 Lite  Key extensibility features (such as multiple types) missing HTML5 also has a number of “semantic” elements<div itemscope itemid=“”><p>My name is <span <video>, <article>… such as <time>, itemprop="name">Neil</span>.</p><p>My band is called <span itemprop="band">Four Parts Water</span>. I was born on <time itemprop="birthday" datetime="2009-05-10"> May 10th 2009</time>. <img itemprop="image" src=”me.png" alt=”me”></p></div>
  40. 40. Current state of metadata on the Web 31% of webpages, 5% of domains contain some metadata  Analysis of the Bing Crawl (US crawl, January, 2012)  RDFa is most common format  By URL: 25% RDFa, 7% microdata, 9% microformat  By eTLD (PLD): 4% RDFa, 0.3% microdata, 5.4% microformat  Adoption is stronger among large publishers  Especially for RDFa and microdata See also  P. Mika, T. Potter. Metadata Statistics for a Large Web Corpus, LDOW 2012  H.Mühleisen, C.Bizer.Web Data Commons - Extracting Structured Data from Two Large Web Corpora, LDOW 2012
  41. 41. Exponential growth in RDFa data Another five-fold increase between October 2010 and January, 2012 Five-fold increase between March, 2009 and October, 2010Percentage of URLs with embedded metadata in various formats
  42. 42. Crawling and indexing
  43. 43. Crawling the Semantic Web Linked Data  Similar to HTML crawling, but the the crawler needs to parse RDF/XML (and others) to extract URIs to be crawled  Semantic Sitemap/VOID descriptions RDFa  Same as HTML crawling, but data is extracted after crawling  Mika et al. Investigating the Semantic Gap through Query Log Analysis, ISWC 2010. SPARQL endpoints  Endpoints are not linked, need to be discovered by other means  Semantic Sitemap/VOID descriptions
  44. 44. Data fusion Ontology matching  Widely studied in Semantic Web research, see e.g. list of publications at  Unfortunately, not much of it is applicable in a Web context due to the quality of ontologies Entity resolution  Logic-based approaches in the Semantic Web  Studied as record linkage in the database literature  Machine learning based approaches, focusing on attributes  Graph-based approaches, see e.g. the work of Lisa Getoor are applicable to RDF data  Improvements over only attribute based matching Blending  Merging objects that represent the same real world entity and reconciling information from multiple sources
  45. 45. Data quality assessment and curation Heterogeneity, quality of data is an even larger issue  Quality ranges from well-curated data sets (e.g. Freebase) to microformats  In the worst of cases, the data becomes a graph of words  Short amounts of text: prone to mistakes in data entry or extraction  Example: mistake in a phone number or state code Quality assessment and data curation  Quality varies from data created by experts to user-generated content  Automated data validation  Against known-good data or using triangulation  Validation against the ontology or using probabilistic models  Data validation by trained professionals or crowdsourcing  Sampling data for evaluation
  46. 46. Indexing Search requires matching and ranking  Matching selects a subset of the elements to be scored The goal of indexing is to speed up matching  Retrieval needs to be performed in milliseconds  Without an index, retrieval would require streaming through the collection The type of index depends on the query model to support  DB-style indexing  IR-style indexing
  47. 47. IR-style indexing Index data as text  Create virtual documents from data  One virtual document per subgraph, resource or triple  typically: resource Key differences to Text Retrieval  RDF data is structured  Minimally, queries on property values are required
  48. 48. Horizontal index structure Two fields (indices): one for terms, one for properties For each term, store the property on the same position in the property index  Positions are required even without phrase queries Query engine needs to support the alignment operator ✓ Dictionary is number of unique terms + number of properties Occurrences is number of tokens * 2
  49. 49. Vertical index structure One field (index) per property Positions are not required  But useful for phrase queries Query engine needs to support fields Dictionary is number of unique terms Occurrences is number of tokens ✗ Number of fields is a problem for merging, query performance
  50. 50. Distributed indexing MapReduce is ideal for building inverted indices  Map creates (term, {doc1}) pairs  Reduce collects all docs for the same term: (term, {doc1, doc2…}  Sub-indices are merged separately  Term-partitioned indices Peter Mika. Distributed Indexing for Semantic Search, SemSearch 2010.
