Answering Product-related Questions with Heterogeneous Information

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 696–705 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 696 Answering Product-related Questions with Heterogeneous Information∗ Wenxuan Zhang, Qian Yu, Wai Lam The Chinese University of Hong Kong {wxzhang,yuqian,wlam}@se.cuhk.edu.hk Abstract Providing instant response for product-related questions in E-commerce question answering platforms can greatly improve users’ online shopping experience. However, existing prod- uct question answering (PQA) methods only consider a single information source such as user reviews and/or require large amounts of labeled data. In this paper, we propose a novel framework to tackle the PQA task via ex- ploiting heterogeneous information including natural language text and attribute-value pairs from two information sources of the concerned product, namely product details and user re- views. A heterogeneous information encod- ing component is then designed for obtaining unified representations of information with dif- ferent formats. The sources of the candidate snippets are also incorporated when measur- ing the question-snippet relevance. Moreover, the framework is trained with a specifically de- signed weak supervision paradigm making use of available answers in the training phase. Ex- periments on a real-world dataset show that our proposed framework achieves superior per- formance over state-of-the-art models. 1 Introduction To help potential consumers address their con- cerns during online shopping, many E-commerce sites now provide a community question answer- ing (CQA) platform, where users can post ques- tions for a specific product, and others can vol- untarily answer them. Very often, it takes a long time for an asker to wait for an answer on such platforms. Therefore, automatically providing a proper response to a product-related question can greatly improve user online shopping experience and stimulate purchase decisions. ∗ The work described in this paper is substantially sup- ported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Code: 14200719). Several efforts have been made to tackle such product-related question answering (PQA) task (McAuley and Yang, 2016; Yu et al., 2018a; Gao et al., 2019; Chen et al., 2019b; Deng et al., 2020b). The existing methods can be generally categorized regarding the involved information source, i.e., from where the responses are obtained. A pio- neer work by McAuley and Yang (McAuley and Yang, 2016) investigates answer selection via de- tecting clues from user reviews. From then on, the review set becomes a commonly used auxil- iary information for predicting the answer types or distinguishing true answers from randomly sam- pled ones (Wan and McAuley, 2016; Yu and Lam, 2018). However, these methods are not feasible for newly-posted questions without candidate answers. A recent approach for PQA task is to directly ex- tract review sentences as the response for a given question (Chen et al., 2019a). But it requires a large number of labeled question-review pairs, whose an- notation is a time-consuming and laborious work. Other information sources, such as existing QA collections, are also …

Content truncated to 500 words