使用过训练提升词性标注依存句法联合模型的速度

Improving the Efficiency for Joint POS-Tagging and Dependency Parsing with Uptraining

词性标注和依存句法分析是自然语言处理领域中句子级别基本分析技术的两个重要任务,一般来说词性标注是依存句法分析的一个前提条件。基于联合分析的方法将这两个任务在一个统一的统计模型中联合处理能避免错误传播这类问题的发生,因此这种联合模型能取得比较好的性能。但是这种联合模型会带来算法上的时间复杂度的额外开销,因此导致联合分析的方法,速度非常慢。本文提出一种基于过训练的方法,通过极少量的性能损失,使得联合模型的解码速度提升了6倍。

POS tagging and dependency parsing are basic tasks of sentence -level natural language processing. Generally POS - tagging is a necessary prerequisite for dependency parsing. The joint models which link the two tasks together and process them by a unified model have achieved improved performances, because joint modeling can avoid the error - propagation problem. However, the time complexity of joint models can be always so large, thus yields much slower speed. This paper proposes a method based on uptraining technique to improve the speed of joint models, with only very little loss in performances.