Research On Dependency-based Chinese Semantic Role Labeling

Semantic analysis is an important task in Natural Language Processing (NLP), Se-mantic Role Labeling (SRL) is a way to achieve shallow semantic analysis, and it only labels semantic roles which related to predicate of the sentence.Current semantic role labeling mainly focuses on using phrase structure trees. Re-search on constituent structure-based semantic role labeling has been relatively mature, and have achieved good performance. However, the bottleneck problem is increasingly outstanding accompanying with the development of this method, which cause it hard to further improve its performance. Therefore, began to have scholars explore the use of dependency parse tree for semantic role labeling. However the research of Chinese do-main limited by some factors such as the size of corpus, the development was not fast as other languages. So, we start out from dependency parse tree, to research on Chinese semantic role labeling.The main contents are as follows:Firstly, we build a baseline system in this paper, which includes predicate labeling (PL), semantic role labeling and evaluation system in three parts. Predicate labeling is divided into two parts:predicate identification and predicate classification, semantic role labeling is also divided into semantic role identification and semantic role classification. The following research are all based on this baseline system.Secondly, we analyzes the features which used in baseline system, and more targeted to get more effective features. Because of features has always been an important factor to determine semantic role labeling system performance. Whether we can discover more effective features directly determines system performance.Finally, we redefine semantic role classification as sequence labeling through theo-retical analysis in this paper, and replace maximum entropy classification model with conditional random fields (CRF) model. At the same time, according to the error analysis, we propose to add null category to semantic role classification model, so that this stage has to be able to rectify some semantic role identification phase remnant of errors.The contribution of our work includes the following:1) research and build a seman-tic role labeling baseline system using Chinese dependency parsing tree;2) to improve system performance via add new features;3) by redefining semantic role classification as sequence labeling and adding null category, to improve the performance of the system. The final results show that the methods that we used in this paper can be greatly improve the performance of semantic role labeling. The experiments on automatic predicate label-ing show our system can achieve85.79%/85.85%/85.82on precision, recall, and labeled F1score.