Upload
melinda-townsend
View
214
Download
0
Tags:
Embed Size (px)
Citation preview
Domain-specific Web Corpora and their Applications
Gregor Erbach
Saarland University
Project COLLATE
(funding: BMBF 01 IN A01 B)
Outline
Part I: Web Corpora
Part II: Applications of Web Corpora
Part III: LT-World Web Corpus
Part IV: Research in COLLATE
Part I: Web Corpora
1. Formal Properties of the Web
2. Web Corpus
3. Document and Hyperlink Database
4. TREC web track
Formal Properties of the Web
• Hypertext/Hypermedia• Directed graph with cycles• Edges = hyperlinks• Nodes = documents ???• Nodes often have internal tree structure (HTML,
XML)
Web Corpus
A web corpus consists of• a database of documents• a database of hyperlinks
Document Database
Information for each document:– URL/URN
– Full Text (possibly with linguistic annotation such as POS, named entities, phrases)
– Full Text Index
– Metadata• Author, Language, Date, MIME type … (Dublin Core)
• Category, Abstract, Keywords, Type of Page …
Fields of Hyperlink Database
• source anchor URL
• source anchor position on web page (percentage)
• source anchor position in document structure (HTML element path)
• source anchor type (text or image)
• source anchor text and context
• target anchor URL
• target anchor position on web page
• target anchor MIME type
Derived Properties of Hyperlinks
• Same document?• Same server?• Same 2nd/3rd level domain?• Ascending of descending in directory structure• Source is within a list of links• Navigation link (up, previous, next …)
TREC web track
• Construction of a web corpus (WT10g) according to the following criteria:
– Broadly representative of web data in general
– Many inter-server links
– Contains all available pages from a set of servers
– Contains an interesting set of meta-data
– Contains few binary, non-English or duplicate documents
– Size: 10 GB
P. Bailey, N. Craswell and D. Hawking. Engineering a multi-purpose test collection for Web retrieval experiments. IP&M, to appear.
Part II: Applications of Web Corpora
1. Web Mining
2. Information Retrieval
3. Clustering and Categorisation
4. Summarisation
5. Discovery of Relations
6. Terminology Extraction
7. Information Extraction
8. Ontology Learning
Useful Methods
• Machine Learning and Data Mining• Natural Language Processing• Information Retrieval• Ontologies and Semantic Web• Bibliometrics (citation analysis ~ link analysis)
Web Mining
• Web Content Mining– Discovery of terminology, acronyms, concepts
• Web Structure Mining– Discovery of relations, communities …
• Web Usage Mining– Discovery of navigation patterns
Information Retrieval
• Usage of hyperlinks for determining popularity of web pages
• Hub and authority pages• Widely used: Google PageRank• Mixed results in TREC web track
Jon M. Kleinberg (1997) Authoritative Sources in a Hyperlinked Environment. Journal of the ACM
Sergey Brin, Lawrence Page (1998) The Anatomy of a Large-Scale Hypertextual Web Search Engine. Computer Networks and ISDN Systems
Clustering
• Standard clustering algorithms form clusters by iteratively grouping documents/clusters, according to a distance measure
• Content-based methods measure distance by counting terms/concepts (often TF/IDF)
• Connectivity-based distance measures make use of hyperlinks
Categorisation
• Categorisation algorithms determine the membership of a document in a pre-defined thematic category
• Content-based categorisation methods measure distance from a representative of the category
• Connectivity-based distance measures are based on the assumption that certain types of hyperlinks lead to documents of the same category
Summarisation / Keyword Extraction
• Source anchor text has been used to generate short summaries of target web pages.
Discovery of Relations
• Hyperlink structure reflects relations between web resources (e.g. between personal homepage, project page, organisation page)
• Relations can be discovered by content-based methods and by connectivity-based methods
Terminology Extraction
• Content-based: extraction of domain terminology by statistical analysis (TF/IDF …) and/or phrasal chunking
• Applicability of connectivity-based methods?
