Reversing the NLP Pipeline: What Do LMs Have to Offer Linguistics?

ACL ARR 2024 April Submission108 Authors

Abstract: Language models (LMs) have profoundly reshaped the field of natural language processing (NLP) in recent years. This paper presents the first investigation of the implications of LMs for linguistics, examining their role as a source of linguistic data. Despite their limitations, LMs have shown remarkable fluency and proficiency in modeling human language, suggesting their potential as linguistic artifacts. By proposing approaches to uncover these encoded linguistic structures, the paper explores how LMs provide opportunities for linguists. Embracing LMs as linguistic artifacts reopens the door for collaboration between NLP and linguistics, overcoming limitations of previous paradigms and providing new opportunities for the field of linguistics research.

Paper Type: Short

Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics

Research Area Keywords: linguistic theories, reproducibility, NLP tools for social analysis, NLP in resource-constrained settings, probing

Contribution Types: Position papers, Theory

Languages Studied: English

Submission Number: 108