MatchACNN: A Multi-Granularity Deep Matching Model

This paper discusses a deep learning approach to ranking relevance in information retrieval (IR). In recent years, deep neural networks have led to exciting breakthroughs in speech recognition, computer vision, and natural language processing (NLP) tasks. However, the multi-granularity deep matching model has yielded few positive results. Existing deep IR models use the granularity of words to match the query and document. According to the human inquiry process, matching should be done at multiple granularities of words, phrases, and even sentences. MatchACNN, a new deep learning architecture for simulating the aforementioned human assessment process, is presented in this study. To solve the aforementioned problems, our model treats text matching as image recognition, extracts features from different dimensions, and employs a two-dimensional convolution neural network and an attention mechanism in image recognition. Experiments on Wiki QA Corpus, NFCorpus, and TREC QA show that MatchACNN can significantly outperform existing deep learning methods.