Information Extraction
• Automatic extraction of meta-data• Extraction of named entities for concept-based
indexing• Extraction of templates/relations for relation-based
indexing, and question answering
Ontology Learning
• Extraction of candidates by frequency of occurrence in similar contexts
• Usage of textual clues (“such as”, “sogar” …)• Applicability of connectivity-based methods?
• Definition and acronym mining
Part III: LT-World Web Corpus
1. Content of LT World
2. Ontology
3. Hyperlinking within LT World
4. Construction of the corpus
LT World: Idea and Context
• The virtual information center is a comprehensive WWW-based information and knowledge service for the entire area of language technology.
• LT World is a “virtual” center in the sense that most information will physically remain with their creators or with other service providers.
• The virtual information center has been online since October 2001 under the name „LT World“ for „Language Technology World“ (www.lt-world.org)
Virtual Information Center - LT World
• Information and Knowledge– Technical and Scientific Information
• Players and Teams– Persons, Projects, Organisations
• Resources and Results– Research Systems, Commercial Products
• Communication and Events– News, Conferences
LT World Ontology
Publications
Products Projects People
Layer 2: Specific Ontologies
Corpora etc.
Layer 1: Dublin Core
Layer 3: Ontology for CL & LT
LT World Ontology
• Dimensions– Linguality (monolingual, multilingual, cross-language)
– Application
– Computational/mathematical methods
– Linguistic Models / Theories
– Level of linguistic description/processing
– Technologies
– Language(s)
• Ontology is modelled in RDF with Protégé 2000
LT World: Coverage
• 99 topic nodes• 300 NLP tools and products• 1800 people• 850 organisations• 500 projects
Data Acquisition Process
• Manual collection, categorization and annotation of URLs by students and staff
• Sources: conference proceedings and journals, lists of links on the web,
• Self-registration and correction of data by users of the service
• Technical/scientific information in topic nodes has been provided by domain experts
LT World: Topic Nodes
Topic nodes are the main information unit of the Area “Knowledge and Information”. They are organized in a shallow slightly multidimensional hierarchy following the chapter plan of the second edition of the Language Technology Survey.
Example of the shallow hierarchy:Information Extraction
• Named Entity Recognition
• Terminology Extraction
• Relation Extraction
• Answer Extraction
Information for each Topic
• Name
• Acronyms
• aka‘s, Term Translations
• Short Definition
• Overview Article (from HLT Survey)
• Topic Websites
• R&D Prototypes/Products
• Projects
• People
• Literature
Hyperlinking between Sections
Corpus Construction
• Start from URLs in LT-World collection• Expand document set by recursively following outgoing
hyperlinks using a webspider (e.g., GNU wget)• Expand document set by following incoming hyperlinks
(“link” query to search engine)• Expand document set by search engine queries with
domain terminology• Construct document database and link database• (Filter out irrelevant documents)• Publish Corpus
Part IV: Research Directions
Categorisation / Information Extraction
Discovery of Relations for Hyperlinking
Other
Categorisation and Information Extraction
• Research objectives – find method for categorising documents according to
LT-World ontology
– find method for extraction of meta-information
• Compare and combine content-based and connectivity-based methods
• If successful, it will contribute to semi-automatic extension of the coverage of LT-World
Discovery of Relations
• Objective: develop method for finding pairs of related documents, e.g. personal page – organisation page.
• Content-based and connectivity-based methods are applicable
• If successful, it will enable a significant improvement of LT-World (resource discovery, resource annotation)
Other
• Objective: compare and combine content-based and connectivity-based clustering methods
• Applications:1. Information Retrieval
2. Clustering
3. Summarisation
4. Terminology Extraction
5. Ontology Learning
Conclusion
• Main research interest: comparison and combination of content-based and connectivity-based methods
• Main application impact: going from a set of “seed” web pages to a domain-specific information system