Multi-perspective Inductive Answering of Subjective Questions on Products

This article focuses on the topic of answering subjective questions about products by multi-perspective induction. Unlike traditional factoid QA, whose answers are unique and can be extracted directly from the text, the answers to subjective QA are not a simple text span, but involve various viewpoints and facts. A good answer should incorporate both objective facts and subjective opinions from various sources. Facts often involve multiple fine-grained product aspects, some of which are implicit and not explicitly mentioned. The opinions are complex, hidden, and scattered. Questions may be vague, sometimes requiring commonsense inference. This challenging task has wide-ranging applications, but limited research has studied it. To address this problem, we propose a new model to answer subjective questions in an inductive way. Given a question, we first retrieve the missing but necessary commonsense knowledge to supplement its implicit aspects and hidden relations, to better understand the ask points and users' needs. We then parse and build an aspect tree from the factual data, such as product descriptions, to incorporate all aspects and their parent-child relations. Based on the tree, we infer the question-related aspects from the retrieved content. For each aspect, we aggregate its scattered opinions and objective facts to yield a summary. To deal with diverse types adaptively, we construct an aspect-controlled generation model. Each aspect summary would add a prefix that is learned through a trainable gating mechanism and fused into a multi-aspect decoder to derive a comprehensive answer. The whole model with a retriever and generator is jointly trained by a reinforced framework. Extensive experiments are conducted on our created large-scale SupQA dataset, and the results show the effectiveness of our approach.