Deep and shallow features learning for short texts matching

摘要

Short texts matching problem is a special issue in natural language matching. Different from common natural language, short texts have their own characteristices, such as casual expressions and limited lengths, especially in the sentences from social media. Previous works usually use rule-based model and retrieval-based model to match short texts. These models merely focus on word-level similarity between short texts and can not capture deep matching relation of them. To boost the performance of short texts matching, we investigate a basic con-volutional neural network model to learn the sentence-level deep matching relation between short texts. Subsequently, we propose a hybrid model to merge sentence-level deep matching relation with shallow features to generate the final matching score. We evaluate our model on a dataset of short-text conversation based on real-world instances from Sina Weibo. The experimental results show that our model outperforms the previous state-of-art work on this task.