Skip to Main content Skip to Navigation
New interface
Theses

Exploration des relations terminologiques entre les termes multi-mots dans les modèles de sémantique distributionnelle

Abstract : A term is a lexical unit with specialized meaning in a particular domain. The organization of terms reflects the structure of domain knowledge, which is based on the relationships between domain concepts, i.e., between single terms (STs) or multi-word terms (MWTs). Structured terminology resources, such as specialized dictionaries, terminology banks, and databases, are designed to meet the needs of research, translation, technical writing, etc. However, the relationships between MWTs are often underrepresented. On the other hand, we have seen the massive and successful use of distributional semantic models (DSMs) in many semantics works over the last few years. Our research problem is on the intersection of these two themes. It can be broken down into two questions: is it possible to capture relational information between MWTs using DSMs? What are the relationships between MWTs that these models can best identify? We conducted a set of experiments in the French environment domain to answer these two questions. Our general working hypothesis is that static and contextual DSMs allow us to identify terminological relations between MWTs. We explored the possibilities of these models using two methods. The first one adopting lexical substitution is based on the predictions of a masked language model (MLM). The second one captures lexico-semantic relations by analogy between term representations generated by a FastText model. These methods are tested on two datasets. First, we rely on a dataset composed of French synonymous MWTs of the environment domain provided by the IATE database. Due to the lack of resources integrating various semantic relations between MWTs, we built a second dataset by semantic projection from the single terms and their relations listed in the dictionary of environmental terms, DiCoEnviro. A manual annotation using the contexts of the MWTs is performed to check the preservation of the inferred relationship between the MWTs. The experimental results we obtained validate our hypothesis regarding the possibility of capturing terminological relations between MWTs by DSMs. Analogy involving a FastText model performs better than lexical substitution performed with an MLM in capturing synonymy, antonymy, and hyponymy. The results also show that the conditioning strategy results in MLM predictions that are more closely related to the masked word and that the performance of analogy is improved when variants of MWTs are treated as occurrences of MWTs. Another notable result is that the semantic composition of MWTs is modelled by analogy and captured partially by MLM. This result confirms the advantage of combining distributional and compositional approaches for identifying semantic relations between MWTs. The best results were obtained with the analogy, with an MRR of 0.793 for synonymy, 0.720 for antonymy, 0.613 for hypernymy, and 0.579 for hyponymy. Overall, this thesis is one of the first attempts to identify lexical relations between MWTs in a specialized domain, that of the environment, by exploring DSMs. We have constructed a dataset of MWTs linked by various lexical relationships, which also made available for further research purpose. Ideally, this work is hope to provide a roadmap for applying DSMs for the terminology structuring task.
Complete list of metadata

https://theses.hal.science/tel-03835888
Contributor : Yizhe WANG Connect in order to contact the contributor
Submitted on : Tuesday, November 1, 2022 - 1:17:35 PM
Last modification on : Wednesday, December 7, 2022 - 3:16:25 AM

File

thesis_Yizhe_redaction-6.pdf
Files produced by the author(s)

Licence


Distributed under a Creative Commons Attribution 4.0 International License

Identifiers

  • HAL Id : tel-03835888, version 1

Collections

Citation

Yizhe Wang. Exploration des relations terminologiques entre les termes multi-mots dans les modèles de sémantique distributionnelle. Linguistique. Université Toulouse - Jean Jaurès, 2022. Français. ⟨NNT : ⟩. ⟨tel-03835888⟩

Share

Metrics

Record views

0

Files downloads

0