Semantic Access to Data from the Web
description
Transcript of Semantic Access to Data from the Web
Semantic Access to Data from the WebRaquel Trillo*, Laura Po+, Sergio Ilarri*, Sonia Bergamaschi+ and E. Mena*
1st International Workshop on Interoperability through Semantic Data and Service Integration (ISDSI’09)
Cagmoli (Genova), Italy, 25th June 2009
Distributed Information Systems Group, http://sid.cps.unizar.es University of Zaragoza, Spain
Databases Group, http://www.dbgroup.unimo.it/ Univ. Of Modena e Reggio Emilia, Italy
* +
Outline Introduction.
Basic Architecture of the system: Discovering the Semantics of User Keywords. Semantics-Guided Data Retrieval.
Improvements to the Basic Architecture: Probabilistic Word Sense Disambiguation. Retrieval of Synonyms of User Keywords.
Conclusions and Future Work.
ISDSI’09 Cagmoli (Italy), 25th June 2009
Introduction
ISDSI’09 Cagmoli (Italy), 25th June 2009
Search engines have become the best allies of users. They index most no hidden Web. They succeed when users ask for popular information on
the Web.
Traditional Search engines are based on syntactic techniques (no semantics): Polysemous Words: with several meanings
(senses/interpretations). Example: Mouse (animal, Mickey Mouse, input device,
etc).
Synonymous Words: Different representations (words) with the same meaning.
Example: automobile or car Example: lorry or truck
Introduction
ISDSI’09 Cagmoli (Italy), 25th June 2009
Truck
172,000,000
Lorry
4,760,000
Introduction: Semantic Search
ISDSI’09 Cagmoli (Italy), 25th June 2009
Semantic Search engines can overcome the problems of traditional search engines. Consider the semantics of keywords and not
only its representation (how they are written). Our proposal:
Classify the results of traditional search in different categories by considering their possible meanings.
Considers the synonyms of the user keywords to retrieve more pages.
Introduction: Web Clustering
ISDSI’09 Cagmoli (Italy), 25th June 2009
Along last decades, different techniques to cluster documents have appeared: Traditional clustering algorithms cannot be
applied to search result clustering. Features that a clustering for web search should:
Separate relevant pages for the user from irrelevant ones.
Provide browsable summaries of each cluster.
Be applied to snippets and not to whole pages. Be incremental and provide results ASAP. Allow the overlapping between groups.
OutlineIntroduction.Basic Architecture of the system:
Discovering the Semantics of User Keywords: Obtaining the possible keyword senses (meanings). Selecting the most probable sense of each user keyword.
Semantics-Guided Data Retrieval: Lexical annotations of results of a traditional search. Categorization of results.
Improvements to the Basic Architecture:Conclusions and Future Work.
ISDSI’09 Cagmoli (Italy), 25th June 2009
Basic Architecture of the system
Discovering the semantics of User keywords
Semantics-Guided Data Retrieval
Extraction of keyword senses
Disambiguation of keyword senses
Selection of the most probable intended category
Categorization of hits
Lexical annotation of hits: title and snippet
Search keywords in traditional search engines
Possible keyword senses
Selected senses
Hits (results of a traditional SE)
Annotated Hits by considering the Possible Keyword ss
Clusters or categories of hits
Semantic Cluster of Hits
Keywords
Goal: Discover the intended meaning of each user keyword. How: Word Sense Disambiguation Algorithm performs in two phases:
Phase 1: Discover the possible meanings (senses) from semantic resources such as Ontologies, Thesaurus, etc. Phase 2: For each keyword select one intended meaning by considering the context.
Obtaining the Possible Keyword Senses of each User Keyword
ISDSI’09 Cagmoli (Italy), 25th June 2009
Consulting a well-known general-pupose shared thesaurus such as WordNet: Advantages: It is fast and provides a reliable set of
senses. Disadvantages: It does not cover with the same detail
different domains of knowledge. Ex: The meaning of developer as “sb who designs and implements software” does not appear.
Consulting the knowledge stored in different pools of ontologies available on the Web and using synonym probability measures to remove redundant interpretations: Advantages: The more ontologies consulted, the more
chances to find the semantics assigned by the user. Disadvantages: It could introduce noise and irrelevant
information.
