Multilingual Production Principles of Lexical Collocations

Informazioni relative al progetto

MaPPLexiC

ID dell’accordo di sovvenzione: 101201497

DOI

10.3030/101201497

Data della firma CE 1 Settembre 2025

Data di avvio 1 Novembre 2025

Data di completamento 31 Ottobre 2030

Finanziato da

• European Research Council (ERC)

Costo totale

Nessun dato

Contributo UE

€ 2 488 590,00

Coordinato da

UNIVERSIDAD POMPEU FABRA
Spain

Periodic Reporting for period 1 - MaPPLexiC (Multilingual Production Principles of Lexical Collocations)

Periodo di rendicontazione: 2025-11-01 al 2026-04-30

Sintesi del contesto e degli obiettivi generali del progetto

The language of native speakers is idiosyncratic. Thus, we know that the early bird gets the worm, we check for the best before date of the grocery we purchase, we jump the line, we face the challenge, and we are sometimes deeply sorry. The MaPPLexiC project focuses on one type of idiosyncratic constructions, which stands out in terms of its frequency of use and as a measure of language proficiency of a learner: (lexical) collocations. The above cited jump [the] line, face [the] challenge, and deeply sorry are examples of them. According to the definition common in lexicography, a collocation is a semi-compositional idiosyncratic co-occurrence of two lexical items that form a syntactico-semantic pattern, where the occurrence of one of them (the “base”) is freely chosen by the Speaker, while the other (the “collocate”) is predetermined by the base and conditioned by the meaning that the Speaker intends to express by the collocate. In face [the] challenge, challenge is the base, face the collocate, ‘deal with’ the intended meaning, and the base is the Direct Grammatical Object of the collocate; in deeply sorry, sorry is the base, the intended meaning is ‘intense’ and deeply is the collocate, which is a modifier of the base.
Collocations are often referred to as “prefabricated phrases”, but the fact that they are prefabricated does not mean that their construction is ad hoc, i.e. not guided by covert
semantic, cognitive, contextual, stylistic and other criteria. Small scale studies show that bases that share primary semantic features also tend to share collocates. There is also evidence from cognitive and socio-cultural studies that collocation creation follows covert, but well-defined criteria. However, the current state of the art still lacks a systematic account of how collocations are constructed across languages. This has far-reaching consequences for the theory of language and for its application alike.
The theoretical consequence is that the claim, advanced by a substantial number of modern linguistic theories, that there is no principled boundary between grammar and the lexicon has so far been supported mainly by evidence from regular syntactic structures involving collocations. It has not yet been complemented by a dynamic grammar of collocation production.The practical consequence is that language learners are still largely required to memorize collocations or look them up in dictionaries, without access to tools that would help them understand how such expressions are constructed.

MaPPLexiC sets out to derive and make explicit the production principles underlying the creation of collocations on a large scale, drawing upon neural large language model (LLM)
technologies. “Collocation production principles” are in this context inherent rules of the language system that determine which semantic, lexical, contextual, and other features of the base words and the surrounding discourse prompt the selection of specific collocates to convey the intended meaning.

In more specific terms, MaPPLexiC targets the following objectives:

O1: Retrieve from the internal neural representations of the LLM human interpretable features that are decisive for the collocation status and collocation type of a given word
combination.
O2: Develop a model for distilling collocation production principles for common collocation patterns from the obtained human interpretable features.
O3: Carry out language-contrastive studies and explore cross-language transfer of collocation production principles, to account for less resourced languages.
O4: Build a test workbench and test extensively the collocation production principles and their cross-language transfer.

The collocation production principles should eventually cover all collocation patterns. However, due to feasibility considerations, MaPPLexiC starts with the 20 most recurrent
patterns in the available language data. To facilitate cross-language studies, MaPPLexiC works with pairs of Germanic (English, German), Romance (French and Spanish), Finno-Ugric (Finnish and Hungarian), and Slavic (Czech and Russian) languages.

Lavoro eseguito dall’inizio del progetto fino alla fine del periodo coperto dalla relazione e principali risultati finora ottenuti

During the reporting period (the first six months of the project), the work on the project focused on WPs 1 and 2, and, in particular, on Activities 1.1 (collocation list compilation), 1.2 (Corpus acquisition and preprocessing), 1.3 (DNN property vector derivation), and 2.2 (Discovery of covert features).

In the context of Activity 1.1 the list of English collocations available prior to the project has been extended and validated and initial lists have been compiled for all languages of the project. Further extension is in progress. For English, extensive experiments have been carried out on collocation generation in interaction with a Large Language Model (LLM). The experiments resulted in comprehensive definitions of 11 different collocation patterns (lexical functions), which were used to retrieve new collocation samples. The outcome of this research has been submitted to the International Journal of Lexicography (currently under review).

In the context of Activity 1.2 a minimal pair (light verb constructions vs. full-verb constructions) seed corpus has been generated for English, and its use in the exploration of the representation of collocational information in LLMs has been demonstrated in experiments. The insights of this preliminary research has been submitted to the International Conference on Empirical Methods in Natural Language Processing (currently under review).

In the context of Activity 1.3 work has been carried out on the improvement of collocation extraction and classification models. A framework for an LLM-based multiple agent team has been designed and implemented, which simulates the dedicated tasks of the members of a team of lexicographers compiling a collocation dictionary. At the current stage, the framework is evaluated and fine-tuned. Preliminary experiments show that it is able to provide fine-grained collocation patterns (in terms of lexical functions) in English of high quality. A journal submission on the results of this research is currently in preparation.

In the context of Activity 2.2 experiments have been carried out on clustering of hidden representations of different collocation patterns in order to assess whether at which layers state-of-the-art LLMs separate collocational meanings. Furthermore, experiments on the extraction and lexicalization of collocation features using Sparse Autoencoder models have been carried out. A conference submission on the outcome of this research is currently in preparation.

Progressi oltre lo stato dell’arte e potenziale impatto previsto (incluso l’impatto socioeconomico e le implicazioni sociali più ampie del progetto fino ad ora)

The research carried out during the first six months of the project already led to results beyond the state of the art. This concerns, more specifically:

- Extensive multilingual collocation lists, classified in terms of the fine-grained taxonomy of lexical functions. These lists are unique and constitute a valuable resource for the research community.

- Minimal pairs corpus on English containing contrastive light-verb vs. full-verb constructions. This corpus will allow deep analyses of the broadly-studied linguistic phenomenon of light-verb constructions, which were not possible so far.

- LLM-based multiple agent framework for collocation extraction and classification. The framework provides a flexible instrument for compilation of collocation dictionaries. Given its universal and task-independent design, it can be easily adapted to the investigation of other linguistic phenomena as well as other complex tasks that imply the need for different skills and expertise.