【摘要】 神经网络模型可以有效地处理通用领域命名实体识别,然而在标注语料匮乏和包含大量噪声的特定领域,其性能通常会下降。针对这一问题,提出一种迁移学习神经网络模型TL-BiLSTM-CRF。利用双向长短时记忆网络提取具有字符级别形态特征的字符向量,结合具有语义、语序等特征信息的词向量作为输入,构建基本模型;在基本模型中引入词适应层,通过典型相关性分析算法弥合源域和目标域词向量特征空间的差异,对基本模型进行迁移。在社交媒体文本中的实验结果表明,该算法在Twitter数据集上F1值为64.87%,优于目前最好的模型。更多还原

【Abstract】 The neural network model can effectively deal with domain-general named entity recognition. However, in the specific domain lacking tagged corpora and containing a lot of noises, its performance usually decreases. Aiming at this problem, the TL-BiLSTM-CRF transfer model is proposed. The basic model was constructed through integrating with the character embeddings with the morphological features of character-level extracted by bidirectional long-term memory network and word embeddings with semantic, word order and other feature information; the basic model was transferred through introducing the word adaption layer into it and bridging the gap between the source and target embeddings spaces employing the method of canonical correlation analysis. The experimental results in the social media text show that the F1 value on the Twitter dataset is 64.87%, proving that the TL-BiLSTM-CRF transfer model is superior to the currently best method.更多还原