The Grammar of Transformers: A Systematic Review of Interpretability Research on Syntactic Knowledge in Language Models
This review identifies gaps and biases in interpretability research for NLP researchers, highlighting an over-focus on English and BERT, making it incremental by synthesizing existing studies.
The authors conducted a systematic review of 337 articles to assess the syntactic abilities of Transformer-based language models, finding that models perform well on form-oriented phenomena like part of speech but show weaker results on syntax-semantics interface tasks such as binding dependencies.
We present a systematic review of 337 articles evaluating the syntactic abilities of Transformer-based language models, reporting on 1,015 model results from a range of syntactic phenomena and interpretability methods. Our analysis shows that the state of the art presents a healthy variety of methods and data, but an over-focus on a single language (English), a single model (BERT), and phenomena that are easy to get at (like part of speech and agreement). Results also suggest that TLMs capture these form-oriented phenomena well, but show more variable and weaker performance on phenomena at the syntax-semantics interface, like binding or filler-gap dependencies. We provide recommendations for future work, in particular reporting complete data, better aligning theoretical constructs and methods across studies, increasing the use of mechanistic methods, and broadening the empirical scope regarding languages and linguistic phenomena.