Motivation:

We aim to build an intelligent chat bot for answering both general questions (chit-chat) and summer camp (CISS) related questions.

Model:

Fig.1: The overview of our framework. It contains two modules, one is a seq2seq-based generative model for answering chit-chat questions, the other is retrieval-based model for selecting the most relevant answer for the given summer camp (CISS) question.

Fig.2: The seq2seq-based generative model for answering chit-chat questions.

Fig.3 The retrieval-based model for selecting the most relevant answer for the given summer camp (CISS) question. Given a question and a answer pair, we use two different LSTMs to encode them respectively. A matching attention mechanism is adopted to better capture their relations and model their semantic meanings. Then, a score function (e.g., cosine sim, bilinear) is applied to calculate the similarity between the question and the answer. We adopt pairwise ranking function to train the whole module.

Our system (build on Telegram):