针对目前各种基于长短期记忆网络 LSTM 的句子情感分类方法没有考虑词的词 性信息这一问题，将词性与自注意力机制相结合，提出一种面向句子情感分类的神经网络 模型 PALSTM(Pos and Attention-based LSTM)． 首先，结合预训练词向量和词性标注工具 分别给出句子中词的语义词向量和词性词向量表示，并作为 LSTM 的输入用于学习词在 内容和词性方面的长期依赖关系，有效地弥补了一般 LSTM 单纯依赖预训练词向量中词 的共现信息的不足;接着，利用自注意力机制学习句子中词的位置信息和权重向量，并构 造句子的最终语义表示;最后由多层感知器进行分类和输出． 实验结果表明，PALSTM 在 公开语料库 Movie Ｒeviews、Internet Movie Database 和 Stanford Sentiment Treebank 二元分 类及五元情感上的准确率均比一般的 LSTM 和注意力 LSTM 模型有一定的提升．

Aiming at the problem that most of existing LSTM-based methods for sentence sentiment classification don’t take into account the part-of-speech (POS) information of words，a neural network model，PALSTM， which combines POS and self-attention mechanism，was proposed and applied to sentence sentiment classification． Firstly，PALSTM used pre-trained word vectors and POS tagging tool to give the semantic and POS word vector rep- resentations of words in the sentences，and then took them as the inputs of a LSTM so as to capture the long-term dependence of words on content and part-of-speech，which effectively compensates for the common LSTM networks relying solely on the co-occurrence information of words in pre-trained word vectors． Secondly，the self-attention mechanism was used to learn the position information about words in the sentences and build the corresponding po- sition weight matrix，which yields the final semantic representations of sentences． Finally，the results was classi- fied and outputted via a multi-layer perceptron． The experiments show that PALSTM outperforms common LSTM and attentional LSTM models on some open corpus，i. e． Movie Reviews，Internet Movie Database，Stanford Sen- timent Treebank binary and fine-grained classification．