Improving short-text representation in convolutional networks by dependency parsing

摘要

Automatic question answering (QA) system is the inevitable trend of future search engines. As the essential steps of QA, question classification and text retrieval both require algorithms to capture the semantic information and syntactic structure of natural language. This paper proposes dependency-based convolutional networks to learn a representation of sentences. First, we use dependency layer to map discrete word depth on the dependency tree of a sentence into continuous real space. Then, the mapping result serves as weight of word vectors and convolutional kernels are employed as feature extractors for further specific tasks. The method proposed allows convolutional networks to take the advantage of higher representational ability of dependency structure. Experiments involving three tasks including text classification, duplicate classification and text pairs ranking confirm the advantages of our model.