📄 Proceedings of the First International Workshop on Construction Grammars and NLP (CxGs+NLP, GURT/SyntaxFest 2023)

Summary

This volume contains the proceedings of the first workshop to bring together Construction Grammar (CxG) and natural language processing. The workshop was held in March 2023 at Georgetown University (Washington, D.C.) as part of GURT/SyntaxFest, together with Depling, TLT and UDW. The organizers wanted to bring the two fields closer. CxG holds that linguistic structures at every level carry meaning. Modern NLP, by contrast, relies on pretrained language models and often does without linguistic theory. All four events shared a joint review process, and the program committee chairs made the final assignment of accepted papers among them. The volume includes papers on a range of topics: an analysis of a computationally derived constructicon and the manual annotation of constructions in a treebank. Several papers test whether neural language models acquire constructions. Others deal with tokenization, vector representations and Fluid Construction Grammar. There are also studies of Chinese and Brazilian Portuguese. For projects like Mova, the volume is useful as an overview of ways to test what models know about form–meaning pairings and to annotate such phenomena in corpora.

Key points

• Joint review: 33 long papers were submitted and 27 accepted (82%), of which 9 went to CxGs+NLP. Of 16 short papers, 9 were accepted (56%), and the workshop received 2 of them.
• The workshop chairs are Claire Bonial (U.S. Army Research Lab) and Harish Tayyar Madabushi (University of Bath). Publisher: ACL, ISBN 978-1-959429-35-7.
• Resources: the Argument Structure Construction Treebank (Kyle, Sung) and a UMR annotation scheme for Chinese verb compounds.
• Model evaluation: CALaMo (Pannitto, Herbelot) proposes evaluating language models from a constructionist perspective. Weissweiler et al. show that CxG helps in understanding neural language models.
• Other topics: "constructivist" tokenization of English, an overview of Fluid Construction Grammar, empathy classification in medical essays, and negative results on the automatic detection of clausal constructions in Brazilian Portuguese.