Active learning approach using a modified least confidence sampling strategy for named entity recognition

One of the important subtasks of information extraction is named entity recognition (NER). Its aim is to identify and to classify the named entities in the textual data into predetermined categories. There are a large number of supervised learning and deep learning models being developed for the entity recognition task, which performs well in the presence of a labeled training set. The availability of the labeled training set requires the labeling of large unlabeled data, which is both expensive and time taking. Active learning is an iterative approach that provides a way to minimize labeling cost without affecting performance. This approach uses a sampling strategy that selects the appropriate unlabeled data instances, an oracle to label the selected data instances, and a machine learning model (base classifier). In this work, a modified least confidence-based query sampling strategy for the active learning approach for named entity recognition task has been proposed, which considers different numbers of uncertain words present within the sentences to compute the final least confidence score of the sentence for comparison. To evaluate the effectiveness of the proposed approach, the comparison of the performance is made among the active learning approaches with the proposed sampling strategy, random sampling strategy, and two other well-known existing uncertainty query sampling strategies. Real-world scenario for active learning approach is simulated for experiment, and the total amount of labeled data required for training of active learner to reach the stop condition while using different sampling strategies is recorded. The experiment is carried for the development and the test set of the three different biomedical corpora and a Spanish language NER corpus. It is found that with the proposed active learning approach, there is a minimal requirement of labeled data for training to reach the above performance level in comparison with the other approaches. The performance of the proposed approach is found to be slightly better than the existing sampling approach, and the performance of all the approaches is far better than the random sampling approach.