Reasoning Requirements for Indirect Speech Act Interpretation

Proceedings of the 28th International Conference on Computational Linguistics, pages 4937–4948 Barcelona, Spain (Online), December 8-13, 2020 4937 Reasoning Requirements for Indirect Speech Act Interpretation Vasanth Sarathy, Alexander Tsuetaki, Antonio Roque, Matthias Scheutz Human-Robot Interaction Laboratory Tufts University Medford, MA 02155 {vasanth.sarathy, alexander.tsuetaki, antonio.roque, matthias.scheutz}@tufts.edu Abstract We perform a corpus analysis to develop a representation of the knowledge and reasoning used to interpret indirect speech acts. An indirect speech act (ISA) is an utterance whose intended meaning is different from its literal meaning. We focus on those speech acts in which slight changes in situational or contextual information can switch the dominant intended meaning of an utterance from direct to indirect or vice-versa. We computationalize how various contextual fea- tures can influence a speaker’s beliefs, and how these beliefs can influence the intended meaning and choice of the surface form of an utterance. We axiomatize the domain-general patterns of reasoning involved, and implement a proof-of-concept architecture using Answer Set Program- ming. Our model is presented as a contribution to cognitive science and psycholinguistics, so representational decisions are justified by existing theoretical work. 1 Introduction Understanding natural language utterances inherently involves resolving ambiguity associated with the meaning of linguistic expressions. Resolving linguistic ambiguity involves choosing one amongst many possible meanings, at the word level, at the sentence level, and at the discourse level. Word and sentence embeddings are current topics of active research and focus on selecting suitable word or sentence rep- resentations that best capture meaning. Current transformer architectures such as BERT (Devlin et al., 2018) are used to pretrain language representations so word and sentence embeddings account for nearby linguistic context. These systems are based on two key insights that we explore in this paper. The first insight is that word meanings (in the case of BERT) and span meanings (in the case of SpanBERT (Joshi et al., 2020)) cannot be understood separate from their context, which in these cases is primarily linguis- tic context. Using these contextualized language embeddings have led to state-of-the-art performance in several downstream tasks. However, there is still the open problem of how to incorporate extra-linguistic context. Specifically, it is unclear how to categorize and use knowledge derived from common sense knowledge about how the world works as well as situational knowledge about aspects of the physical environment in a systematic way for NLU. The second insight is that the way context influences meaning can be entirely captured statistically from data. That is, a word encountered together with context in a particular way is likely to mean the same as the same word encountered in a similar linguistic context. The assumption is that systems with sufficient training data, or at least with powerful pretrained models are likely to interpret language correctly. However, Cohen (2019) and others have argued that pattern recognition by itself might be insufficient, and such an association might require logical reasoning beyond learning statistical similarity. Recent work by Sarathy et al. (2019) explored both these insights and argued that for the …

Content truncated to 500 words