基于相关主题模型和多层知识表示的文本情感分析

Text Sentiment Analysis Based on Correlated Topic Model and Multi-layer Knowledge Representation

将相关主题模型和多层知识表示方法相结合开展文本情感分析研究。首先,针对传统分割算法的不足和主题间相关关系,采用相关主题模型对文本进行主题特征分割,构造主题先验信息输入预训练语言模型;其次,基于主题先验信息和相关关系向量,采用预训练的语言模型嵌入进行文本词的动态表示,能有效解决一词多义的问题;最后,使用双向长短期记忆模型对文本句子进行表示,考虑每个词的前后信息来捕捉句子的位置信息,在句子表示向量的信息抽取中融入注意力机制,使用多头抽取考虑全局的方式,可以抽取更全面的文本信息。

Correlated topic model and multi-layer knowledge representation method were integrated to carry out text sentiment analysis.Firstly,considering the shortcomings of traditional segmentation algorithms and topic correlation,the correlated topic model was used to segment the text features,and topic prior information was built as the input of pre-trained language model.Secondly,the pre-trained embeddings from language model was used to dynamically express the words of texts based on topic prior information and correlation vector,which could effectively solve the polysemy problem.Finally,bidirectional long short-term memory model was adopted to represent the sentences of text,and the information before and after each word was considered to capture sentence position.Attention mechanism was integrated into information extraction of sentence representation vector,which used multi-head extraction method to extract more comprehensive information of texts.