Semantic representation and dependency information are of great importance for matching a response with its multi-turn context. In this paper, we propose an enhanced matching network (EMN) to enhance the matching ability for the multi-turn response selection system in terms of both constructing semantic representation and extracting dependency information. First, the commonly used recurrent neural network (RNN) is replaced with gated convolutional neural network (GCNN) in the matching network to construct more expressive semantic representations of sentences. Second, local inference modeling and inference composition in the enhanced sequential inference model (ESIM) are utilized here to capture enhanced interactive information between the response and each utterance in the context. Finally, EMN is based on a similar multi-turn structure to sequential matching network (SMN) for extracting turns' dependency information in the chronological order. We furthermore propose a combined model (EMN-SMN) to integrate SMN into EMN for distilling more important dependency information from sentence pairs. Experiments are carried out on Ubuntu Corpus and Douban Conversation Corpus. The results show that EMN can outperform the state-of-the-art methods and the combined model can further improve overall performance.

In this paper, we propose a novel framework by combining the phonetic temporal neural model (PTN) with an improved LSTM (IM-LSTM). This is achieved by using an up-down connection from the time t to t+1 in the LSTM structure, which aims to capture the latent information from the previous time step. This updated structure can perform better to discriminate the frame-level phonetic information produced by PTN. On the AP16-OLR language identification dataset, our final model achieves relative growth rate 5.04%, 2.19%, 2.73% on EER and 6.55%, 5.81%, 2.23% on C-avg in 1s, 3s and fulllength utterance condition than the standard PTN, respectively. The proposed framework receives a better performance than the standard PTN and other proposed models, particularly in 1s condition. This shows the efficacy and flexibility of the proposed method.