基于顺序遗忘编码和Bi-LSTM的命名实体识别算法

为了使长短时记忆网络(Long Short-Term Memory,LSTM)更精确地提取句子较远的特征信息,提出一种融合顺序遗忘编码(Fixed-size Oradinally Forgetting Encoding,FOFE)结合循环神经网络的命名实体识别算法。利用FOFE可以保留任意长度句子信息的编码方式来增强LSTM对句子特征的提取能力。利用Bi-LSTM和FOFE编码分别对向量化表示的文本进行特征提取和编码表示。结合得到的两个特征向量,通过注意力机制对Bi-LSTM的输入与输出之间的相关性进行计算,最后利用条件随机场学习标签序列的约束。该算法分别在英文和中文两种语言的数据集中进行了对比实验,F1值分别达到了91.30和91.65,验证了该方法的有效性。

NAMED ENTITY RECOGNITION ALGORITHM BASED ON ORDINALLY FORGETTING ENCODING AND BI-LSTM

In order to make long short term memory(LSTM)more accurate in extracting feature information of far sentences,we propose a named entity recognition algorithm which combines fixed-size ordinally forgetting encoding(FOFE)with recurrent neural network.Using FOFE can retain the encoding method of sentence information of any length to enhance the ability of LSTM to extract sentence features.We used Bi-LSTM and FOFE coding to extract and code the vectorized text respectively.Then we combined the two eigenvectors and calculated the correlation between the input and output of Bi-LSTM by the attention mechanism,and used the constraints of conditional random field to learn label sequences.The algorithm is tested in English and Chinese datasets,and the F1 values are 91.30 and 91.65respectively,which proves the effectiveness of the method.