基于语言模型的中文话头自足句识别方法

NT clause recognition method based on language model

根据话头自足句理论进行标点句的话头自足句识别任务,探究语言模型在该任务中的应用效果,进一步将该任务划分为单个标点句的话头自足句识别和标点句序列的话头自足句识别。在单个标点句的话头自足句识别任务中,对该任务在形式上进行了完整的定义,将其形式化为机器阅读理解任务。从结果来看,在《围城》语料上取得了83.46%的准确率,在"鱼类百科"语料上的准确率达到91.21%。在标点句序列的话头自足句识别任务中,利用单个标点句的话头自足句识别任务中训练好的模型,以流水线的方式进行标点句序列的话头自足句识别,在"鱼类百科"语料上的准确率达到82.03%。实验结果表明,提出的方法能够有效地提高中文话头自足句识别的准确率。

Based on the NT clause theory,this paper carried out the NT clause identification task of punctuation sentences and explored the application effect of the language model in this task.Furthermore,it divided this task into single NT clause identification and sequential NT clauses identification task.In the single NT clause identification task,this paper formalized it into a machine reading comprehension task.From the results,it obtained the accuracy rate of 83.46%on the Fortress Besieged corpus,and reached the accuracy rate of 91.21%in the fish encyclopedia corpus.In the sequential NT clauses identification task,using the single NT model,it fed the punctuation sentences sequence in a pipelined manner and got the sequential NT clauses.The accuracy rate in fish encyclopedia corpus reached 82.03%.The experimental results show that the proposed method can effectively improve the accuracy of Chinese NT clause identification.