Combining Feature-based Kernel with Tree kernel for Extracting Relations

This report proposes a composite kernel method of semantic relation extraction between named entities within natural language documents. Using two kernel methods, the benefits from both are combined to gain an increase in performance compared to earlier approaches. 1) A linear kernel processing linguistic features, such as word-span, orderof-entities and word-type. 2) A tree kernel computing the similarity of a relation type with the shortest path-enclosed tree between a pair of candidate entities. Experiments are done using previous implemented methods, such as context sensitiveness and latent annotations to measure their impact on the performance. Evaluating on a dataset for the relation extraction task at the Conference on Computational Natural Language Learning from 2004, the results obtained are on par with previous state-of-the-art approaches on the same dataset.