Obtaining the Possible Keyword Senses of each User Keyword
ISDSI’09 Cagmoli (Italy), 25th June 2009
Option 1: Consulting a well-known general-pupose shared thesaurus such as WordNet.
Option 2: Consulting the knowledge stored in different pools of ontologies and using synonym probability measures to remove redundant interpretations.
The trade-off between the two approaches is not totally clear: Implement both options beginning by the
Wordnet one. Perform experimental evaluation to decide
which approach to consider.
Discovering the semantics of User keywords
Extraction of keyword senses
Disambiguation of keyword senses
Possible keyword senses
Selected senses
ISDSI’09 Cagmoli (Italy), 25th June 2009
Selecting the most probable sense of each User Keyword
Goal: Select the most probable intended meaning for each user keyword.
How: Using Word Sense Disambiguation techniques: Many features can be considered in the context of
written document, but here the process is more complex. No syntax of whole sentences, few keywords (<5), etc.
ISDSI’09 Cagmoli (Italy), 25th June 2009
Selecting the most probable sense of each User Keyword Try to emulate the behaviour of a human by considering
the possible meanings of the rest of keywords: If star appears in the context “Star Hollywood”, then
the most probable intended meaning is “famous actor/actress”.
If star appears in the context “Star Sky”, then the most probable intended meaning is “celestial body”.
The architecture proposed does not depend on a particular Word Sense Disambiguation technique: Probabilistic Word Sense Disambiguation techniques
that combine different algorithms.
OutlineIntroduction.Basic Architecture of the system:
Discovering the Semantics of User Keywords:Obtaining the possible keyword senses (meanings).Selecting the most probable sense of each user keyword.
Semantics-Guided Data Retrieval: Lexical annotations of results of a traditional search. Categorization of results.
Improvements to the Basic Architecture:Conclusions and Future Work.
ISDSI’09 Cagmoli (Italy), 25th June 2009
Semantics-Guided Data Retrieval
Semantics-Guided Data Retrieval
Selection of the most probable intended category
Categorization of hits
Lexical annotation of hits: title and snippet
Search keywords in traditional search engines
Possible keyword senses
Selected senses
Hits (results of a traditional SE)
Annotated Hits by considering the Possible Keyword ss
Cluster or categories of hits
Semantic Cluster of Hits
Keywords
Goal: Select hits relevant for the user and filter irrelevant ones.
Phase 1: Retrieval by using traditional techniques.Phase 2 and 3: Lexical annotations of hits and classification of them by using Word Sense Disambiguation.Phase 4: Selection of the category corresponding to the selected senses.
How:
Goal: Associated to each user keyword that appears in each returned hit (title, URL and snippets) a meaning by considering the possible the meaning of the keyword.
Cleasing each hit to remove stopwords and mark without semantic information.
Performing WSD by considering the context of the words (its neighbour words in a window).
How:
Lexical Annotation of the Results of a Traditional Search Engine
Cleasing of hitsPossible keyword senses
Hits (results of a traditional SE). For each hit title, URL and Snippet
Annotated Hits by considering the Possible Keyword Senses
Lexical Annotation
Lexical Annotation of the Results of a Traditional Search Engine
Cleasing of hits
Possible keyword senses
Hits (results of a traditional SE). For each hit title, URL and Snippet
Annotated Hits by considering the Possible Keyword Senses
Lexical Annotation
Only information from snippets is used to perform the lexical anotation
New senses for words appears but only when they are widespreaded they are integrated in semantic resources
Categorization of the Annotated Results
ISDSI’09 Cagmoli (Italy), 25th June 2009
Hit1(s11, s21), Hit2 (s11, s22), Hit3(s11, s22), Hit4(s11, ?),…
K1 (Hollywood): S11K2 (Star): S21(Celestial body), S22 (Actor/Actres)
C1(S11, S21): Hit1, …C2(S1U, S21): Hit4, ...C3(S11, S22): Hit2, Hit3,... C4(S1U, S22): ...C5(S1U, S2U): ...
Goal: Associated to annotated hit a category. How:
Defining the categories considering the possible keyword senses.
Associated to each hit a category by considering its annotations.
Categorization of the Annotated Results
ISDSI’09 Cagmoli (Italy), 25th June 2009
C1(S11, S21): Hit1, …
C2(S1U, S21): Hit4, ...
