Proceedings of the 21st Workshop on Multiword Expressions (MWE 2025)

Atul Kr. Ojha, Voula Giouli, Verginica Barbu Mititelu, Mathieu Constant, Gražina Korvel, A. Seza Doğruöz, Alexandre RademakerNAACL 2025

TL;DR

This document presents the proceedings of the 21st Workshop on Multiword Expressions (MWE 2025), showcasing research on the complex linguistic phenomena of multiword expressions (MWEs). The papers explore challenges and advancements in MWE processing, identification, and representation, particularly in the context of large language models (LLMs) and low-resource languages.

What Problem Does This Paper Address?

Multiword expressions (MWEs) – such as idioms, phrasal verbs, and collocations – pose significant challenges for computational linguistics due to their lexical, syntactic, semantic, and pragmatic idiosyncrasies. These expressions often defy standard linguistic modeling and NLP tasks like parsing and machine translation, hindering the development of robust natural language understanding systems. This workshop addresses the ongoing need for better methods to process, identify, and represent MWEs, especially in diverse linguistic contexts including low-resource languages and the rapidly evolving landscape of large language models.

Methods

The papers presented at MWE 2025 employ a variety of methods to tackle MWE-related challenges. These include:

  • Lexical Resource Development: Creating and enhancing lexica and corpora specifically designed for MWEs, often with a focus on low-resource languages or specific linguistic phenomena (e.g., verbal idioms, ambiguous noun-adjective MWEs).
  • Model Training and Evaluation: Utilizing and adapting existing NLP models, including transformer-based models like BERT and LLMs, for MWE identification, interpretation, and disambiguation tasks. This involves fine-tuning models on specific MWE-related datasets and exploring how different model layers capture MWE properties.
  • Corpus Annotation: Developing new annotation schemes and creating annotated treebanks (e.g., Universal Dependencies treebanks) to facilitate empirical investigation of MWEs across languages.
  • Cross-lingual Transfer: Investigating methods for transferring MWE knowledge and disambiguation capabilities across languages, particularly for idiomatic expressions.
  • Compositionality Analysis: Developing methods to rate and analyze the compositionality of MWEs, providing gold standards for intrinsic evaluation.

Datasets

The workshop features research utilizing and introducing several datasets:

  • Standard Modern Greek UDv.2 Treebank: Used for VMWE identification.
  • European Portuguese Corpus: Annotated for verbal idioms.
  • MultiCoPIE: A new multilingual corpus of potentially idiomatic expressions in Catalan, Italian, and Russian.
  • Galician Dataset: Contains 240 ambiguous noun-adjective MWEs, contextualized in sentences, with ratings for compositionality, frequency, ambiguity, and productivity.
  • Irish Named Entity Corpus: A small, gold-standard corpus annotated with named entities.
  • Synthetic Idiom Datasets: Generated using LLMs for training task-specific models.

Key Insights and Ideas

  • LLMs as a Double-Edged Sword: LLMs show promise in advancing idiom corpus construction and can capture lexical and syntactic properties of MWEs. However, they also present new challenges, such as the need for robust tokenization strategies in the era of LLMs.
  • The Importance of Low-Resource Languages: There's a continued focus on developing resources and tools for endangered and low-resource languages, highlighting the need for MWE research to be inclusive and linguistically diverse.
  • Compositionality as a Key Evaluation Metric: The workshop emphasizes the role of compositionality in evaluating MWE representations and models, suggesting it as a crucial standard for intrinsic evaluation.
  • Cross-lingual MWE Understanding is Crucial: The development of multilingual resources and the exploration of cross-lingual transfer for MWE disambiguation underscore the growing importance of handling MWEs across different languages.
  • MWEs as a Testbed for Linguistic Theory: The study of MWEs continues to inform and challenge linguistic theories, particularly Construction Grammar, by providing empirical data on how form and meaning interact in complex ways.

Key Results

  • The workshop features a diverse range of accepted papers, with 75% of submissions being accepted after peer review, indicating a high level of research activity in the field.
  • Several papers introduce new lexical resources and annotated corpora for various languages, including Portuguese, Irish, and a multilingual set for Catalan, Italian, and Russian.
  • Research demonstrates the application of LLMs for tasks like generating synthetic idiom datasets and improving MWE identification.
  • Studies on specific languages like Galician and Greek showcase advancements in understanding and processing MWEs within those linguistic contexts.

Limitations and Future Work

  • Low-Resource Language Challenges Persist: While progress is being made, the development of comprehensive MWE resources for all low-resource languages remains a significant challenge.
  • LLM Interpretation and Control: Further research is needed to fully understand and control how LLMs represent and process MWEs, especially concerning their latent representations and potential biases.
  • Tokenization in the LLM Era: The panel discussion on tokenization in the era of LLMs suggests this is an active area of research with ongoing challenges and no single, universally accepted solution.
  • Cross-lingual Transfer Effectiveness: While promising, the effectiveness and limitations of cross-lingual transfer for MWE disambiguation require continued investigation.
  • Standardization of Evaluation: The ongoing development of gold standards for intrinsic evaluation, particularly concerning compositionality, suggests a need for continued efforts in standardizing evaluation methodologies.

Summary updated

BibTeX
@inproceedings{ojha2025proceedings,
  author = {Atul Kr. Ojha and Voula Giouli and Verginica Barbu Mititelu and Mathieu Constant and Gražina Korvel and A. Seza Doğruöz and Alexandre Rademaker},
  title = {{Proceedings of the 21st Workshop on Multiword Expressions (MWE 2025)}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure