Winner-Takes-All based Multi-Strategy Learning for Information Extraction

Abstract

This paper proposes a winner-takes-all based multi-strategy learning for information extraction. Unlike the majority of multi-strategy approaches that commonly combine the prediction of all base learnings involved, our approach takes a different strategy by employing only the best, single predictor for a specific information task. The best predictor (among other predictors) is identified during training phase using k-fold cross validation, which is then retrained on the full training set. Empirical evaluation on two benchmarks data sets demonstrates the effectiveness of our strategy. Out of 26 information extraction cases, our strategy outperforms other information extraction algorithms and strategies in 16 cases. The winner-takes-all strategy in general eliminates the difficult situation in multi-strategy learning when the majority of base learners cannot make correct prediction, resulting in incorrect prediction on its output. In such a case, the best predictor with correct prediction  in our strategy will take over for the overal prediction.