C3(S11, S22): Hit2, Hit3,...
C4(S1U, S22): ...
C5(S1U, S2U): ...C6(S11, S2U): ...
Select the category (cluster) that correspond to the selected senses of the user.
The hits of each category are orderd following the ranking return by the search engine.
Unknown Sense for Hollywood
Unknown Sense for star
Problems of Basic Architecture
ISDSI’09 Cagmoli (Italy), 25th June 2009
Problem 1: The system only selects the most probable intended category but the user can be interested in other one.
Problem 2: Sometimes, even for a human it is very difficult to decide which is the meaning which is being used for a word.
Problem 3: The system is not considering the synonyms of the keywords
OutlineIntroduction.Basic Architecture of the system:
Discovering the Semantics of User Keywords. Semantics-Guided Data Retrieval.
Improvements to the Basic Architecture: Probabilistic Word Sense Disambiguation. Retrieval of Synonyms of User Keywords.
Conclusions and Future Work.ISDSI’09 Cagmoli (Italy), 25th June 2009
ISDSI’09 Cagmoli (Italy), 25th June 2009
Probabilistic Word Sense Disambiguation
Show more intrepretations to the user: Instead of only showing to the user the category corresponding to the most probable senses, showing him/her all the categories sorted by considering the probability associated to each category.
C1(S11, S21): Hit1, …
C2(S1U, S21): Hit4, ...
C3(S11, S22): Hit2, Hit3,...
C4(S1U, S22): ...
C5(S1U, S2U): ...C6(S11, S2U): ...
C3(S11, S22): Hit2, Hit3,...
C4(S1U, S22): ...
C1(S11, S21): Hit1, …
C2(S1U, S21): Hit4, ...
C6(S11, S2U): ...C5(S1U, S2U): ...
Probabilistic Word Sense Disambiguation Probabilistic Word Sense Disambiguation:
It is based on a probabilistic combination of different WSD algorithms so the process is not affected by the effectiveness of a single algorithm.
Associate a probability to each lexical annotation that indicates the reliability level of the annotation.
So, each hit will be associated to several categories with a certain probability.
0,750,20
0,05
C3(S11 (hollywod), S22(star)): Hit2, Hit3,...
ISDSI’09 Cagmoli (Italy), 25th June 2009
Retrieval of Synonyms of User Keywords
Probabilistic Word Sense Disambiguation: Associate to each hit the product of the probabilities of
its annotations and use this value to rank the hits clasiffied inside a category (group of cluster).
Enrichment of the clusters with retrieval of synonyms of the senses that represent that category.
Celebrity, actor/actress
OutlineIntroduction.Basic Architecture of the system:
Discovering the Semantics of User Keywords. Semantics-Guided Data Retrieval.
Improvements to the Basic Architecture: Probabilistic Word Sense Disambiguation. Retrieval of Synonyms of User Keywords.
Conclusions and Future Work.ISDSI’09 Cagmoli (Italy), 25th June 2009
ISDSI’09 Cagmoli (Italy), 25th June 2009
Related Work
There exist several techniques for clustering the results of a web search, but most of them based only on statistics techniques.
Some approaches consider semantics, such as:
Hao et al. 2008: Uses only WordNet and assumes a predefined set of categories.
Hemayati et al. 2007: Limited to queries with a single keyword and does not allow overlapping categories.
ISDSI’09 Cagmoli (Italy), 25th June 2009
Conclusions and Future Work We have proposed an architecture to group the results of a standard search engine in different categories:
The categories are defined by the senses of the input keywords.
The system has desirable features in this kind of systems.
Non-popular searches do not remain hidden.
Next steps:
Implementation of the system proposed.
Design a set of experiments with users to evaluate it.
Semantic Access to Data from the WebRaquel Trillo, Laura Po+, Sergio Ilarri*, Sonia Bergamaschi+, E. Mena*
1st International Workshop on Interoperability through Semantic Data and Service Integration ISDSI’09
Cagmoli (Genova), Italy, 25th June 2009
http://sid.cps.unizar.es Univ. Zaragoza
http://www.dbgroup.unimo.it/ Univ. Of Modena e Reggio Emilia
Grazie Mille!Thank you very much!Questions and suggestions.
* +