基于注意力机制的循环神经网络评价对象抽取模型

针对评论文本中评价对象的抽取任务,需要设计特征模板,而抽取结果往往受特征模板影响大的问题,提出一种端到端的神经网络评价对象抽取模型。分析条件随机场CRF在评价对象抽取任务中的特征模板设计;使用词向量嵌入模型在语义空间表示词语,并分析注意力机制在神经网络模型中的作用;将条件随机场模型与循环神经网络模型LSTM相结合,形成基于注意力机制的LSTM-CRF-Attention模型。在NLPCC2012和NLPCC2013两个数据集上进行实验,该模型的F值比CRF模型分别提高8.15%和11.03%。实验结果也同时验证词向量具备表示词语特征的能力,注意力机制能够有效提高神经网络模型中的评价对象抽取效果。

EVALUATION OBJECT EXTRACTION MODEL OF RECURRENT NEURAL NETWORK BASED ON ATTENTION MECHANISM

Aiming at the problem that feature templates were needed to extract evaluation objects from comment texts, and the extraction results were often greatly affected by feature templates, we proposed an end-to-end neural network evaluation object extraction model. We analyzed the feature template design of conditional random field in the evaluation object extraction task. Then we used the word vector embedding model to represent words in the semantic space, and analyzed the role of attention mechanism in the neural network model. Combining the conditional random field model with the cyclic neural network model LSTM, the LSTM-CRF-Attention model was formed. Experiments on NLPCC2012 and NLPCC2013 show that the F value of the proposed model is 8.15% and 11.03% higher than that of CRF model respectively. The experimental results also verify that word vectors have the ability to represent word features. Attention mechanism can effectively improve the extraction effect of evaluation objects in the neural network model.