Discourse structure interacts with reference but not syntax in neural language models
This reveals shortcomings in LMs' ability to acquire abstract linguistic knowledge, which is important for improving natural language processing systems.
The study tested whether neural language models (LMs) learn interactions between linguistic representations, specifically how discourse structure (implicit causality) influences reference and syntax, finding that LMs only condition reference on discourse structure, not syntax, unlike humans.
Language models (LMs) trained on large quantities of text have been claimed to acquire abstract linguistic representations. Our work tests the robustness of these abstractions by focusing on the ability of LMs to learn interactions between different linguistic representations. In particular, we utilized stimuli from psycholinguistic studies showing that humans can condition reference (i.e. coreference resolution) and syntactic processing on the same discourse structure (implicit causality). We compared both transformer and long short-term memory LMs to find that, contrary to humans, implicit causality only influences LM behavior for reference, not syntax, despite model representations that encode the necessary discourse information. Our results further suggest that LM behavior can contradict not only learned representations of discourse but also syntactic agreement, pointing to shortcomings of standard language modeling.