关键词: 文本相似度, 信息交互, 双向长短时记忆, 自注意力机制, 协同注意力机制

Abstract: Text similarity calculation is one of  core tasks in natural language processing. Traditional text similarity calculation methods only consider structured features or semantic of the text. The lack of in-depth analysis of multiple text  features leads to low performance. Therefore, a text similarity calcu- lation method based on multiple related information interaction is proposed, which adds cosine correlation characteristics to the text embedding matrix. The self-attention mechanism is used to consider the text relevance of itself and word dependence. Further, the alternate co-attention machanism is used to extract the se-mantic interaction information between texts, so deeper and richer text representations from different perspectives indicate obtained. The experimental results on two datasets show that the F1 values of the proposed method are 0.916 1 and 0.769 5 respectively, and indicate the proposed method outperforms the benchmark method.

Key words: text similarity, information interaction, bi-directional long and short-term memory, self-attention mechanism, co-attention mechanism