Review-Driven Answer Generation For Product-Related Questions In E-Commerce

With the popularization of information technology,E-commerce has developed vigorously.While online shopping brings convenience,the drawbacks are gradually emerging.The narrow and limited approach to information makes it impossible for users to intuitively obtain product-related information,resulting in a large amount of time consumption and a poor user experience.In order to reduce the customer churn rate,leading E-commerce websites explored the use of intelligent question answering systems to alleviate the time cost of information seeking.However,the existing intelligent question answering system can not be widely used in E-commerce due to a series of problems(e.g.response efficiency and relying on a structured knowledge base).To address the demanding information need,this paper,we attempt to generate a natural answer for a product-realted question base on the relevant information provided in the reviews.Specifically,we propose a reivew-driven framework for answer generation,named RAGE,which is built on the basis of a multi-layer gated convolutional neural network.The model improves the response efficiency and the quality of generated answer mainly from four aspects of basic structure,review extraction,review representation and fusion method.In terms of basic structure,RAGE replaces the recurrent neural network with a multi-layer gated convolutional neural network,while untilizing position information and POS tags to enable the model to be aware of the portion it is currently dealing with and to enhance the ability of syntactic and grammatical representation.Then,we use sliding window and Word Mover's Distance to extract high-quality review snippets.To tackle the adverse impact of the nosieprone of reviews,we propose a weighted strategy which considers the word frequency and semantic relatedness to highlight the relevant words appearing in the review snippets.Then,we untilize both hierarchical attention and gate mechanism to inject the relevant information provided by the reivew snippets to guide to answer generation.Finally,we evaluate the RAGE and existing state-of-the-art alternative generation models over two real-world E-commerce QA datasets,and comprehensively measure the performance of the models from multiple perspectives.The experimental results demonstrate that RAGE can identify the relevant information from the noisy reviews and supervise the answer generation to produce more accurate and informative answer in natural language.