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

Proceedings of the 24th Conference on Computational Natural Language Learning, pages 396–407 Online, November 19-20, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 396 Discourse structure interacts with reference but not syntax in neural language models Forrest Davis and Marten van Schijndel Department of Linguistics Cornell University {fd252|mv443}@cornell.edu Abstract Language models (LMs) trained on large quan- tities 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 in- teractions between different linguistic repre- sentations. In particular, we utilized stim- uli from psycholinguistic studies showing that humans can condition reference (i.e. corefer- ence resolution) and syntactic processing on the same discourse structure (implicit causal- ity). 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 neces- sary discourse information. Our results fur- ther suggest that LM behavior can contradict not only learned representations of discourse but also syntactic agreement, pointing to short- comings of standard language modeling. 1 Introduction Neural network language models (LMs), pretrained on vast amounts of raw text, have become the dom- inant input to downstream tasks (Peters et al., 2018; Radford et al., 2018; Devlin et al., 2019). Com- monly, these tasks involve aspects of language com- prehension (or understanding). One explicit exam- ple is coreference resolution, wherein anaphora (e.g., pronouns) are linked to antecedents (e.g., nouns) requiring knowledge of syntax, semantics, and world-knowledge to match human-like com- prehension. Recent work has suggested that LMs acquire abstract, often human-like, knowledge of syntax (e.g., Gulordava et al., 2018; Futrell et al., 2018; Hu et al., 2020). Additionally, knowledge of gram- matical and referential aspects linking a pronoun to its antecedent noun (reference) have been demon- strated for both transformer and long short-term memory architectures (Sorodoc et al., 2020). Hu- mans are able to modulate both referential and syntactic comprehension given abstract linguistic knowledge (e.g., discourse structure). Contrary to humans, we find that discourse structure (at least as it pertains to implicit causality) only influences LM behavior for reference, not syntax, despite model representations that encode the necessary discourse information. The particular discourse structure we examined is governed by implicit causality (IC) verbs (Gar- vey and Caramazza, 1974). Such verbs influence pronoun comprehension: (1) a. Sally frightened Mary because she was so terrifying. b. Sally feared Mary because she was so terrifying. In (1), she agrees in gender with both Sally and Mary, so both are possible antecedents. However, English speakers overwhelmingly interpret she as referring to Sally in (1-a) and Mary in (1-b), despite the semantic overlap between the verbs. Verbs that have a subject preference (e.g., frightened) are called subject-biased IC verbs, and verbs with a object preference (e.g., feared) are called object- biased IC verbs. In addition to pronoun resolution, IC verbs also interact with relative clause (RC) attachment: (2) a. John babysits the children of the musi- cian …

Content truncated to 500 words