Chinese Frame Identification using T-CRF Model

As one of the important tasks of SemEval Evaluation, Frame Semantic Structure Extraction based on the FrameNet has received much more attention in NLP field. This task is often divided into three sub-tasks: recognizing target words which are word expressions that evoke semantic frames, assigning the correct frame to them, namely, Frame Identification (FI), and for each target word, detecting and labeling the corresponding frame elements properly. Frame identification is the foundation of this task. Since the existence of links between frame semantics and syntactic features, we attempt to study FI on the basis of dependency syntax. Therefore, we adopt a tree-structured conditional random field (T-CRF) model to solve Chinese frame identification based on Dependency Parsing. 7 typical lexical units which belong to more than one frame in Chinese FrameNet were selected to be researched. 940 human annotated sentences serve as the training data, and evaluation on 128 test data achieved 81.46% precision. Compared with previous works, our result shows obvious improvement.