Combining Structured and Flat Features by a Composite Kernel to Detect Hedges Scope in Biological Texts

To distinguish factual and uncertain information in biological texts, hedged information detection has received considerable interest in the biomedical natural language processing, which remains a challenging task due to the complexity of the syntactic and semantic analysis. This paper presents an approach to hedges scope detection using a composite kernel which combines structured and flat features. The composite kernel consists of two individual kernels: a polynomial kernel that exploits the flat features widely used in hedges scope detection and a tree kernel that captures the syntactic structured features. Four structured features over a parse tree are explored for hedges scope learning to investigate the effect of the structured features. Experiments on the CoNLL-2010 evaluation data show that our model achieves F-scores of 87.34% on hedge identification and 57.47% on scope detection respectively, which are better than those of the previous reported systems. The analysis results show that structured syntactic features with the tree kernel is more effective for hedges scope detection than the traditional flat syntactic features without the labor of detailed features designing.