Survey on Lexical Resources Focused on Multiword Expressions for the Purposes of NLP

Verginica Mititelu, Voula Giouli, Gražina Korvel, Chaya Liebeskind, Irina Lobzhanidze, Rusudan Makhachashvili, Stella Markantonatou, Aleksandra Markovic, Ivelina StoyanovaNAACL 2025

TL;DR

This paper presents a comprehensive survey of computational lexical resources for Multiword Expressions (MWEs), updating a previous survey from 2016. It analyzes over 60 resources across various languages, highlighting trends in their creation, usage, and accessibility, and identifying critical gaps in support for less-resourced languages.

What Problem Does This Paper Address?

Multiword Expressions (MWEs) – like "kick the bucket" or "spill the beans" – are notoriously tricky for Natural Language Processing (NLP) systems. Their meaning often can't be deduced from the individual words, making automatic identification and understanding a significant challenge. To tackle this, computational lexica dedicated to MWEs are crucial. However, the landscape of these resources is fragmented and constantly evolving. This paper aims to provide a clear, up-to-date overview of these MWE lexica, identifying what exists, how it's structured, and where the gaps are, particularly concerning language coverage and accessibility.

Methods

The authors conducted a systematic survey of MWE lexical resources. They defined specific inclusion criteria, focusing on computational lexica rather than general dictionaries or terminological databases. Their data collection involved searching major repositories like the European Language Grid (ELG) and the ACL Anthology, as well as consulting specialized databases and book series. They also reached out to the community for information on newly developed resources. The collected information was then analyzed across several axes: time span, intended usage, linguality type, languages covered, MWE types, acquisition method, accessibility, and linkage to other resources.

Datasets

This paper is a survey and does not introduce new datasets for training or evaluation. Instead, it catalogs and analyzes existing MWE lexical resources.

Key Insights and Ideas

  • The MWE Lexicon Landscape is Growing, but Unevenly: While the number of MWE resources has increased since the last survey, there's a significant imbalance in language coverage, with well-resourced languages like English dominating.
  • Focus Shifting to Less-Resourced Languages: A positive trend observed is an increasing effort to develop MWE resources for languages with less digital support, often driven by EU-funded initiatives.
  • Accessibility is Good, but Documentation Needs Work: A majority of the identified resources are freely available, often through international catalogs. However, the detailed description of their contents is frequently lacking, making it hard to assess their precise utility.
  • Manual Curation Still Key: Despite advances in automation, a substantial portion of MWE resources still requires manual development, underscoring the complexity of capturing MWE nuances.
  • Linking is Crucial, but Underdeveloped: Most MWE lexica are not linked to corpora or other lexical resources, limiting their applicability in downstream NLP tasks.

Key Results

The survey identified 66 MWE-related resources, a decrease from the 107 reported in a previous survey (2016), but with a focus on computational lexica.

  • Usage: 78% of resources are purely computational, and 19% serve both computational and human users.
  • Languages: 51 resources are monolingual, covering 37 languages. 10 are bilingual, and 5 are multilingual. English is the most represented language.
  • Acquisition: Resources are primarily developed semi-automatically (39.9%) or manually (31.8%).
  • Accessibility: 74% of resources are free or free for academic use.
  • Linking: Only 33.3% of resources are linked to corpora, and 24.2% to other lexical resources.

Limitations and Future Work

The authors acknowledge that the survey's comprehensiveness is limited by the availability and discoverability of resources. They also note that the description of resource contents often lacks the detail needed for precise understanding. Future work will focus on exploring the types of linguistic information provided by these resources, identifying best encoding practices, and further investigating the challenges of representing MWEs, especially for less-resourced languages and language varieties.

Summary updated

BibTeX
@inproceedings{mititelu2025survey,
  author = {Verginica Mititelu and Voula Giouli and Gražina Korvel and Chaya Liebeskind and Irina Lobzhanidze and Rusudan Makhachashvili and Stella Markantonatou and Aleksandra Markovic and Ivelina Stoyanova},
  title = {{Survey on Lexical Resources Focused on Multiword Expressions for the Purposes of NLP}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure