ENACT - Adapter les modèles de traitement automatique des langues et les modèles multimodaux pour aider les apprenant(e)s de langues à comprendre les expressions polylexicales

Keywords

Large Language Models, Small Language Models, multiword expressions, L2 language learners, text-to-image generation, definition simplification and rephrasing

Subject details

The heterogeneity of levels of learners is a challenge for the teachers in terms of personalized instruction and assessment. Natural language processing (NLP) has demonstrated significant potential for enhancing language learning, particularly within the field of Computer-Assisted Language Learning. The PhD thesis will explore and evaluate computational methods for vocabulary acquisition of Multiword Expressions (MWE) in a reading context. MWEs are sequences of several words functioning as single units (Constant et al., 2017). Their lexical, syntactic, semantic, and pragmatic idiosyncrasies (Savary et al., 2021; Kim & Baldwin, 2010) make them challenging for L2 learners. To enhance language learning, MWEs should be explained and defined to learners as their meaning cannot be fully derived from their parts: e.g. donner sa langue au chat (FR), lit. 'give her/his tong to the cat' trans. 'give up', rire jaune (FR) lit. 'laugh yellow' trans. 'hollow laugh'. NLP studies related to personnalized vocabulary acquisition have examined computational methods for measuring text readability (Wilkens et al 2022) and word complexity (North et al. 2023) or for creating targeted exercises and learning materials, particularly graded lexicons aligned with the Common European Framework of Reference for Languages (CEFR), e.g. Tack et al (2016). Regarding MWEs, many theoretical works in phraseodidactics (González-Rey, 2019) have been published. Notably, some pedagogical methods propose to provide images to explain MWEs (Kühn, 1996) or definitions and paraphrases (Milicevic & Hamel, 2007). Nonetheless, only few graded MWE resources and tools exist for language learning purposes, such as PhraseoText (González-Rey, 2019), EWA (Santos 2023). In addition, LLMs and chatbots are revolutionizing language learning with adaptive dialogues (e.g. Tyen et al., 2022), but support for non-English languages is still limited. The PhD candidate will develop and evaluate methods to automate the process of creating such MWE-based resources in a dynamic way according to the learner's profile (e.g. CEFR level) and reading context, by combining textual and multimodal LLMs (Fu, et al, 2025) in a multilingual approach. With the help of textual LLMs, one can generate explanations and definitions of MWEs adapted for the learner's level to support vocabulary acquisition. For instance, the research may build on a paraphrase dataset that is available for PARSEME 2.0 Multilingual Shared Task. To illustrate the sense of the MWEs, one can also generate or select images in the specific textual context, based on the resulting explanations and definitions. This study could rely on several datasets that are available to relate static images with specific MWE senses (Pickard et al, 2025) within the Semeval shared task AdMIRe (Advancing Multimodal Idiomaticity Representation). One important challenge will be to adapt explanations and definitions to the learner's level and therefore to automatically identify their complexity for rephrasing purposes as well as the CEFR level of the MWE for selection purposes, e.g. relying on supervised methods and LLMs (Kelious et al, 2024). To improve LLM performances, one can provide the learner's profile, and static graded lexical resources for several languages (e.g. NT2Lex (Tack et al, 2016) for Dutch, SVALex (François et al, 2016) for Swedish , PolylexFLE (Todirascu et al, 2024) for French). The Ph.D. candidate will develop and evaluate methods to aggregate and integrate these diverse multilingual information sources to enhance the performance and accuracy of LLMs. Alongside LLMs, the PhD candidate shall investigate the use of Small Language Models (SLM) in order to create dynamic resources accessible for language learners which could be used on laptops with smaller ecological cost. Tested methods could include fine-tuning, multilingual transfer learning, ensemble methods, multi-agent strategies.