Social Commonsense Reasoning with Multi-Head Knowledge Attention

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2969–2980 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2969 Social Commonsense Reasoning with Multi-Head Knowledge Attention Debjit Paul Research Training Group AIPHES Institute for Computational Linguistics Heidelberg University paul@cl.uni-heidelberg.de Anette Frank Research Training Group AIPHES Institute for Computational Linguistics Heidelberg University frank@cl.uni-heidelberg.de Abstract Social Commonsense Reasoning requires un- derstanding of text, knowledge about social events and their pragmatic implications, as well as commonsense reasoning skills. In this work we propose a novel multi-head knowledge attention model that encodes semi- structured commonsense inference rules and learns to incorporate them in a transformer- based reasoning cell. We assess the model’s performance on two tasks that require different reasoning skills: Abductive Natural Language Inference and Counterfactual Invariance Pre- diction as a new task. We show that our proposed model improves performance over strong state-of-the-art models (i.e., RoBERTa) across both reasoning tasks. Notably we are, to the best of our knowledge, the first to demon- strate that a model that learns to perform coun- terfactual reasoning helps predicting the best explanation in an abductive reasoning task. We validate the robustness of the model’s rea- soning capabilities by perturbing the knowl- edge and provide qualitative analysis on the model’s knowledge incorporation capabilities. 1 Introduction Humans are able to understand natural language text about everyday situations effortlessly, by re- lying on commonsense knowledge and making inferences. For example in Figure 1, given two observations: Dotty was being very grumpy and She felt much better afterwards – we can come up with a plausible explanation about what could have provoked the change in Dotty’s emotion. We can also construct alternative hypotheses that will not change Dotty’s emotion. In order to judge the plausibility of such explanations, we need to have information about mental states and social norms, i.e., a form of commonsense knowledge. Such in- formation includes that calling a close friend, in Dotty was being very grumpy. Observation1: She felt much better afterwards. Observation2 : Dotty call some close friends to chat. Dotty ate something bad. Dotty is having a bad day cause Dotty feels annoyed effect Dotty wants to feel better wants Dotty seen as social seen as Dotty needed to have a good time intent Dotty feels friendly effect Dotty feels sick Dotty wanted to eat effect Dotty wants to eat something else wants Dotty seen as pleased Dotty feels relieved effect ❌ ✔ ✔ ⁉ seen as motivation Figure 1: Motivational example: The top and bottom blue boxes show two observations. The green and red box contain a plausible and an implausible hypothesis, respectively. A green line denotes that an event is likely to follow, the yellow line that an event is somewhat un- likely to follow, the red line something unlikely. general, makes people feel happy. This kind of inference goes beyond the broadly studied textual entailment task (Bowman et al., 2015) in that i) it re- quires a specific form of knowledge, namely knowl- edge about mental states …

Content truncated to 500 words