Discourse structure interacts with reference but not syntax in neural language models

TL;DR

This study investigates how neural language models process discourse structure, revealing that they encode discourse information affecting reference but not syntax, highlighting limitations in their linguistic abstraction capabilities.

Contribution

It demonstrates that current neural language models encode discourse information influencing reference but fail to integrate it with syntactic processing, unlike humans.

Findings

Models encode discourse info affecting reference resolution.

Discourse info does not influence syntax in models.

Models show limitations in linguistic abstraction.

Abstract

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.