When a sentence does not introduce a discourse entity, Transformer-based models still sometimes refer to it

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 969 - 982 July 10-15, 2022 ©2022 Association for Computational Linguistics When a sentence does not introduce a discourse entity, Transformer-based models still sometimes refer to it Sebastian Schuster Center for Data Science Department of Linguistics New York University schuster@nyu.edu Tal Linzen Center for Data Science Department of Linguistics New York University linzen@nyu.edu Abstract Understanding longer narratives or participat- ing in conversations requires tracking of dis- course entities that have been mentioned. In- definite noun phrases (NPs), such as a dog, frequently introduce discourse entities but this behavior is modulated by sentential operators such as negation. For example, a dog in Arthur doesn’t own a dog does not introduce a dis- course entity due to the presence of negation. In this work, we adapt the psycholinguistic assessment of language models paradigm to higher-level linguistic phenomena and intro- duce an English evaluation suite that targets the knowledge of the interactions between sen- tential operators and indefinite NPs. We use this evaluation suite for a fine-grained inves- tigation of the entity tracking abilities of the Transformer-based models GPT-2 and GPT-3. We find that while the models are to a certain extent sensitive to the interactions we investi- gate, they are all challenged by the presence of multiple NPs and their behavior is not sys- tematic, which suggests that even models at the scale of GPT-3 do not fully acquire basic entity tracking abilities. 1 Introduction In order to understand longer narratives or to par- ticipate in conversations, humans and natural lan- guage understanding systems have to keep track of the entities that have been mentioned in the dis- course. For example, when someone tells you that Arthur owns a dog, they have introduced the entity of a person named Arthur and the entity of a dog owned by Arthur into the discourse. Once entities have been introduced to the discourse, it is natural to refer back to them with noun phrases or pro- nouns to elaborate further, e.g., by saying It has a red collar to elaborate on the dog. While no fully-specified account exists of how humans achieve this feat, many existing theories are based on the idea that humans maintain mental files (e.g., Heim, 1982; Murez and Recanati, 2016), i.e., explicit memory representations for each entity that encode all properties of an entity and its rela- tion to other entities. When engaging in a conver- sation or reading a longer narrative, humans then update these representations as they encounter new entities or new information about existing entities. Large pre-trained language models (LMs) such as GPT-2 (Radford et al., 2019) and GPT-3 (Brown et al., 2020), which in recent years have become the dominant foundation for natural language un- derstanding and generation tasks, lack explicit rep- resentations of discourse entities. It remains largely an open question to what extent LMs can match human behavior in tracking discourse entities. The most extensive …

Content truncated to 500 words