Chinese Semantic Role Labeling Based On Conditional Random Fields

In recent years, semantic role labeling applied more and more widely in information retrieval, question answering system and other natural language processing areas, but the majority of the studies aimed at the English, other languages, because of the lack of well-marked corpus, the study was relatively small.Not long ago, the University of Pennsylvania released a manual tagging Chinese corpus:Chinese Propbank. This article, we try to use the Conditional Random Fields machine-learning algorithm for Chinese Propbank 1.0 to do a supervised learning. In addition to using some basic features of the existing, we also proposed four new features, and through a large number of experiments to verify the validity of each feature。At last, we can see that each new feature play a positive role to the results in experimental.In order to shorten the training time and improve results, we have adopted a two-stage classification method.In the first phase, we carried out the dual classification, determine whether a syntactic component is a semantic role, then do multi-classification to assign a specific semantic role.Taking into account the accuracy and recall rate of the first phase:binary classification, we select the appropriate value as the threshold probability to carry out the non-semantic-role filtering, which to some extent, to recall some of the elements which are wrong-predicted to improve the final results.Ultimately, we selected a large number of thresholds, from the outcome, when the threshold is 0.5,its results can be achieved the best:accuracy rate is over 80%,F-score is also close to 80%.This is a satisfactory outcome in the Chinese semantic role labeling area, we believe that if more distinguishing features can be able to be joined, experimental results will be further enhanced.