Multi-Granularity Neural Sentence Model for Measuring Short Text Similarity

摘要

Measuring the semantic similarities between short texts is a critical and fundamental task because it is the basis for many applications. Although existing methods have explored this problem through enriching the short text representations based on the pre-trained word embeddings, the performance is still far from satisfaction because of the limited feature information. In this paper, we present an effective approach that combines convolutional neural network and long short-term memory to exploit from character-level to sentence-level features for performing the semantic matching of short texts. The proposed approach nicely models the feature information of sentences with the multiple representations and captures the rich matching patterns at different levels. Our model is rather generic and can hence be applied to matching tasks in different language. We use both paraphrase identification and semantic similarity tasks for evaluating our approach. The experimental results demonstrate that the proposed multiple-granularity neural sentence model obtains a significant improvement on measuring short texts similarity compared with the existing benchmark approaches.