Abstract: In order to improve the accuracy for semantic similarity of short texts, this paper introduces a new method based on neural language model and compositionality in distributional semantics. Firstly the neural language model learns word representations with the local context and the global context, and better captures the semantics of words. And then it uses a dependency parse visualization tool, to get the dependence tree of the text, then using a semantic compositionality model to get the text semantic representation. According to the semantic representation, it uses VSM to judge the semantic relevance between the texts. The experimental results show this method can improve the accuracy for semantic similarity of short texts.

Key words: semantic analysis, neural network, semantic compositionality, text similarity, word representation

摘要： 为了更好地提高短文本语义相似度分析能力，提出了基于神经网络和组合语义的短文本语义相似度分析算法。利用神经网络构建词义表示模型，结合局部和全局上下文信息学习词语在实际语境下的表示；通过句法分析，得到文本的依存关系，并依此构建组合关系树，使用组合语义模型得到整个文本的语义表示；通过计算两个语义表示结果之间的相似性来计算文本的语义相似度。实验分析结果表明，该方法能够在一定程度上提高文本语义分析能力。

关键词: 语义分析, 神经网络, 组合语义, 文本相似度, 词义表示