摘要： 中文短文本通常使用单词序列而非字符序列进行语义匹配，以获得更好的语义匹配性能。然而，中文分词可能是错误或模糊的，容易引入噪声或者错误传播，从而损害模型的匹配性能。此外，多数中文词汇具有一词多义的特点，短文本由于缺少上下文环境，相比一词多义的长文本更难理解，这对于模型正确捕获语义信息是一个更大的挑战。提出一种短文本匹配模型，使用词格长短期记忆网络（Lattice LSTM）融合字符和字符序列的多粒度信息。引入外部知识HowNet解决多义词的问题，使用软注意力机制获取2个句子间的交互信息，并利用均值池化和最大池化算法进一步提取句子的特征信息，获取句子级语义编码表示。在数据集LCQMC和BQ上的实验结果表明，与ESIM、BIMPM和Lattice-CNN模型相比，该模型能有效提升中文短文本语义匹配的准确率。

关键词: 短文本语义匹配, 词格长短期记忆网络, 多粒度信息, 外部知识, 软注意力机制

Abstract: Chinese short text semantic matching generally uses word sequences, rather than character sequences, to achieve higher semantic-matching performance.However, Chinese word segmentation may be inaccurate or ambiguous, and word segmentation errors can introduce noise and lead to error propagation, thus deteriorating the final matching performance.Meanwhile, most Chinese words are characterized by polysemy.Short texts are more difficult to understand than long texts due to lack of context, which presents a greater challenge for the model to correctly capture semantic information.To solve this problem, a short text matching model is proposed.This model uses the Lattice Long Short Term Memory(Lattice LSTM) to integrate the multi-granularity information of characters and character sequences, introduces external knowledge HowNet to solve the problem of polyonyms, and employs a soft-attention mechanism to capture the interaction between two sentences.Furthermore, mean and maximum pooling algorithms are used to extract sentence features and obtain semantic encoding representations at the sentence level.Experiments on LCQMC and BQ datasets show that the model effectively improves the performance of Chinese short text semantic matching compared with the ESIM, BIMPM, and Lattice-CNN models.

Key words: short text semantic matching, Lattice Long Short Term Memory(Lattice LSTM), multi-granularity information, external knowledge, soft-attention mechanism

中图分类号: 

• TP18