  51. 51. Query Processing / Matching
  52. 52. Structure Taxonomy of search approaches Query processing / matching techniques for Semantic Search Types of semantic data Formalisms for querying semantic data Approaches  General task: hybrid graph pattern matching  Matching keyword query against text  Matching structured query against structured data  Matching keyword query against structured data  Matching structured query against text (a hybrid case) Main tasks, challenges and opportunities
  53. 53. Taxonomy of search approaches (1) The search problem  A collection of resources, called data  Information needs expressed as queries  Search is the task of efficiently computing results from data that are relevant to queries Document data retrieval vs. structured data retrieval  Differences in query and data representation and matching  Efficiently retrieve structured data that exactly match formal information needs expressed as structured queries  Effectively rank textual results that match ambiguous NL / keyword queries to a certain degree (notions of relevance) Semantic search: ranked retrieval of document and
  54. 54. Taxonomy of search approaches (2) Query engines (of databases)  Exact  Complete Query  Sound • Approximate Matching • Not complete • Not sound • Ranked Data • Best effort • Top-k Search engines (stand-alone/database extensons)Query processing mainly focuses on efficiency of matchingwhereas ranking deals with degree of matching (relevance)!
  55. 55. Query processing for Semantic Search (1) Resources represented by semantic data ranging from  Structured data with well defined schemas  Semi-structured data with incomplete or no schemas  Data that largely comprise text  Hybrid / embedded data Information needs of varying complexity, captured using different formalisms and querying paradigms  Natural language texts and keywords  Form-based inputs  Formal structured queries (Search is end-user oriented paradigm, requires “natural”, intuitive querying interfaces) Semantic search: efficiently computing results (query processing) from data that are relevant to queries (ranking)
  56. 56. Query processing for Semantic Search (2) NL Form- / facet- Structured Queries Keywords Questions based Inputs (SPARQL)Ambiquities Query Matching Data RDF data Semi- OWL ontologies with Structured embedded in Structured rich, formal RDF data text (RDFa) RDF data semanticsAmbiquities: confidence degree, truth/trust
  57. 57. Query processing for Semantic Search (3) Textual Data Structured query on Keyword query textual data on textual , e.g. querying data, e.g. Web Search target Semantic extension for search systems group of different search users, information systems?Unstructured needs, and types Structured of data. StructuredQuery Query Query processing for on Keyword query query on structured search is hybrid semantic structured data data, e.g. e.g. standard combination of techniques! search querying extensions for interface for databases databases / RDF stores Structured Data
  58. 58. Types of data models (1) Textual  Bag-of-words  Represent documents, text in structured data,…, real- world objects (captured as structured data)  Lacks “structure”  in text, e.g. linguistic structure, hyperlinks, (positional information)  Structure in structured data representation term (statistics) In combination with combination Cloud Computing Cloud technologies, promising Computing solutions for the Technologies management of `big solutions data have emerged. management Existing industry `big data solutions are able to industry support complex solutions queries and analytics support tasks with terabytes of complex data. For …… example, using a Greenplum.
