Sentence-based undersampling for named entity recognition using genetic algorithm

Abstract

Named entity recognition (NER), as one of the crucial tasks of information extraction (IE), has important effect on the quality of its subsequent applications such as answering the question, co-reference resolution, relation discovery, etc. NER can be considered as a kind of classification problem, which has to deal with its own challenging issues. Class-Imbalanced Problem (CIP) is one of the important problems in classification domain from which almost all NER tasks also suffer, because usually, the number of entity mentions of interest in the given text is much less than undesired entities. The quality of the IE’s subtasks for which NER is the basis is directly affected by any improvement on the performances of NER systems. In this research, an effort has been made to increase the overall performance of NER systems by decreasing the curse of CIP as much as possible. A new heuristic approach based on the genetic algorithm has been devised to undersample the training data which is used for NER. Regarding the fact that given training patterns for NER are of individual sentence forms, in the developed approach, this issue is considered as well and it was applied to individual sentences from training data. The proposed method has been applied on two different corpuses: CoNLL corpus from newswire domain and JNLPBA from biomedical context to see its impact on different type of contexts. By increasing the performance in terms of F-score for both data sets, our proposed method outperforms the baseline systems using original data. Furthermore, in comparison with random undersampling, it results in better outcomes. In addition, the effect of considering sentences of training data individually in sampling process and taking all of them together has been investigated.