Research On Automatic Semantic Role Labeling Of Chinese FrameNet

Currently, the biggest obstacle of natural language processing techniques is automatic semantics analysis. But semantic role labeling is a realization way of shallow semantic analysis, which has been successfully applied to question answering system and machine translation, etc. Whether the semantic role labeling is appropriate would affect modeling method of semantic analysis directly, as well as the accuracy and robustness of application systems, such as question answering system and machine translation.The innovation of this paper is to use the new semantic features, which is the first time to add the features to semantic role labeling of Chinese FrameNet respectively, the features including Tongyici Cilin information and distributed word representation, which trained by the method of deep learning. In this paper, the task can be seen as a sequence labeling tasks which regards word as label unit. With using some basic features such as word and part of speech, we research the influence of Tongyici Cilin information features and the distributed word representation features on performance of semantic role labeling in Chinese FrameNet.In this paper, we research the semantic role labeling based on the features of Tongyici Cilin information. The word feature is very important feature in semantic role labeling, but the sparsity of word feature has a great impact in performance of model. This paper adds some Cilin information to former features such as word, part of speech, position of target word and so on, and uses the semantic knowledge of Chinese FrameNet (CFN) which was self-developed by Shanxi University as our corpus to study the influence of different Cilin information on semantic role boundary identification and classification. The results show that Cilin information could improve the performance of semantic role labeling, especially for the semantic role classification.In this paper, we also research the semantic role labeling based on the distributed word representation. Because of the relatively small size of corpus, the features defined manually are difficult to cover all words, so using this features to improve the result of semantic role labeling is relatively limited. In today’s information explosion age, using information of big data effectively will be unprecedented convenient, deep learning is based on this concept, which uses unsupervised methods to learn features through the large natural text first, then adds the features (called distributed word representation) into the machine learning models to supervised learning a new model. In this paper, we learn the word distributed representation by the method of deep learning, and then based on the distributed word vector representation to build new feature template, and use the model of CRFsuite to study the automatically semantic role labeling in Chinese FrameNet. Because it is the first explore, the results of experiments increased only on part of frames. The reasons of the phenomenon might be two points:the scale of corpus and parameter adjustment, we must focus on these question in the future.