基于Bi-LSTM-CRF网络的语义槽识别  

Slot filling based on Bi-LSTM-CRF

作　　者：徐梓翔 车万翔[1] 刘挺[1] 

机构地区：[1]哈尔滨工业大学计算机科学与技术学院,哈尔滨150001

出　　处：《智能计算机与应用》2017年第6期91-94,共4页Intelligent Computer and Applications

摘　　要：自然语言理解任务的主要目标是运用自然语言处理的相关方法,对用户发出的语句进行解析,转化成结构化的语义表示。本文重点研究了基于语义槽抽取的自然语言理解方法。在基于语义槽抽取的自然语言理解任务中,任务的输入是用户的指令型语句,输出为指令的语义槽实体标注序列,如出发日期、出发地点等,故可将语义槽抽取任务看作类似于命名实体识别任务,以序列标注任务的方法解决。本文研究提出了基于Bi-LSTM-CRF模型的语义槽抽取方法,在英文语料ATIS上进行了实验。实验结果表明,基于Bi-LSTM-CRF网络的方法相比于传统机器学习的基准方法,结果得到了大幅度的提升。对于模型识别的结果,研究中采用F1值进行评价。The main purpose of natural language understanding is to transform natural language to structural representation.One of the methods of natural language understanding is slot filling. In the slot filling task,the input is natural language and the output is slots which is pre-defined based on the specified intent,such as from location,to location in flight intent. So the paper treats slot filling task as sequence labeling task,then makes experiments with Bi-LSTM-CRF model on slot filling task. The result shows that Bi-LSTM-CRF model has the significant improvement comparing with traditional statistical methods like CRF. Concretely,F1 is used to evaluate the model results.

关 键 词：语义槽抽取 自然语言理解 Bi-LSTM-CRF网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]