1. GRISP: A Massive Multilingual Terminological Database for Scientific and Technical Domains
Patrice Lopez and Laurent Romary
INRIA & HUB – IDSL

2. Overview
GRISP (Generic Research Insight in Scientific and technical Publications)
Multiple scientific and technical fields
Multilingual (en, fr, de)
Built from the compilation of open resources
Sound conceptual model
Mapping across a variety of domains
Use of structural constraints
Machine learning techniques for controlling the fusion process
Our sources: MeSH, UMLS, Specialist Lexicon, Gene Ontology, ChEBI, WordNet, WOLF, SUMO, IPC, Wikipedia
Result: several millions terms, concepts, semantic relations and definitions.

3. Why are we doing all this?
Terminology is the main vehicle by which technical and scientific units of knowledge are represented and conveyed (30-80%; Ahmad, 1996)
Application to a large collection of multilingual and multi-domain patent documents
Two underlying considerations:
Cost of manually maintained terminological resources
Cf. Biosis, IATE, TermScience
Khayari et al., 2006: Modeling the heterogeneity of resources
A lot of available resources online, based on heterogeneous organizational principles
Underlying vision: Integrating knowledge engineering into current state of the art information retrieval and classification systems

4. Merging terminological resources
Related to the fusion of ontologies
Ontologies are usually relatively small in size
Semi-automatic methods: McGuinness et al., 2000
Fully automatic method
Madhavan et al., 2001: exploit structural and linguistic matching
Doan et al., 2001: Machine learning techniques (concepts and properties)
Gal et al., 2005: fuzzy logic methods
Existing work on merging classification systems
Wang et al., 2008: Merging of subject headers in Digital Libraries
Automatic merging techniques for heterogeneous terminologies has not been yet investigated
Much richer linguistic content
No formal organization of concepts
Do not model facts or assertions

5. A quick reminder Terminological resources
Approximation of lexical semantics in specialized fields
Based on a concept to term (onomasiological) model
Naturally multilingual (term grouping according to languages)
Existing standards
ISO 704: editorial principles for building up a terminological resource
ISO 16642: Abstract model for representing terminological databases
Romary, 2001
ISO 30042: A concrete XML syntax (TBX)
Note: terminology standards do not standardize terminologies!

6. Target terminological model
Multiple languages
Multiple terms
Variants, abbreviation, inflexions
Multiple descriptions
E.g. multiple definitions, complementing each other
Additional information: illustrations, formulae, etc.
Basic conceptual relations
Local metadata
Provides management information attached to the various terminological description levels (e.g. origin, validation level, register)
Allows the creation of views (e.g. all MeSH entries; cf. Khayari et al., 2006)
And yes, ISO (TMF) can all this!
Main issue: identifying the relevant data category in the various source terminologies

7. Merging terminologies, merging models
TMF model 1
TMF model 2
Target model
TMF model 2
TMF model 2

8. TMF in a nutshell
Metadata (sources, revisions)
Ontological relations, definition
Terminological Data Collection (TDC)
Terminological Entry
Language Section
Term Section
Terminological Entry
Language Section
Term Section
Dialectal information, definition
Terminological Entry
Language Section
Term Section
Grammatical information, register, … definition
Terminological Entry
Language Section
Term Section
+ any kind of local metadata (origin, certainty, accessibility)

9. Merging terminologies, merging models
/definition/
TMF model 1
Data category mapping
TMF model 2
/definition/
Target model
TMF model 2
TMF model 2

10. Identifying domains Theoretical background GRISP
Non-ambiguity of a term within a domain
E.g. 129 domains in MESH
GRISP
Set of 76 reference domains (see table 1)
Scientific and technical domains of Wordnet Domains (Magnini and Cavaglià, 2000)
Organised as a hierarchy
Manual mapping from resource specific domains to our reference set

11. Merging concepts
Identification of common concepts across terminological sources – core principles
Baseline: same term + same domain = same concept
Difficulties: Conflicting domain mapping, high polysemy of term variants and incorrectly positioned concepts (e.g. Wikipedia)
Wrongly merged concepts
Lost in precision for concept description
Revised: same preferred term + same domain = same concept
Source conformance rule: separated concepts in a given source cannot be further merged (by transitivity)
Not applied to Wordnet, IPC and Wikipedia
Smoothing down the rules: using machine learning techniques

12. Concept merging as a machine learning process
Concept pool
Concept
Concept
Concept
Concept
Concept
Concept
Concept
Concept
Concept
Concept
Concept
Concept
Features
Merging decision
SVM (Support Vector Machine) and MLP (Multi-Layer Perceptron) binary classification models

13. Training process Training features Training data
(f1-2) sources (e.g. S1=“MeSH”, S2=“Wikipedia”)
(f3) Number of common domains between the two concepts
(f4) Number of same source-specific categorizations
(f5) Boolean indicating if both preferred terms are identical
(f6) Boolean indicating if both preferred terms are identical after stemming
(f7) Ratio of identical terms given all terms
(f8) Similarity measure of the definition texts, after stemming and based on negative KL divergence
(f9) Number of domains of the merged concept
(f10) Number of words of the longest common terms
Training data
Wikipedia – MeSH mapping
Pascal database (INIST)

14. Result overview Observations:
Merger
Concepts
Terms
Sem. Rel.
Aggregation
1,503,818
3,140,726
970,864
Merg. Rule 1
1,457,538
3,157,179
1,022,303
Merg. Rule 2
1,476,508
3,114,711
971,218
SVM
1,450,688
3,195,118
1,088,446
MLP
1,451,710
3,192,325
1,081,955
Overall content:
596,865 definitions
1,321,988 source specific categorizations of concepts
20,000 acronyms
14,268 chemical formulas and
12,375 chemical structure identifiers.
Observations:
Small number of actual merges (cf. product names, chemical and medical entities)
Merging relevant for frequently used concepts

15. Evaluation Merger Wiki/MeSH PASCAL Merging Rule 1 cov. 0.6464
acc
cov
acc
Merging Rule 2
cov
acc
cov
acc
SVM
cov
acc
cov
acc
MLP
cov
acc
cov
acc
Random subset of 10% of the merging examples extracted from Wikipedia/MeSH mappings and from the PASCAL terminology
Merging Rule 2 produces almost perfect merging but with a very low coverage
Rule 1 extends the coverage at the price of a relatively high rate of merging error
Machine Learning approaches further extend the coverage while maintaining a high precision

16. GRISP browser: radial engine
rendering
rendering
rendering

18. Application: Patatras
PATATRAS (PATent and Article Tracking, Retrieval and AnalysiS)
Context: CLEF-IP competition
Prior art search task (EPO documents)
1,9 million documents in English, French and German (more than 3 billion words)
Ranked first for all subtasks of the evaluation track among 14 participants (Roda et al., 2009)
Conceptual indexing of the CLEF-IP corpus
Development of a term annotator based on GRISP
Term variant matching after POS + lemmatization
Concept disambiguation based on IPC classes of the documents
1.1 million different terms identified
176 million annotations

19. Results: Patatras Significant accuracy improvements for CLEF-IP
Combination of a word-based and concept-based ranked results with a regression model
Based on 10,000 queries

20. Epilogue Online tool Free resource Constant evolution
Contact:
Free resource
Based on the freely available subset of resources
Constant evolution
Maintenance according to evolution of our sources
Addition of further sources