J. Turmo, 2006 Adaptive Information Extraction Summary • Information Extraction Systems • Multilinguality • Introduction • Language guessers • Machine Translators • Translingual architectures • Information integration in MIE systems • Evaluation • Adaptability
35
Embed
J. Turmo, 2006 Adaptive Information Extraction Summary Information Extraction Systems Multilinguality Introduction Language guessers Machine Translators.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
J. Turmo, 2006 Adaptive Information Extraction
SummarySummary
• Information Extraction Systems
• Multilinguality• Introduction
• Language guessers
• Machine Translators
• Translingual architectures
• Information integration in MIE systems
• Evaluation
• Adaptability
• Information Extraction Systems
• Multilinguality• Introduction
• Language guessers
• Machine Translators
• Translingual architectures
• Information integration in MIE systems
• Evaluation
• Adaptability
J. Turmo, 2006 Adaptive Information Extraction
IntroductionIntroductionMultilinguality
• Multilingual IE (MIE) tasks:
The textual information contained in the output templates is wanted to be presented in a different language than the input documents
• Tipically: • input documents written in one language • output templates written in another one
J. Turmo, 2006 Adaptive Information Extraction
IntroductionIntroductionMultilinguality
• Relatively little research in MIE• LRE program in Europe
• ECRAN, FACILE, AVENTINUS, SPARKLE, …• tools and components for IE in different languages
• TIDES program in USA• PROTEUS, RIPTIDES, CREST, …• fast machine translation and information access
J. Turmo, 2006 Adaptive Information Extraction
• Up to now Multilingual IE evaluation just for NE tasks. Two recent scenarios:
• CoNLL 2002-2003:• Language-independent NE recognition
• ACE 2007: • Arabic input documents• English output NE mentions
• Fei Huang (2005). Multilingual NE Extraction and Translation from text and speech. PhD. Thesis
IntroductionIntroductionMultilinguality
Open researchline
J. Turmo, 2006 Adaptive Information Extraction
IntroductionIntroductionMultilinguality
• Basic elements of MIE architectures:• language guessers• monolingual architectures
• Classical approches:• use of Machine Translation with monolingual IE architectures• extension of monolingual architectures to translingual architectures
J. Turmo, 2006 Adaptive Information Extraction
IntroductionIntroductionMultilinguality
• Basic elements of MIE architectures:• language guessers• monolingual architectures
• Classical approches:• use of Machine Translation with monolingual IE architectures• extension of monolingual architectures to translingual architectures
J. Turmo, 2006 Adaptive Information Extraction
SummarySummary
• Information Extraction Systems
• Multilinguality• Introduction
• Language guessers
• Machine translators
• Translingual architectures
• Information integration in MIE systems
• Evaluation
• Adaptability
• Information Extraction Systems
• Multilinguality• Introduction
• Language guessers
• Machine translators
• Translingual architectures
• Information integration in MIE systems
• Evaluation
• Adaptability
J. Turmo, 2006 Adaptive Information Extraction
Language guessersLanguage guessersMultilinguality
• Goal: identify the language of a document
• Linguistic approach:• based on a vocabulary of keywords• idea: at least one word from a tipical sentence written in some language should be included in the corresponding vocabulary• manually built
J. Turmo, 2006 Adaptive Information Extraction
Language guessersLanguage guessersMultilinguality
• Stochastic approach:• most widely used• based on:
• generate a frequency table of elements per language• compare frequencies of elements in the document with those in the table.• elements = or special characters or word sequences or char sequences(different approaches)
J. Turmo, 2006 Adaptive Information Extraction
Language guessersLanguage guessersMultilinguality
• Stochastic approach:• Pros: good results (over 95% accuracy)• Cons: short texts [Zhdanova,02] copes with this problem
• Try to overcome the ineficiency the MIE architectures based on MT• Fussion of IE and interlingua MT
• Idea: when dealing with a particular domain, it is possible to build a language-independent conceptual model of the particular scenario of extraction [Gaizauskas et al. 97]
• For each source language requires:• Use of different lexical preprocessors • Use of different syntactico-semantic parsing • Use of different sets of IE patterns (if the MIE system is based on pattern matching)
• Possible use of language-independent processors (e.g., NERC)
• M-TURBIO system [Turmo et. al 99]• EuroWordNet (EWN)• Sets of IE-patterns for each source language• Mappings from IE-patterns to ILIs in EWN• Add a new source language, involves
• Add new IE-patterns • Add new tagger and parser• …
J. Turmo, 2006 Adaptive Information Extraction
SummarySummary
• Information Extraction Systems
• Multilinguality• Introduction
• Language guessers
• Machine Translators
• Translingual architectures
• Information integration in MIE systems
• Evaluation
• Adaptability
• Information Extraction Systems
• Multilinguality• Introduction
• Language guessers
• Machine Translators
• Translingual architectures
• Information integration in MIE systems
• Evaluation
• Adaptability
J. Turmo, 2006 Adaptive Information Extraction
Information Integration in MIEsInformation Integration in MIEsMultilinguality
• The most general architecture• Input documents in different source languages not aligned• Output templates in different target languages
• Possible approaches:• MIE system + II system• MIE/II system
J. Turmo, 2006 Adaptive Information Extraction
Information Integration in MIEsInformation Integration in MIEsMultilinguality
• Pros:• Versatil• An instance can occur just in one document written in a specific language.• Can be easier to extract an instance expressed in one language than another
• better processors or resources
• Cons:• Problems inherent to II
• inconsistent values, similar values, generalizations, …
J. Turmo, 2006 Adaptive Information Extraction
SummarySummary
• Information Extraction Systems
• Multilinguality
• Evaluation• Introduction
• Metrics
• Data sets
• Adaptability
• Information Extraction Systems
• Multilinguality
• Evaluation• Introduction
• Metrics
• Data sets
• Adaptability
J. Turmo, 2006 Adaptive Information Extraction
IntroductionIntroductionEvaluation
• The evaluation of the performance of an IE system depends on different factors:
• The IE task: domain, language, document style, …
• The user needs: software use, human use, just some clues about the relevant facts, the context in which they occur, …
What does correctly extracted means?What are the right metrics?What are the best data sets?
J. Turmo, 2006 Adaptive Information Extraction
IntroductionIntroductionEvaluation
The president of ALP in Spain will leave his job tomorrow night
NP NP
The president of ALP in Spain will leave his job tomorrow night
NP
Exact extraction
?
The president of ALP in Spain will leave his job tomorrow night
NP
The president of ALP in Spain will leave his job tomorrow night
NP
Exact extraction
?
J. Turmo, 2006 Adaptive Information Extraction
SummarySummary
• Information Extraction Systems
• Multilinguality
• Evaluation• Introduction
• Metrics
• Data sets
• Adaptability
• Information Extraction Systems
• Multilinguality
• Evaluation• Introduction
• Metrics
• Data sets
• Adaptability
J. Turmo, 2006 Adaptive Information Extraction
MetricsMetricsEvaluation
• Different evaluation frameworks with different points of view of what is correctly extracted:
• PASCAL: • correct = exact extraction• Same metrics as in MUC6
• ACE:• correct = partial extraction (more sophisticated than MUC)
J. Turmo, 2006 Adaptive Information Extraction
MetricsMetricsEvaluation
ACE metric
Idea: How well match the information extracted by a system with that of the reference model?
• Given a system output, s, and a reference model, m, find the global optimum of function Value(s,m) that maximizes the matchings between instances in s and instances in m