Minding Language Models’ (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13960–13980 July 9-14, 2023 ©2023 Association for Computational Linguistics Minding Language Models’ (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker Melanie Sclar1 Sachin Kumar2 Peter West1 Alane Suhr3 Yejin Choi1,3 Yulia Tsvetkov1 1Paul G. Allen School of Computer Science & Engineering, University of Washington 2Language Technologies Institute, Carnegie Mellon University 3Allen Institute for Artificial Intelligence msclar@cs.washington.edu Abstract Theory of Mind (ToM)—the ability to reason about the mental states of other people—is a key element of our social intelligence. Yet, de- spite their ever more impressive performance, large-scale neural language models still lack ba- sic theory of mind capabilities out-of-the-box. We posit that simply scaling up models will not imbue them with theory of mind due to the in- herently symbolic and implicit nature of the phe- nomenon, and instead investigate an alternative: can we design a decoding-time algorithm that enhances theory of mind of off-the-shelf neu- ral language models without explicit supervi- sion? We present SYMBOLICTOM, a plug-and- play approach to reason about the belief states of multiple characters in reading comprehen- sion tasks via explicit symbolic representation. More concretely, our approach tracks each en- tity’s beliefs, their estimation of other entities’ beliefs, and higher-order levels of reasoning, all through graphical representations, allowing for more precise and interpretable reasoning than previous approaches. Empirical results on the well-known ToMi benchmark (Le et al., 2019) demonstrate that SYMBOLICTOM dramatically enhances off-the-shelf neural networks’ theory of mind in a zero-shot setting while showing ro- bust out-of-distribution performance compared to supervised baselines. Our work also reveals spurious patterns in existing theory of mind benchmarks, emphasizing the importance of out-of-distribution evaluation and methods that do not overfit a particular dataset. 1 Introduction Reasoning about other people’s intentions, desires, thoughts, and beliefs is a cornerstone of human social intelligence. Children naturally develop an understanding of every individual’s unique mental state and how it might impact their actions (Frith et al., 2003). Known as Theory of Mind (ToM) (Premack and Woodruff, 1978), this ability is cru- cial for efficient and effective communication. Alice and Bob are in a room with a basket and a box. Alice puts some celery in the basket and leaves the room. Bob then moves the celery into the box. Where will Bob search for the celery? (*) Where does Bob think that Alice will look for the celery when she returns? (**) Alice Bob is in is in is in room basket box celery is in is in is in room basket box celery (*) (**) Figure 1: A simple story requiring theory of mind. Note that Alice’s belief of the celery’s location differs from reality (i.e. Alice holds a false belief). Readers must reason that Alice will look for the celery where she left it, and that Bob will make that same assumption. Questions shown require different depths of mental state modeling. Cognitive and literary studies have extensively argued theory of mind’s …

Content truncated to 500 words