Automated Assistance in E-commerce: An Approach Based on Category-Sensitive Retrieval

摘要

This paper aims towards building an automated conversational assistant to help customers in an e-commerce scenario. Our dataset consists of live chat messages between human agents and buyers. These chats belong to many different issue types and we build a multi-instance SVM classifier to automatically classify these chats into the corresponding issue types. We further use this insight to append the category information obtained from the classifier to an LSTM based architecture to be able to provide appropriate responses given an utterance by a human agent. We find that using class information along with the base dual encoder model helps in improving the quality of the retrieved responses in terms of BLEU scores. Human judgement experiments validate that using class information is able to bring out relevant messages in top-3 and top-5 responses much more number of times than the base model that does not use the class information.