Skip to content

A pipeline approach to automatically extract terminological concept systems from text. We use multilingual neural language models to extract terms and their relations on a an intra-sentence level.

Notifications You must be signed in to change notification settings

Text2TCS/Towards-Learning-Terminological-Concept-Systems

Repository files navigation

Towards Learning Terminological Concept Systems from Multilingual Natural Language Text

Reference

Wachowiak, L., Lang, C., Heinisch, B., & Gromann, D. (2021). Towards Learning Terminological Concept Systems from Multilingual Natural Language Text. In 3rd Conference on Language, Data and Knowledge (LDK 2021). Schloss Dagstuhl-Leibniz-Zentrum für Informatik. Chicago

Try it Out

If you just want to try out the service without using the code provided here you can use our implementation made available on the European Language Grid. However, the implemenation on the European Language Grid utilizes a slightly improved architecture as well as a different dataset for model-training; an updated description will be made available soon.

Architecture

Architecture for extracting terminological concepts systems from natural language. PicArchitecture

Example Output

The resulting terminological concept system is returned in a TBX format as well as connected graph (see below).
PicExampleGraphOutput

Term Extraction Scores

Dataset Precicion Recall F1
TermEval2020 EN 54.9 62.2 58.3
TermEval2020 FR 65.4 51.4 57.6
TermEval2020 NL 67.9 71.7 69.8
ACL RD-TEC Annotator 1 74.4 77.2 75.8
ACL RD-TEC Annotator 2 80.1 79.3 80.0

Relation Extraction Scores

Relation Type Precicion Recall F1
synonymy 0.85 0.76 0.80
activityRelation (e1,e2) 0.93 0.97 0.95
activityRelation (e2,e1) 0.00 0.00 0.00
associativeRelation 0.90 0.92 0.91
causalRelation (e1,e2) 0.90 0.95 0.92
causalRelation (e2,e1) 0.92 0.91 0.91
genericRelation (e1,e2) 0.90 0.93 0.92
genericRelation (e2,e1) 0.46 0.41 0.43
instrumentalRelation (e1,e2) 0.72 0.68 0.70
instrumentalRelation (e2,e1) 0.85 0.88 0.86
none 0.69 0.44 0.54
originationRelation (e1,e2) 0.83 0.89 0.86
originationRelation (e2,e1) 0.84 0.83 0.83
partitiveRelation (e1,e2) 0.90 0.85 0.87
partitiveRelation (e2,e1) 0.77 0.77 0.77
spatialRelation (e1,e2) 0.90 0.91 0.91
spatialRelation (e2,e1) 0.90 0.82 0.86

About

A pipeline approach to automatically extract terminological concept systems from text. We use multilingual neural language models to extract terms and their relations on a an intra-sentence level.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published