Extending the Transformer with Context and Multi-dimensional Mechanism for Dialogue Response Generation

The existing work of using generative model in multi-turn dialogue system is often based on RNN (Recurrent neural network) even though the Transformer structure has achieved great success in other fields of NLP. In the multi-turn conversation task, a response is produced according to both the source utterance and the utterances in the previous turn which are regarded as context utterances. However, vanilla Transformer processes utterances in isolation and hence cannot explicitly handle the differences between context utterances and source utterance. In addition, even the same word could have different meanings in different contexts as there are rich information within context utterance and source utterance in multi-turn conversation. Based on context and multi-dimensional attention mechanism, an end-to-end model, which is extended from vanilla Transformer, is proposed for response generation. With the context mechanism, information from the context utterance can flow to the source and hence jointly control response generation. Multi-dimensional attention mechanism enables our model to capture more context and source utterance information by 2D vectoring the attention weights. Experiments show that the proposed model outperforms other state-of-the-art models (+35.8% better than the best baseline).