Active Learning For Chinese Frame Semantic Roles Labeling

Semantic role labeling is one of the key in the field of natural language processing research problems.As a kind of implementation method of shallow semantic analysis,it can directly services in automatic question answering and machine translation,information extraction,and other related tasks in the field of natural language processing.In this paper,with the aid of Chinese FrameNet semantic knowledge resources,we studyed the Chinese semantic role labeling,namely using conditional random field model and first introduced dependent path for feature selection problem to increase the rate of recall.On the basis,active learning is applied to construction of Chinese FrameNet sentences library.The current Chinese Frame semantic role labeling study is given priority to the supervised machine learning methods,so a certain size and manual annotation examples of high quality is needed as the training corpus.But the current Chinese FrameNet sentences resources is relatively small and the manual annotation cost high.Then in this article,active learning method is introducted to reduce the costs of the construction of Chinese FrameNet sentence library.The characteristic of path was added to Chinese Frame semantic role labeling in the paper,and the effect of different paths on the results was analyzed.Then it introduces how the active learning is applied to construction of Chinese FrameNet sentence library,so as to reduce the labor costs of building sentence library and improve Chinese sentence semantic role labeling performance.This active learning take uncertainty of sampling and voting of the committee as the method of determining confidence.The method choose the examples of the labeling models predict poor,in order to make the labeling model achieve the same results only need with less training.This paper proposes and compares four tradeoff rules of predicts credibility of Chinese Frame semantic role labeling model.Experimental results show that the addition of a feature path,effectively increasing the rate of recall of the Chinese Frame semantic role labeling.And the increases of recall rate play an important role to the active learning.About active learning methods,uncertainty simpler and more efficient method than the committee vote.But both obtain better results compared to passive learning methods.First of all,it makes Chinese frame semantic role labeling can be a 30 percent reduction in time to achieve the same results up to the amount of manual annotation;Second,Compared with passive learning random selection,when using the same amount of training sentences,active learning has improved greatly than passive learning,the highest performance upgrade 5.07 percentage points.