抄録
In the past decades, kernel methods are enthusiastically explored for relation extraction. This paper proposes a penalty tree similarity algorithm by extending the dependency tree kernel. Dependency tree kernel computes the similarity of two parse trees by enumerating their matched sub-trees. The penalty tree similarity, however, not only consider the similar structures of the parse trees, but also count in their influences by exploring relative position information between the sub-trees to the target (the entity pairs that generate the tree). An experiment is conducted and the comparison between the dependency tree kernel and the penalty tree similarity is also done. The results show that the former achieves a better performance.
