📄 Proceedings of the Second International Workshop on Construction Grammars and NLP

Summary

This is the proceedings volume of the Second International Workshop on CxGs+NLP. The workshop brings together Construction Grammars (CxG) and natural language processing. CxG describes linguistic knowledge as a network of form–meaning pairs at any level of abstraction and does not separate the lexicon from grammar. As a result, the theory covers compositional and non-compositional phenomena equally well. The organizers see CxG as a tool for testing what large language models (LLMs) actually know. At the same time, NLP methods help scale up constructionist theory itself. The first workshop took place shortly after ChatGPT appeared. The second focused on generative LLMs. The volume includes papers on computational frameworks and tools for CxG, on the constructional knowledge of LLMs and how to evaluate it, as well as empirical and theoretical studies. Among them are a construction-based NLI benchmark, an analysis of predicative constructions in UD, a Python package for analyzing argument structure constructions, and the modeling of sign languages. For Ukrainian NLP, the volume is useful as an overview of methods for testing the grammatical generalizations of language models at a deeper level than individual rules.

Key points

• The workshop took place on 24 September 2025 in Düsseldorf, co-located with the 16th IWCS conference. On 25 September, a separate community-building day was held, with panels and joint work on resources and data sharing.
• There were 35 submissions. Accepted were 17 archival papers presented in person, 5 online papers in the form of short talks with posters, and 9 non-archival posters. The volume's table of contents lists 21 papers.
• Invited speakers: Adele Goldberg (Princeton), Laura A. Michaelis (University of Colorado Boulder), and Thomas Hoffmann (Katholische Universität Eichstätt-Ingolstadt).
• Resources and papers include: the ASC analyzer package for English, verbal predication constructions in Universal Dependencies (Croft, Nivre), a constructional NLI benchmark and NLI fine-tuning for evaluating LLM generalizations, CxG-informed PEFT fine-tuning, and minimal pairs for Chinese resultative constructions.
• Workshop chairs: Claire Bonial (Georgetown University, Army Research Lab) and Harish Tayyar Madabushi (University of Bath). The volume was published by ACL, ISBN 979-8-89176-318-0.