An Extractive Chat Summary Generation Method for Ecommerce Chatbots

The usage of chatbot for handling the conversations is gradually increasing. By using this technique, companies can provide a real-time and efficient channel to deliver the information to their users. Chatbot can response significant volume of frequent-asked and typical questions, which is bound to reduce the company human resource consumption. However, sometimes a chatbot cannot satisfy users with the provided answers, in such situation, chatbot should be transferred to an agent in order to handle the rest of the conversation. A concise summary of the user utterances helps the agent to handle the rest of the chat without raising annoying delay. Because the dialog between chatbots and user is fragmented and informal when compared to classic summarization dataset, high complexity models often have difficulty in achieving ideal results. In this paper, we present an extractive chat summarization system to provide a concise summary of the discussed topics in chat. We conclude three groups of critical and universally applicable features dedicated to summarization tasks and use the features to extract keywords for ranking sentences’ importance in the chat. We evaluate supervised and unsupervised methods for keyword extraction and generate the summary by combining selected sentences with specific rules. We compare our obtained results with two state-of-art, deep learning based methods over a new chatlog dataset. Experimental results reveal our method to be more effective and efficient than currently popular, high complexity methods.