Relevant CommonSense Subgraphs for “What if...” Procedural Reasoning

Findings of the Association for Computational Linguistics: ACL 2022, pages 1927 - 1933 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Relevant CommonSense Subgraphs for "What if..." Procedural Reasoning Chen Zheng Michigan State University zhengc12@msu.edu Parisa Kordjamshidi Michigan State University kordjams@msu.edu Abstract We study the challenge of learning causal rea- soning over procedural text to answer "What if..." questions when external commonsense knowledge is required. We propose a novel multi-hop graph reasoning model to 1) effi- ciently extract a commonsense subgraph with the most relevant information from a large knowledge graph; 2) predict the causal answer by reasoning over the representations obtained from the commonsense subgraph and the con- textual interactions between the questions and context. We evaluate our model on WIQA benchmark and achieve state-of-the-art perfor- mance compared to the recent models. 1 Introduction In recent years, large-scale pre-trained language models (LMs) have made a breakthrough progress and demonstrate a high performance in many NLP tasks, including procedural text reasoning (Tandon et al., 2019; Rajagopal et al., 2020). There is a large amount of knowledge that is stored implicitly in language models that help in solving various NLP tasks (Devlin et al., 2019b). When we rea- son over text, sometimes, the knowledge contained in a given text is sufficient to predict the answer, as it is shown in the question 1 of Figure 1. This knowledge is directly encoded and used by LMs models (Tandon et al., 2019). However, there are many cases in which the required knowledge is not included in the procedural text itself. For example, for the question 2 in Figure 1, the information about the “nutrient” on the seeds does not exist in the pro- cedural text. Therefore, the external commonsense knowledge is required. There are several existing resources that contain world knowledge and commonsense. Examples are knowledge graphs (KGs) like ConceptNet (Speer et al., 2017) and ATOMIC (Sap et al., 2019). Look- ing back at the question 2, we observe that through providing the external knowledge triplets (nutrient, Procedural Text: 1. A plant produces a seed. 2. The seed falls to the ground. 3. The seed is buried. 4. The seed germinates. 5. A plant grows. 6. The plant produces flowers. 7. The flowers produce more seeds Questions and Answers: 1. suppose plants will produce more seeds happens, how will it affect less plants. (A) More (B) Less (C) No effect 2. suppose the soil is rich in nutrients happens, how will it affect more seeds are produced. (A) More (B) Less (C) No effect 3. suppose The sun comes out happens, how will it affect less plants. (A) More (B) Less (C) No effect Figure 1: WIQA contains procedural text, and different types of questions. The bold choices are the answers. relatedto, soil) and (soil, relatedto, seed) derived from ConceptNet, we can build an explicit reason- ing chain and choose an explainable answer. Two challenges exist in procedural text reason- ing and using external KBs. The first challenge is effectively extracting the most …

Content truncated to 500 words