  59. 59. Types of data models (2) Textual Structured  Resource Description Framework (RDF)  Represent real-world objects, services, applications, …. documents  Resource attribute values and relationships between resources  Schema Picture creator Person Bob
  60. 60. Types of data models (3) Textual Structured Hybrid  RDF data embedded in text (RDFa)
  61. 61. Types of data models – RDFa (1)…<div about="/alice/posts/trouble_with_bob"> <h2 property="dc:title">The trouble with Bob</h2> <h3 property="dc:creator">Alice</h3> Bob is a good friend of mine. We went to the same university, and also shared an apartment in Berlin in 2008. The trouble with Bob is that he takes much better photos than I do: <div about=""> <img src="" /> <span property="dc:title">Beautiful Sunset</span> by <span property="dc:creator">Bob</span>. </div></div>… adopted from :
  62. 62. Types of semantic data – RDFa (2)Bob is a good friend of mine. We contentwent to the same university, andalso shared an apartment in Berlinin 2008. The trouble with Bob isthat he takes much better photosthan I do: content adopted from :
  63. 63. Types of semantic data - conclusionSemantic data in general can be conceived as a graph with text and structured data items as nodes, and edges represent different types of relationships including explicit semantic relationships andvaguely specified ones such as hyperlinks!
  64. 64. Formalisms for querying semantic data (1)Example information need“Information about a friend of Alice, who sharedan apartment with her in Berlin and knowssomeone working at KIT.” Unstructured queries Fully-structured queries Hybrid queries: unstructured + structured
  65. 65. Formalisms for querying semantic data (2)Example information need“Information about a friend of Alice, who sharedan apartment with her in Berlin and knowssomeone working at KIT.” Unstructured  NL  Keywords shared apartment Berlin Alice
  66. 66. Formalisms for querying semantic data (3)Example information need“Information about a friend of Alice, who sharedan apartment with her in Berlin and knowssomeone working at KIT.” Unstructured Fully-structured  SPARQL: BGP, filter, optional, union, select, construct, ask, describe  PREFIX ns: <> SELECT ?x WHERE { ?x ns:knows ? y. ?y ns:name “Alice”. ?x ns:knows ?z. ?z ns: works ?v. ?v ns:name “KIT” }
  67. 67. Formalisms for querying semantic data (4) Fully-structured Unstructured Hybrid: content and structure constraints “shared apartment Berlin Alice” ?x ns:knows ? y. ?y ns:name “Alice”. ?x ns:knows ?z. ?z ns: works ?v. ?v ns:name “KIT”
  68. 68. Formalisms for querying semantic data (5) Fully-structured Unstructured Hybrid: content and structure constraints “shared apartment Berlin Alice” ?x ns:knows ? y. ?y ns:name “Alice”. ?x ns:knows ?z. ?z ns: works ?v. ?v ns:name “KIT”
  69. 69. Formalisms for querying semantic data - conclusionSemantic search queries can be conceived as graph patterns with nodes referring totext and structured data items, and edges referring to relationships between these items!
  70. 70. Processing hybrid graph patterns (1) Example information need “Information about a friend of Alice, who shared an apartment with her in Berlin and knows someone working at KIT.”apartment shared Berlin Alice ?x ns:knows ?z. ?z ns: works ?v. ?v ns:name “KIT” ?y ns:name “Alice”. ?x ns:knows ? y trouble with bob FluidOps 34 Peter sunset.jpg Bob is a good friend Beautiful of mine. We went to Sunset the same Germany Semantic Alice Search university, and also shared an apartment in Berlin in 2008. The trouble Germany 2009 Bob with Bob is that he Thanh takes much better photos than I do: KIT
  71. 71. Processing hybrid graph patterns (2) Matching hybrid graph patterns against data
  72. 72. Matching keyword query against text • Retrieve documents • Inverted list (inverted index) keyword  {<doc1, pos, score, ...>, <doc2, pos, score, ...>, ...} • AND-semantics: top-k joinshared Berlin Alice shared Berlin Alice D1 D1 D1 shared = berlin = alice shared
