Text Matching as Time Series Matching

Text matching plays a fundamental and vital role in many aspects of natural language processing (NLP), where multi-level text matching is the most challenging task. Recently, the prevailing use of deep neural networks for text matching to fuse the multi-granularity semantic features has been witnessed. In this study, we adopt a two-stage text matching frame from rough to fine, which combines the recall stage and the ranking stage to extract different levels of matching information. In particular, the word embeddings of Query and Document are novel as time series in the ranking stage, and in response to this idea, we propose a simple and efficient neural network to extract matching information of signal trend patterns, which is built on Residual connection module, spatial Attention mechanism module and Multi-scale convolution module, namely RAM-CNN. The extensive experimental results show that the efficiency of the RAM-CNN model exceeds those of shallow models. And we finally demonstrate the effectiveness of the two-stage text matching model on a practical book-matching task.