Case-Based Abductive Natural Language Inference

Proceedings of the 29th International Conference on Computational Linguistics, pages 1556–1568 October 12–17, 2022. 1556 Case-Based Abductive Natural Language Inference Marco Valentino, Mokanarangan Thayaparan, Andr´e Freitas Idiap Research Institute, Switzerland University of Manchester, United Kingdom marco.valentino@idiap.ch Abstract Most of the contemporary approaches for multi-hop Natural Language Inference (NLI) construct explanations considering each test case in isolation. However, this paradigm is known to suffer from semantic drift, a phe- nomenon that causes the construction of spuri- ous explanations leading to wrong conclusions. In contrast, this paper proposes an abductive framework for multi-hop NLI exploring the retrieve-reuse-refine paradigm in Case-Based Reasoning (CBR). Specifically, we present Case-Based Abductive Natural Language In- ference (CB-ANLI), a model that addresses un- seen inference problems by analogical transfer of prior explanations from similar examples. We empirically evaluate the abductive frame- work on commonsense and scientific question answering tasks, demonstrating that CB-ANLI can be effectively integrated with sparse and dense pre-trained encoders to improve multi- hop inference, or adopted as an evidence re- triever for Transformers. Moreover, an empir- ical analysis of semantic drift reveals that the CBR paradigm boosts the quality of the most challenging explanations, a feature that has a direct impact on robustness and accuracy in downstream inference tasks. 1 Introduction Multi-hop inference is the task of composing two or more pieces of evidence from external knowledge resources to address a particular reasoning prob- lem (Thayaparan et al., 2020). In the context of Natural Language Inference (NLI), this task is of- ten used to develop and evaluate explanation-based systems, capable of performing transparent multi- step reasoning with natural language (Wiegreffe and Marasovic, 2021; Jansen et al., 2018; Cam- buru et al., 2018). While multi-hop inference has been largely explored for extractive problems such as open-domain question answering (Yang et al., 2018), increasing attention is being dedicated to Similar Case a stick is a kind of object friction is a kind of force h: Two sticks getting warm when rubbed together is an example of a force producing heat magnetic attraction pulls two objects together pull is a force friction causes the temperature of an object to increase Retrieve friction is a kind of force friction causes the temperature of an object to increase Which force produces energy as heat? Friction Reuse and Refine Correct Explanation Spurious Explanation Figure 1: Performing multi-hop inference considering each case in isolation can lead to the construction of spurious explanations. In contrast, we propose the adoption of a Case-Based Reasoning (CBR) paradigm where the construction of new explanations is con- strained by previously solved examples. the abstractive setting, where the models are re- quired to compose long chains of facts expressing abstract commonsense and scientific knowledge (Clark et al., 2018; Valentino et al., 2022). In this setting, multi-hop inference is often framed as an Abductive Natural Language Infer- ence (ANLI) problem (Bhagavatula et al., 2019), where, for a given set of alternative hypotheses H = {h1, h2, . . . , hn}, the goal is to construct an explanation for …

Content truncated to 500 words