  73. 73. Matching structured query against structured data • Retrieve data for triple patterns • Index on tables • Multiple “redundant” indexes to cover different access patterns • Join (conjunction of triples) • Blocking, e.g. linear merge join (required sorted input) • Non-blocking, e.g. symmetric hash-join • Materialized join indexesPer1 ns:works ?v ?v ns:name “KIT” ?x ns:knows ?y. ?x ns:knows ?z. SP-index PO-index ?z ns: works ?v. ?v ns:name “KIT” = = = Per1 ns:works Ins1 Ins1 ns:name KITPer1 ns:works Ins1 Ins1 ns:name KIT
  74. 74. Matching keyword query against structured data • Retrieve keyword elements • Using inverted index keyword  {<el1, score, ...>, <el2, score, ...>,…} • Exploration / “Join” • Data indexes for triple lookup • Materialized index (paths up to graphs) • Top-k Steiner tree search, top-k subgraph exploration Alice Bob KIT Alice Bob KIT ↔ ↔ = =Alice ns:knows Bob Inst1 ns:name KIT Bob ns:works Inst1
  75. 75. Matching structured query against text• Based on offline IE (offline see Peter‟s slides)• Based on online IE, i.e., “retrieve “ is as follows • Derive keywords to retrieve relevant documents • On-the-fly information extraction, i.e., phrase pattern matching “X name Y” • Retrieve extracted data for structured part • Retrieve documents for derived text patterns, e.g. sequence, windows, reg. exp. ?x ns:knows ?y. ?x ns:knows ? ?z ns: works ?v. ?v ns:name “K knows name KIT
  76. 76. Matching structured query against text• Index • Inverted index for document retrieval and pattern matching • Join index  inverted index for storing materialized joins between keywords • Neighborhood indexes for phrase patterns ?x ns:knows ?y. ?x ns:knows ? ?z ns: works ?v. ?v ns:name “K KIT name knows name KIT
  77. 77. Query processing – main tasks  Retrieval  Documents , data Query elements, triples, paths, graphs  Inverted index,…, but also other (B+ tree) Matching  Index documents, triples, materialized paths  Join Data  Different join implementations, efficiency depends on availability of indexes  Non-blocking join good for early result reporting and for “unpredictable” Linked Data / data
  78. 78. Query processing – more tasks  More complex queries: disjunction, aggregation, grouping, an Query alytics…  Join order optimization  Approximate  Approximate the search space Matching  Approximate the results (matching, join)  Parallelization  Top-k Data  Use only some entries in the input streams to produce k results  Multiple sources  Federation, routing  On-the-fly mapping, similarity join  Hybrid
  79. 79. Query processing on the Web - research challenges and opportunities Large amount of semantic data • Optimization, parallelizati Data on inconsistent, redunda • Approximation nt, and low quality • Hybrid querying and data Large amount of data management embedded in text • Federation, routing Large amount of • Online schema sources mappings Large amount of links • Similarity join between sources
  80. 80. Ranking
  81. 81. Structure Problem definition Types of ambiguities Ranking paradigms Model construction  Content-based  Structure-based
  82. 82. Ranking – problem definition Query • Ambiguities arise when representation is incomplete / imprecise Matching • Ambiguities at the level of • elements (content ambiguity) • structure between elements Data (structure ambiguity)Due to ambiguities in the representation of theinformation needs and the underlying resources, theresults cannot be guaranteed to exactly match the query.Ranking is the problem of determining the degree ofmatching using some notions of relevance.
  83. 83. Content ambiguityapartment shared Berlin Alice ?x ns:knows ?z. ?z ns: works ?v. ?v ns:name “KIT” ?y ns:name “Alice”. ?x ns:knows ? y trouble with bob FluidOps 34 Peter sunset.jpg Bob is a good friend Beautiful of mine. We went to Sunset the same Germany Semantic Alice Search university, and also shared an apartment in Berlin in 2008. The trouble Germany 2009 Bob with Bob is that he Thanh takes much better photos than I do: KIT What is meant by “Berlin” in the query? What is meant by “KIT” in the query? What is meant by “Berlin” in the data? What is meant by “KIT” in the data? A city with the name Berlin? a person? A research group? a university? a location?
