Research On E-commerce Product Review Question-answering Syste

Product-Related Question Answering(PRQA),the task aimed at using relevant information in product reviews to answer specific questions,has gained much attention in recent years.Different from traditional Question Answering tasks,Product-Related Question Answering uses real datasets in E-commerce.Mass user-written reviews lack accurate annotations,and it’s hard to establish a connection between questions and reviews directly.Therefore,how to mine and utilize the information related to the question in the reviews is the main challenge facing the research of this task.According to different methods,the PRQA task can be divided into three sub-tasks:(1)The classification-based methods predict a “Yes” or “No” label as the answer;(2)The retrievalbased methods rank all reviews,and the most relevant review to the question is returned;(3)The generation-based methods directly generate the answer to the given question.This paper studies the three sub-tasks of PRQA respectively,including the following three aspects:1.For the PRQA task based on the classification methods,a two-stage answer prediction model based on a multi-task learning framework is proposed.It regards answerability prediction and answer prediction as a two-stage task,and learns by a multi-task learning framework.In addition,a multi-level attention mechanism and a joint-encoding layer are also designed to better integrate information related to the question in the reviews.The experimental results show that this model outperforms the existing best model in most evaluation metrics,and obtains the best performance in the unified evaluation experiment which simulates the real application scenarios.2.For the PRQA task based on the retrieval methods,a review retrieval model based on the relational graph convolutional network is proposed.It enhances the question-related information by explicitly modeling the semantic association among the question,answer,and each review sentence.At the same time,it also introduces more fine-grained keyword representation to integrate fine-grained information by establishing associations between the keywords and the related sentences in the relational graph.The experimental results show that this model outperforms all comparison models,and can accurately retrieve the reviews related to the given question.3.For the PRQA task based on the text generation methods,a fine-grained aspect and opinion aware answer generation model is proposed.It can match the relevant information in questions and reviews more accurately by leveraging fine-grained aspect and opinion information in sentences.What’s more,a new evaluation metric PRENR was introduced to measure the review consistency of the generated answers.Experimental results on three real Ecommerce datasets show that this model not only outperforms the state-of-the-art method in all evaluation metrics but also can generate more accurate and review-consistent answers.