Filter, Rank, and Transfer the Knowledge: Learning to Chat

Abstract

Learning to chat is a fascinating machine learning task with many applications from user-modeling to artificial intelligence. However, most of the work to date relies on designing large hard-wired sets of rules. On the other hand, the growth of social networks on the web provides large quantities of conversational data, suggesting that the time is ripe to train chatbots in a more data driven way. A first step is to learn to chat by ranking the response repository to provide responses that are consistent with the user’s expectations. Here we use a three phase ranking approach for predicting suitable responses to a query in a conversation. Sentences are first filtered, then efficiently ranked, and then more precisely re-ranked in order to select the most suitable response. The filtering is done using part-of-speech tagging, hierarchical clustering, and entropy analysis methods. The first phase ranking is performed by generating a large set of high-level grammatical and conceptual features, exploiting dictionaries and similarity measurement resources such as wikipedia similarity graphs, and by ranking using a boosted regression tree (MART) classifier. The more precise (conceptual) ranking is performed by designing more conceptual features obtained from similarity measurement resources such as query refinement and suggestion systems, sentence paraphrasing techniques, LDA topic modeling and structural clustering, and entropy analysis over wikipedia similarity graphs. The sentences are then ranked according to the confidence of a Transfer AdaBoost classifier, trained using transfer-learning methods in which a classification over a large corpus of noisy twitter and live-journal data is considered as the source domain, and the collaborative ranking of actively collected conversations, which are labeled in an online framework using user feedback, is considered as the destination domain. We give results on the performance of each step, and on the accuracy of our three phase ranking framework.