TPOS Tagging Method Based on BiLSTM_CRF Model

• 作者: Hongwu Yang Lili Wang Ziyan Chen
• 来源期刊: 国际计算机前沿大会会议论文集
• 学科: 社会科学
• 关键词: DEEP LEARNING WORD EMBEDDING MODEL TIBETAN part-of-speech
• 年，卷（期）: 2019,（1）
• 页码范围: 501-503
• 页数: 3页
• 分类号: C

摘要

Part of speech (POS) tagging determines the attributes of each word, and it is the fundamental work in machine translation, speech recognition, information retrieval and other fields. For Tibetan part-of-speech (TPOS) tagging, a tagging method is proposed based on bidirectional long short-term memory with conditional random field model (BiLSTM_CRF). Firstly, the designed TOPS tagging set and manual tagging corpus were used to get word vectors by embedding Tibetan words and corresponding TPOS tags in continuous bag-of-words (CBOW) model. Secondly, the word vectors were input into the BiLSTM_CRF model. To obtain the predictive score matrix, this model using the past input features and future input feature information respectively learned by forward long short-term memory (LSTM) and backward LSTM performs non-linear operations on the softmax layer. The prediction score matrix was input into the CRF model to judge the threshold value and calculate the sequence score error. Lastly, a Tibetan part of speech tagging model was got based on the BiLSTM_CRF model. The experimental results indicate that the accuracy of TPOS tagging model based on the BiLSTM_CRF model can reach 92.7%.