  84. 84. Structure ambiguityapartment shared Berlin Alice ?x ns:knows ?z. ?z ns: works ?v. ?v ns:name “KIT” ?y ns:name “Alice”. ?x ns:knows ? y trouble with bob FluidOps 34 Peter sunset.jpg Bob is a good friend Beautiful of mine. We went to Sunset the same Germany Semantic Alice Search university, and also shared an apartment in Berlin in 2008. The trouble Germany 2009 Bob with Bob is that he Thanh takes much better photos than I do: KIT What is the connection between What is meant by “works”? “Berlin” and “Alice”? Works at? employed? Friend? Co-worker?
  85. 85. Ambiguity Recall: query processing is matching at the level of syntax and semantics Ambiguities arise when data or query allow for multiple interpretations, i.e. multiple matches  Syntactic, e.g. works vs. works at  Semantic, e.g. works vs. employ “Aboutness”, i.e., contain some elements which represent the correct interpretation  Ambiguities arise when matching elements of different granularities  Does i contains the interpretation for j, given some part(s) of i (syntactically/semantically) match j  E.g. Berlin vs. “…we went to the same university, and also, we shared an apartment in Berlin in 2008…” Strictly speaking, ranking is performed after syntactic / semantic matching is done!
  86. 86. Features: What to use to deal with ambiguities?What is meant by “Berlin”? What is theconnection between “Berlin” and “Alice”? Content features  Frequencies of terms: d more likely to be “about” a query term k when d more often, mentions k (probabilistic IR)  Co-occurrences: terms K that often co-occur form a contextual interpretation, i.e., topics (cluster hypothesis) Structure features  Consider relevance at level of fields  Linked-based popularity
  87. 87. Ranking paradigms Explicit relevance model  Foundation: probability ranking principle  Ranking results by the posterior probability (odds) of being observed in the relevant class:  P(w|R) varies in different approaches, e.g., binary independence model, 2-poisson model, relevance model P(D|R) P( D | R) P(w | R) (1 P(w | N )) P(D/N) w D w D k P( w | R) P( w | q1 ,..., qk ) P ( m) P ( w | m) P(qk | m) m M i 1
  88. 88. Ranking paradigms  No explicit notion of relevance: similarity between the query and the document model  Vector space model (cosine similarity)  Language models (KL divergence) Sim(q, d ) Cos(( w1,d ,..., wt , d ), ( w1,q ,..., wk , q )) P(t | q )Sim(q, d ) KL( q || d ) P(t | q ) log( t V P(t | d )
  89. 89. Model construction How to obtain  Relevance models?  Weights for query / document terms?  Language models for document / queries?
  90. 90. Content-based model construction Document statistics, e.g. • An object is more likely  Term frequency about “Berlin”?  Document length • When it contains a Collection statistics, e.g. relatively high number of mentions of the term  Inverse document “Berlin” frequency • When the number of  Background language mentions of this term in models the overall collection is tf relatively low wt , d idf |d | tfP(t | d ) (1 ) P(t | C ) |d |
  91. 91. Structure-based model construction Consider structure of objects during content-based modeling, i.e., to obtain structured content-based model  Content-based model for structured objects, documents and for general tuples P(t | d ) f P(t | f ) f Fd• An object is more likely about “Berlin”? • When one of its (important) fields contains a relatively high number of mentions of the term “Berlin”
  92. 92. Structure-based model construction  PageRank  Link analysis algorithm  Measuring relative importance of nodes  Link counts as a vote of support  The PageRank of a node recursively depends on the number and PageRank of all nodes that link to it (incoming links)  ObjectRank  Types and semantics of links vary in structured data setting  Authority transfer schema graph specifies connection• An object about “Berlin” is more important than one another? strengths • When a relatively large number of objects are linked to it  Recursively compute authority transfer data graph
