Research on Shift Matching to Enhance DAM

Human’s responses in communication depend on the context. Specifically, they are the feedback to a sentence or a word in the context. Further, external knowledge needs to be added to provide appropriate information for the human’s answer. DAM (Deep Attention Matching Network), uses the attention mechanism of transformer to expand utterance and response into multi-level granularity representations, and then calculate the granularity similarity at the same level, which has better effects than using traditional RNN (recurrent neural network). Inspired by DAM, we propose to calculate the similarity between granularities at different levels which can explore more useful information for training and learning in this paper. We call this new matching method "shift matching", which is not limited to enhancing DAM, but can be generalized to other models. Our experiments include two parts: the first part compares the improved model with the base, and then compares the classic model to solve multi-round dialogue problem. The second part is to compare the experimental results of the different shift distances. The results are better than that of the state-of-the-art model.