Research Of Chinese Core Frame Semantic Parsing Technology

As a significant research project in Natural Language Processing Fileds, semantic parsing has attracted much attention at home and abroad in recent years. However, neither one of these approaches was able to achieve the purpose of in-depth parsing, to analysis a sentence deeply and automaticly. Nowadays, although we have acess too many smart products such as Siri, machine translation and information retrieval etc, but these processing technics are incapable of catching up what we really want they do, to be more like human beings. With the aid of Chinese FrameNet based on the theory of frame semantics, we studied the technology of Chinese core frame semantic pasing, which is extract the core frame semantic representation of a Chinese sentence to analyze the semantic content.We solve the problelm by using a three-stage learning model. Taking the tasks’different characteristics into consideration, we choose Maximum Entropy model to take core target in the sentential contexts and predict frame for the core target, choose Conditional Random Fields model to label the frame elements defined in Chinese FrameNet.The main research contents and research results are as follows:(1) For the task of core target identification, the paper proposed a method combining rule-based and statistic-based approaches.We used the synonyms dictionary to get lexical level features and syntactic level features, we also provided a new feature selecion algorithm based on mutual information to performance of the core target identification.(2) Given a target in context,the second stage is disambiguating it to a semantic frame.This model uses Maximum Entropy model to allocate a appropriate frame for the ambiguous target.We organized 47 Common ambiguous target words corpus, by adding long dependency relations in dependency tree and synomyms dictionary features,we set a relatively well frame disambiguation model compared with others.(3) The third stage is find the locally expressed semantic arguments for the target.We proposed a method based on CRFs for frame element annotation.We conversed the task into a sequential taging problem at word-level.We totally extracted 36 features templates combined with open windows strategy.By analyzing the influence of different feature templates, we constructed an optimal model for frame element annotation.We have presented an approach for Chinese core semantic parsing based on a combination of knowledge frome Chinese FrameNet, three probabilistic models trained on full text annotations corpus and used an expedient heuristic feature selection algorithm.We also built a small data set from People’s Daily to test our model and received a performance result.The experimental results validate the importance and effectiveness of core semantic parsing for semantic parsing research.