Structure-Based Convolutional Neural Networks for Chinese Sentence Similarity Measurement
Yang Sun, Fangyu Hu · 2018
Chinese semantics is more complex than English, which is mainly due to that different combinations of the same Chinese characters may have completely different meanings. Interdependency of words can help better understand complex Chinese semantic. Most Chinese sentence similarity calculating merely concentrate on word similarity measurement but cannot process sentence representation, and even in English sentence representation, there are few people paying attention to the interdependency structure of words in sentence. We propose a dependency parsing based MPCNN model (DP-MPCNN) which combines not only the word vector but also the sentence structure information to estimate the similarity between sentence pairs. We first apply the result of dependency parsing from LTP platform to utilize the interdependency relation of words, and different granularity of kernels are used separately to extract the features of sentence and interdependency structure of words. Then we can get the similarity value of sentence pairs through similarity measurement layer and fully connected layer of MPCNN. Our experiments on ChineseSTS show that our proposed method outperforms other baselines in the task of sentence similarity calculating.