Abductive Commonsense Reasoning Exploiting Mutually Exclusive Explanations

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 14883–14896 July 9-14, 2023 ©2023 Association for Computational Linguistics Abductive Commonsense Reasoning Exploiting Mutually Exclusive Explanations Wenting Zhao and Justin T. Chiu and Claire Cardie and Alexander M. Rush Department of Computer Science Cornell University {wz346,jtc257,ctc9,arush}@cornell.edu Abstract Abductive reasoning aims to find plausible ex- planations for an event. This style of reasoning is critical for commonsense tasks where there are often multiple plausible explanations. Ex- isting approaches for abductive reasoning in natural language processing (NLP) often rely on manually generated annotations for supervi- sion; however, such annotations can be subjec- tive and biased. Instead of using direct super- vision, this work proposes an approach for ab- ductive commonsense reasoning that exploits the fact that only a subset of explanations is correct for a given context. The method uses posterior regularization to enforce a mutual ex- clusion constraint, encouraging the model to learn the distinction between fluent explana- tions and plausible ones. We evaluate our ap- proach on a diverse set of abductive reasoning datasets; experimental results show that our ap- proach outperforms or is comparable to directly applying pretrained language models in a zero- shot manner and other knowledge-augmented zero-shot methods. 1 Introduction Abductive reasoning aims to find plausible expla- nations for an event (Paul, 1993). Unlike deduc- tion, which draws a firm conclusion from a set of premises, abduction requires reasoning from an outcome to plausible explanations. Fig. 1 (top) demonstrates the distinction: given only the con- text x, both the blue and the red sentences describe possible subsequent events; however, upon seeing the outcome y only one of the two is a plausible ex- planation (although there may be others). Humans apply abduction in everyday situations (Andersen, 1973) such as reading-between-the-lines (Char- niak and Shimony, 1990) and analyzing causes and effects (Thagard and Shelley, 1997; Pearl and Mackenzie, 2018). Learning to perform abduction is thus an important step towards building human- like machines with commonsense knowledge. x = Emily was stuck in traffic. y = Emily made it to her flight. Her flight was delayed. Her flight left on time. Tuned LiPoR LM p( explanation | x, y ) ? Zero-shot Figure 1: Top: An abductive reasoning example con- sisting of a context x, an outcome y, and two candidate explanations. The goal is to identify the plausible ex- planation given x and y. To predict an explanation, one can apply a pretrained language model (shown as LM) to score y given x and an explanation, and then compute the posterior probability for the explanation. Bottom: Using a LM without fine-tuning (Zero-shot) leads to poor performance, whereas a LM fine-tuned via max-marginal likelihood (Tuned) fails to distinguish the two explanations. LiPoR is trained to partition the explanations in a mutually exclusive manner. Abductive reasoning has been extensively stud- ied in the setting where annotations are avail- able (Storks et al., 2019). However, because de- termining whether an explanation is plausible is …

Content truncated to 500 words