Revisiting the Evaluation of Theory of Mind through Question Answering

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5872–5877, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5872 Revisiting the Evaluation of Theory of Mind through Question Answering Matthew Le Y-Lan Boureau Facebook AI Research New York, NY {mattle,ylan,maxn}@fb.com Maximilian Nickel Abstract Theory of mind, i.e., the ability to reason about intents and beliefs of agents is an important task in artificial intelligence and central to re- solving ambiguous references in natural lan- guage dialogue. In this work, we revisit the evaluation of theory of mind through question answering. We show that current evaluation methods are flawed and that existing bench- mark tasks can be solved without theory of mind due to dataset biases. Based on prior work, we propose an improved evaluation pro- tocol and dataset in which we explicitly con- trol for data regularities via a careful examina- tion of the answer space. We show that state- of-the-art methods which are successful on ex- isting benchmarks fail to solve theory-of-mind tasks in our proposed approach. 1 Introduction Humans interact and communicate with other peo- ple in a highly efficient way, as described for in- stance as Grice’s cooperative principle (Grice et al., 1975). Inferring other people’s mental state is a crucial component of cognitive development, play- ing a role in how humans learn the meaning of words (Bloom, 2002), distinguish beliefs from real- ity (Gopnik and Astington, 1988), predict people’s behavior (Wimmer and Perner, 1983), understand what others refer to, and reduce ambiguity in con- versation (Clark, 1981; Keysar et al., 2000). This ability to reason about the mental state of other agents, called theory of mind, is thus an impor- tant component of an intelligent system aiming to emulate and interact with humans. In developmental psychology, classic tests such as the Sally-Anne test (Baron-Cohen et al., 1985) have been used to assess the ability to infer false beliefs in others. Recently, Grant et al. (2017) and Nematzadeh et al. (2018) proposed to adapt these tests to evaluate the capability of machine learning models to form a theory of mind. The key insight of Grant et al. (2017) was to cast them as ques- tion answering tasks, where a system is given a story and has to answer questions about the beliefs of agents in that story. This allows to adapt the bAbi benchmarking protocol (Weston et al., 2016) to evaluate theory of mind capabilities of modern neural network architectures: stories are automati- cally generated so that a suitably large number of examples can be provided for training. We believe this is a promising approach for ad- vancing research on theory of mind, as it decouples its evaluation from problems such as multi-agent systems, game theory, and meta-learning, which are all components of other evaluation methods (Ra- binowitz et al., 2018; Bard et al., 2019). However, artificial data runs the risk of displaying hidden ar- tifacts and biases that correlate …

Content truncated to 500 words