Research on Relation Extraction Method Based on Active Learning

Abstract

The knowledge in contemporary society has exploded, and the most common knowledge is contained in unstructured natural language texts. Information Extraction technology expresses semantic knowledge in unstructured text through a set of mentioned entities, the relationships between these entities, and the events in which these entities participate. As a key part of information extraction, Relation Extraction technology provides important theoretical basis and use value for text knowledge understanding by judging the relationships between given entities. Currently, relationship extraction based on supervised learning requires a large number of labeled samples. Randomly selecting some data labels is not only a waste of data resources, but also directly affects the final accuracy of the classification model. In fact, with the development of data collection and storage technology, it has become very easy to obtain a large amount of unlabeled natural language text. Therefore, it is of great practical value to design an algorithm that can effectively utilize unlabeled sample sets for relationship extraction. In order to solve the above problems, this paper uses active learning as the starting point to implement a variety of sampling algorithms, mainly including uncertainty, diversity, representativeness and other algorithms. On the basis of verifying that active learning is suitable for relationship extraction tasks, through the fusion of multiple This sampling criterion ultimately yields an active learning sample selection strategy that is still effective under multiple data sets and multiple learning models. Experiments have proven that the multi-criteria fusion sampling strategy proposed in this article is an effective and universal strategy. Compared with multiple single-strategy sampling algorithms, it can achieve equivalent or higher classification accuracy on multiple data sets.