Feature-Based Approach to Chinese Term Relation Extraction

In this paper, we propose a feature-based Chinese term relation extraction approach that combined the advantages of both naive bayes algorithm and perceptron algorithm. A subset of the features was estimated in training data; another subset of the features was trained by discriminative function. The results demonstrate that the proposed hybrid algorithm almost always outperforms the naive bayes algorithms and perceptron algorithms whether the training set is small or not. On the other hand, a novel feature representation was proposed, which included term sequence feature, term appearance features and context information features. Comparing the previous method, long-range dependence was considered in the proposed feature representation, which add the position of feature into vector space model (VSM) and promotes the capability of feature representation. Further, punctuation feature is the important character for terms relation extraction.