Unsupervised Information Extraction in the OntoBasis Project

1
Unsupervised Information Extraction in the
OntoBasis Project

• Marie-Laure Reinberger, Walter Daelemans
• CNTS
• University of Antwerp

2
Context

• OntoBasis Foundation, Construction, Services
  and Applications of Ontologies
• Elaboration and adaptation of semantic knowledge
  extraction tools for the building of specific
  domain ontology

3
Method 1
NP1Subject The/DT patients/NNS NP1Subject
VP1 followed/VBD VP1 use wear remove
mask mask, protective eyewear, face mask,
glove mask, protective eyewear during
operation surgery

• raw text
• ? shallow parser
• parsed text
• ? selection
• classes
• ? similarity
• clusters
• ? pattern matching
• linked clusters

4
Corpus

• SwissProt corpus
• 13M words
• Extracted from Medline abstracts
• Protein and genes

5
Shallow Parsing

• Processing of a large amount of text
• Syntactic analysis -gt semantic information
• Detection of the syntactic structure
  subject-verb-object(demo http//ilk.kub.nl)

6
Example
The patients have followed
a healthy diet .
The/DT patients/NNS have/VBP followed/VBD a/DT
healthy/JJ diet/NN ./.
NP The/DT patients/NNS NP
VP have/VBP followed/VBD VP
NP a/DT healthy/JJ diet/NN NP ./.
NP1Subject The/DT patients/NNS NP1Subject
VP1 have/VBP followed/VBD VP1
NP1Object a/DT healthy/JJ diet/NN NP1Object ./.
7
Selection (1)Syntactic structure verb-direct
object

• Pairs main verb-nominal string, the NS being a
  string of adjectives and nouns
• Building of classes amino_acide_sequence
  deduce predict derive compare
  determine arginine change convert
  replace mitosis exit enter leave initiate
  complete

8
Selection (2)Filtering

• Use of a contrastive corpus Wall Street Journal
• Elimination of the NS containing nouns appearing
  in the WSJ corpus
• Exampleadministrationboard

9
Selection(3)Statistical measures

• Baseline Frequency measure
• Probability measure
• Resnik measure the selectional preference value
  of a verb is strong when this verb combines with
  unfrequent nominal strings

10
Clustering of nominal strings

• Similarity between classes of verbs nominal
  strings combining with a set of common verbs
  share semantic features
• Clusters alanine glycine arginine cysteine
  codon antibody monoclonal_antibody pcr primer
  oligonucleotide clue insight information
  new_insight

11
Pattern Matching

• Nominal string Preposition Nominal
  stringblood_vessel_growth on ribonucleolytic_acti
  vityamino_acid_residue in polymeraseprimer from
  amino_acid_sequence
• Trios organized in classesamino_acid_residue in
  polymerase exon region N-terminal_region
  catalysis

12
Combining clustering with pattern matching

• Enlarging clusters and building links between
  clusters
• combination sequence use of
  oligonucleotide_probe polymerase chain_reaction
  probe set prc primer
• combination sequence use gene of
  byoligonucleotide_probe polymerase
  chain_reaction probe set prc primer of on c
  DNA exon basis sequence
  from amino_acid_sequence sequence region

13
Evaluation

• Use of UMLS (Unified Medical Language System)
  thesaurus and semantic network
• Extraction of pairs of terms related in UMLS
• Recall and precision computed according to the
  quantity of UMLS pairs found in the clusters
• Only a part of the relations is evaluated!

14
Results

• Without filtering, high occurrencesRecall 30
  Precision 16
• Terminology filtered Recall PrecisionFrequen
  cy 59 4Resnik 45 9Probability
  47 7

15
Method 2
NP1Subject The/DT patients/NNS NP1Subject
VP1 followed/VBD VP1 protein of
amino_acid RNA synthetize cDNA

• raw text
• ? shallow parser
• parsed text
• ? pattern matching
• general relations
• ? selection
• highly rated relations
• ? pattern matching
• functional relations

16
Shallow parser

• Trained on new terminology including protein
  names, to get rid of SVO constructions such
  as gene murine chromosome expression
  acetyltransferase gene
• same corpus

17
Filtering

• Selection of patternsnominal string1 prep
  nominal string2
• Statistical measures, probability
  ofNS1-PrepPrep-NS2NS1-Prep-NS2
• Selection of the n NS with the best measures

18
Examples

• Nominal string Preposition Nominal
  stringshort_arm of human_chromosomeoligonucleot
  ide_probe from N-terminal_amino_acid_sequencepr
  imer from amino_acid_sequenceapoptosis during
  neuronal_differentiation

19
Next step

• Selection of patterns subject-vb-object-
  containing the NS selected in the previous step
  - rated according to the previous measure and
  the probability of appearance of the verb- use
  of a (basic) stoplist

20
Examples

• Subject Verb Objectprotein encoding genelysine
  replaced glutamic_acidmeiosis expressed geneRNA
  synthetizes cDNA

21

• valine - signal - cleavage - replace -
  alanine - for - isoleucine - for - alanine -
  to - methionine - for - glycine - for -
  leucine
• Edman_degradation - gave - glycineaspartic_acide
  replaces - glycineglycine - to -
  glutamic_acid - to - aspartic-acid

22

• protein show immunoreactivity - show -
  sequence_similarity - bind copper - represses
  transcription - encode gene - of
  amino_acid - onto membrane
• DNA inhibits - protein_synthesisinduction -
  requires RNA requires -

23
Problems

• Possible negations
• Preposition link not always precise enough
  (part_of relations)

24
Evaluation by experts of the domain

• 261 relations 165 functional/ 96 prepositional
• False/irrelevant 118 95/23
• General info/weak relevance 21 12/9
• Specific info/strong relevance 122 58/64
• Precision 55

25
What next

• Sort and precise part_of relations hydrogenase
  of alcaligenescentral_region of chromosome
• and othersphenylalanine to tyrosine ??
• Pattern Verb-prep-objectguanine changing
  arginine ?

26
Possible applications

• Ontology adaptation to the medical/protein domain
• Document selection
• Information extraction