  93. 93. Taxonomy of ranking approaches Explicitly vs. non-explicitly relevance-based Content-based ranking Structure-based ranking Content- and-structure-based ranking
  94. 94. Result Presentation
  95. 95. Search interface Input and output functionality  helping the user to formulate complex queries  presenting the results in an intelligent manner Semantic Search brings improvements in  Query formulation  Snippet generation  Suggesting related entities  Adaptive and interactive presentation  Presentation adapts to the kind of query and results presented  Object results can be actionable, e.g. buy this product  Aggregated search  Grouping similar items, summarizing results in various ways  Filtering (facets), possibly across different dimensions  Task completion  Help the user to fulfill the task by placing the query in a task context
  96. 96. Query formulation “Snap-to-grid”: suggest the most likely interpretation of the query  Given the ontology or a summary of the data  While the user is typing or after issuing the query  Example: Freebase suggest, TrueKnowledge
  97. 97. Enhanced results/Rich Snippets Use mark-up from the webpage to generate search snippets  Originally invented at Yahoo! (SearchMonkey) Google, Yahoo!, Bing, Yandex now consume markup  Validators available from Google and Bing
  98. 98. Other result presentation tasks Select the most relevant resources within an RDF document  Penin et al. Snippet Generation for Semantic Web Search Engines, ASWC 2010 For each resource, rank the properties to be displayed Natural Language Generation (NLG)  Verbalize, explain results
  99. 99. Aggregated search: facets
  100. 100. Aggregated search:
  101. 101. Related entities Related actors and movies
  102. 102. Adaptive presentation:housing search
  103. 103. Evaluation
  104. 104. Semantic Search challenge (2010/2011) Two tasks  Entity Search  Queries where the user is looking for a single real world object  Pound et al. Ad-hoc Object Retrieval in the Web of Data, WWW 2010.  List search (new in 2011)  Queries where the user is looking for a class of objects Billion Triples Challenge 2009 dataset Evaluated using Amazon‟s Mechanical Turk  Halpin et al. Evaluating Ad-Hoc Object Retrieval, IWEST 2010  Blanco et al. Repeatable and Reliable Search System Evaluation using Crowd-Sourcing, SIGIR2011
  105. 105. Evaluation form
  106. 106. Catching the bad guys Payment can be rejected for workers who try to game the system  An explanation is commonly expected, though cheaters rarely complain We opted to mix control questions into the real results  Gold-win cases that are known to be perfect  Gold-loose cases that are known to be bad Metrics Worker and std. dev onReal  Avg. Known gold-win and gold-loose results Known Good Total Time to  Time to complete bad N complete N Mean N Mean N Mean (sec) badguy 20 2.556 200 2.738 20 2.684 240 29.6 goodguy 13 1 130 2.038 13 3 156 95 whoknows 1 1 21 1.571 2 3 24 83.5
  107. 107. Other evaluations TREC Entity Track  Related Entity Finding  Entities related to a given entity through a particular relationship  Retrieval over documents (ClueWeb 09 collection)  Example: (Homepages of) airlines that fly Boeing 747  Entity List Completion  Given some elements of a list of entities, complete the list Question Answering over Linked Data  Retrieval over specific datasets (Dbpedia and MusicBrainz)  Full natural language questions of different forms  Correct results defined by an equivalent SPARQL query  Example: Give me all actors starring in Batman Begins.
  108. 108. Resources
  109. 109. Resources Books  Ricardo Baeza-Yates and Berthier Ribeiro-Neto. Modern Information Retrieval. ACM Press. 2011 Survey papers  Thanh Tran, Peter Mika. Survey of Semantic Search Approaches. Under submission, 2012. Conferences and workshops  ISWC, ESWC, WWW, SIGIR, CIKM, SemTech  Semantic Search workshop series  Exploiting Semantic Annotations in Information Retrieval (ESAIR)  Entity-oriented Search (EOS) workshop