OntoBasis Project 4th General Meeting July 7, 2003

1
OntoBasis Project4th General MeetingJuly 7,
2003

• CNTS Universiteit Antwerpen
• Marie-Laure Reinberger
• Walter Daelemans

2
Outline

• Task 1.1 - Adapting text analysis tools for lexon
  extraction
• Task 1.2 Evaluation of ontologies
• Task 1.3 Adaptation of ontologies
• Task 2.1 Overview of existing ontology (meta-)
  models
• Prospects

3
Approach

• Retrieving semantic information using
• syntactic structures
• unsupervised methods
• domain specific corpora

4
Data two corpora extracted from Medline

• Key word hepatitis A hepatitis B
• Size ? 4 million words
• Very specific

• Key word blood
• Size ? 7 million words
• More general

5
Improving the syntactic analysis

• Increasing the shallow parser lexicon with
  medical terms
• Training the parser on specific sentences taken
  from Medline

6
Clustering

• Unsupervised method that consists of
  partitioning a set of objects into groups or
  clusters, depending on the similarity between
  those objects

7
Pattern Matching

• Pattern matching consists of finding patterns in
  texts that induce a relation between words, and
  generalizing this pattern to build relations
  between concepts

8
Methods

• Apply pattern matching and cluster the results
• Apply pattern matching and combine the results
  with clusters obtained with another method
• Subject Phrase Verb Object Phrase
• Noun Phrase Preposition Noun Phrase
• Noun Phrases or Terms
• Nouns

9
Evaluation

• Former evaluation Wordnetmissing words and
  termsmissing relations
• UMLS devoted to the medical domainmore
  accuratestill missing relations
  (functional)Example collect draw blood
  specimen sampleRelation blood
  specimen blood sample

10
Clusters of terms, hepatitis corpus before and
after improvement of the SP
11

• But!
• Before evaluation of 148 pairs, 152 UMLS pairs
  formed
• After evaluation of 454 pairs, 381 UMLS pairs
  formed ?more reliable
• And
• Object relations being more specific ? more UMLS
  pairs ? more reliable results

12
Clusters of terms
13
Examples

• Hepatitis ? liver_transplantation
  transplantation orthotopic_liver_transplantation
  ? immunoadsorbent immunosorbent immunoassay
  immunospot immunosorbent_assay
• Blood ? day h month hour min ? Banker
  den_Hollander Knudsen Tanner

14
Method

• Clusters C of terms formed according to
  similarity between classes of co-occurring verbs
• Pattern Matching T1 Prep T2
• Cluster Ci associated to a preposition Prepi and
  a set of terms
• ?Relations Ci Prep Cj

15
Examples

• dose injection vaccination of
  hepatitis_B_vaccine HBV_vaccine vaccine
• recurrence transmission of infection
  hepatitis_B_virus viral_infection HCV hepatitis_B
  HCV_infection disease HBV HBV_infection
  viral_hepatitis
• heparin blood_pressure blood blood_loss during
  aortic_surgery operation apostosis surgery
  coronary_angiography hemipathectomy
  coronary_artery_bypass emergency-surgery
  cardiac_surgery surgical_resection hemodialysis
  procedure dialysis transplantation
• level sign level of platelet molecule
  interleukin cell t_cell CD8 white_blood_cell
  red_blood-cell macrophage risk

16
Patterns hepatitis vs blood
17
Outline

• Task 1.1 - Adapting text analysis tools for lexon
  extraction
• Task 1.2 Evaluation of ontologies
• Task 1.3 Adaptation of ontologies
• Task 2.1 Overview of existing ontology (meta-)
  models
• Prospects

18
Evaluation of UMLS

• Pairs of terms not found or not associated in
  UMLS
• Relation checked on Google, using the mutual
  information measure
• Discovery of missing terms and relations

19
Results

• Transcriptase activity - transcriptase inhibitor
• Transcriptase activity - transcriptase polymerase
  chain reaction
• Transcriptase activity - transcription
• Transcriptase activity - transcription polymerase
  chain reaction
• Transcriptase inhibitor - transcriptase
  polymerase chain reaction
• Mask - protective eyewear
• Face mask - protective eyewear
• Glove - protective eyewear

20
Outline

• Task 1.1 - Adapting text analysis tools for lexon
  extraction
• Task 1.2 Evaluation of ontologies
• Task 1.3 Adaptation of ontologies
• Task 2.1 Overview of existing ontology (meta-)
  models
• Prospects

21
Adaptation of WordNet

• red_blood_cell is_a red_cell is-a cell
• cell hepatoma_cell human_hepatoma_Hep3B_cell
  woodchuck_hepatocyte normal_human_hepatocyte in
  liver
• renal_failure liver-disase liver_disease_secondary
  renal_disease is_a disorder IDC
• patient therapy liver_transplantation
  transplantation for IDC renal_failure
  liver_disease liver_disease_secondary
  renal_disease
• use of mask glove protective_eyewear is_a
  protection with patient subject child adult
  infant person
• face_mask is_a mask
• contamination of mask
• surgical_face_mask is_a face_mask

• red_blood_cell is_a cell
• renal_failure is_a disorder
• face_mask is_a mask
• mask is_a protection

22
Adaptation of Wordnet

• recurrence transmission of HCV hepatitis_B
  HBV hepatitis_A HBV_infection hepatitis_B_virus_in
  fection hepatitis_B_infection HBsAg
  viral_hepatitis infection hepatitis_B_virus
  viral_infection hepatitis_C_virus HCV_infection
  disease virus
• patient with hepatitis hepatitis_A
  hepatitis_B acute_hepatitis hepatitis_type_A
  acute_viral_hepatitis type_A_hepatitis
  type_B_hepatitis hepatitis_B_virus
  infectious_hepatitis chronic_active_hepatitis
  cholestatic_hepatitis chronic_viral_hepatitis_B
  chronic_viral_hepatitis_C post-transfusion_hepatit
  is

23
Outline

• Task 1.1 - Adapting text analysis tools for lexon
  extraction
• Task 1.2 Evaluation of ontologies
• Task 1.3 Adaptation of ontologies
• Task 2.1 Overview of existing ontology (meta-)
  models
• Prospects

24
Ontology meta- models

• A meta-model will determine the level of
  interoperability that can be reached between the
  different applications.
• Different issues
• Ontology growing
• Ontology storing
• Adaptation of ontologies
• Reuse of ontologies

25
Combination of ontologies

• Find to what extent they overlap
• Link equivalent concepts
• Respect consistency, coherence, non-redundancy
• Take care ofsyntactic mismatchessemantic
  mismatchesconceptualization mismatches

26
Systems studied

• On-To-Knowledge and Sesame
• KAON, Text-To-Onto system
• Protégé-2000 (Stanford)
• VOID, the Kactus toolkit
• WebOnto (Open University, England)

27
Prospects

• Improving the building of relations between terms
• Evaluation of UMLS find evidences of anomalous
  relations
• Adaptation of WordNet generalization on the
  whole thesaurus