Caderige Eventbased Information Extraction for the biomedical domain

1
CaderigeEvent-based Information Extraction for
the biomedical domain

• E. Alphonse, S. Aubin, P. Bessières, G. Bisson,
  T. Hamon, S.Lagarrigue, A. Nazarenko, A.-P.
  Manine, C. Nédellec, M. Ould Abdel Vetah, T.
  Poibeau, D. Weissenbacher
• http//www-leibniz.imag.fr/SICLAD/Caderige/
• http//www-lipn.univ-paris13.fr/poibeau/Extra/

2
Caderige partnership

• NLP, knowledge acquisition and machine learning
• LEIBNIZ laboratory (IMAG, CNRS, Grenoble)
• LIPN laboratory (University Paris 13, CNRS,
  Villetaneuse)
• LRI laboratory (University Paris 11, CNRS, Orsay)
• Biology and Bioinformatics
• AÏDA project (IRISA, INRIA-Rennes)
• MIG laboratory (INRA, Jouy-en-Josas.
  Microbiology)
• ENSAR laboratory (INRA, Rennes. Animal genetics)

3
Caderige aim of the project

• Extract structured information from textual
  databases in genomics (i.e. Medline)

.. the GerE protein inhibits transcription in
vitro of the sigK gene encoding sigmaK ..
4
Extraction rules

• Analyze and normalize the text
• Apply extraction rules to fill the extraction
  template
• GerE stimulates cotD transcription and cotA
  transcription , and,
• unexpectedly, inhibits transcription of the
  gene (sigK)
• interaction(X,Y,Z)-
• is-a(X,protein), is-a(Z,gene), is-a(Y,
  interaction), is-a(U, transcription),
  subject(X,Y), DObj(Y,U), NprepN(U,Z).
• How can these rules be semi-automatically
  acquired from the text?

5
Overview

• Overall approach
• NLP for text normalization
• Ontology learning
• Conclusion Toward extraction rule learning

6
Linguistic analysis example

• Multi-level annotations
• Normalization of the original text

7
IR and IE for textual database access

• Naïve bayes
• P 74
• R 85

8
System architecture (IE engine)
9
Analysis principles

• Provide robust and versatile NLP tools
• Provide domain-specific resources
• Existing certified resources
• Acquisition from the corpus
• Provide automatic method for resource acquisition

10
NLP for text normalization

• Named entity analysis
• Terminology acquisition and filtering
• Focused syntactic analysis

11
Named entity analysis

• Group and normalize names corresponding to
  entities of the domain
• Graphical variations sigma K / sigma(K) / sigma-K
• Morphological variations Down syndrom / Down's
  syndrom
• Syntactic variations human cancer / cancers in
  human
• Semantic variation rat somatotropin, rat growth
  hormon
• Synonymy due to renaming SpoIIIG / sigma G.
• Ellipsis EPO mimetic peptide / EPO
• Abbreviations Bacillus subtilis / B. subtilis
• Acronyms chloramphenicol acetyltransferase / CAT

12
Terminology processing

• Terminology is necessary for
• Sentence filtering
• Syntactic analysis
• Ontology learning
• 2 main sources for terminology
• External validated resources (i.e. MeSH Gene
  ontology)
• Acquisition from a representative corpus (Acabit,
  Daille1995)
• How to automatically filter terms?
• Morpho-syntactic variation Jacquemin 2001
• External certified terminology
• Statistical filtering Daille1995

13
Terminology filtering
Terms in corpus
14
Syntactic analysis

• Syntactic analysis is necessary for
• Ontology acquisition
• Structured information extraction
• Dependency vs constituent based grammars
• Constituent grammars efficient to segment the
  text in syntactic phrases
• But fail to extract relevant functional
  relationships betweens phrases
• Our choice the Link Parser (www.link.cs.cmu.edu/l
  ink/)
• Partial dependency-based analysis
• Efficiency proved on useful syntactic relations
• Linguistic resources accessibility

15
Integration of the Link Parser

• Normalized text input to limit linguistic
  ambiguity
• Revised sentence segmentation
• Integration of named entities (e.g. bacillus
  subtilis)
• Integration of terminology (e.g. in vitro)
• Link parser resource tuning for biology
• Addition of unknown words from the biology domain
• Addition of new rules (e.g. omission of the
  determiner before some nouns that require one)
• Separate evaluation of each kind of dependency

16
Semantic annotation

• Semantic annotation is necessary for
• Ontology acquisition
• Structured information extraction
• Resources for biology text annotation has been
  produced
• An annotation tool called Cadixe
• DTD for the biological domain
• Annotated corpora, mainly on Bacillus Subtilis
  (MIG)
• Easy adaptation to new DTD and new domains
• Cadixe currently used by Swiss Prot to annotate
  other biological corpora (with another DTD)

17
The Cadixe annotation tool
18
Ontology learning

• Concept clustering
• Hierarchical learning
• Evaluation

19
Hierarchy of conceptual cluster learning

• Learning goal conceptual hierarchies where nodes
  are classes of semantically-related terms
• Method conceptual clustering based on
  distributional analysis
• Semantic distance between terms is based on the
  number of common contexts shared, in the training
  corpus.
• Common contexts of two terms
• Co-occurrences in a window or in a document
• Co-occurrences in syntactic contexts (Grishman
  et al., 78, Dagan, 98)

20
Syntactic-based distance

• Corpus parsing
• Syntactic dependencies between heads and their
  arguments
• Learning examples for Asium Faure Nédellec,
  1999

21
Learning conceptual hierarchies by clustering

• Concepts correspond to classes of terms
  co-occurring in different syntactic contexts

22
Clustering results

• Tentative evaluation
• Ontology acquisition tested on a scientific
  corpus (Agrovoc)
• Comparison with other distances (Greedy, Dagan)
• Asium Semantic distance produces less induced
  examples (lower recall) but has a higher
  precision than comparable methods

recall
precision
23
Conclusion toward extraction rules learning

• Summary
• Perspectives

24
Summary on Caderige

• The approach text normalization for information
  extraction
• A two-tier architecture
• Production tool a chain of configurable NLP
  tools
• Acquisition tools corpus-based knowledge
  acquisition, based on ML techniques and text
  normalization
• Current activities
• Complete the NLP integration
• Machine learning to acquire extraction rules

25
Example of learned IE rule

• Annotated sentence
• The ltagent typeproteingt GerE lt/agentgt
  protein ltinteraction typenegativegtinhibits
  lt/interactiongt lttarget typeexpressiongttranscripti
  on of ltsource typegenegtthe sigK genelt/sourcegt
  encoding ltproductgt sigmaKlt/productgt lt/targetgt
• Example of learned rule (Propal Alphonse et al.,
  01)
• interaction (X, Y)-
• isa(X,protein), isa(Y,gene), isa(Z,
  neg_interaction), subject(X,Z), DObj(Z, U).