A Research On Weakly Supervised Relation Extraction

Most contents in real world are unstructured texts,which are difficult to utilize directly.Relation extraction is one of the most important task is organizing the unstructured data into structured ones.The supervised relation extraction deeply relies on the resource exhausted human annotated data,while the weakly supervised learning method,which automatically labels data,is an effective method to alleviate the cost from human annotation.The weakly supervised method faces wrong label problem and relation overlapping.Hereby are listed the major works towards these issues.An end-to-end neural network is utilized to learn relations.The handcraft features are hard to meet the real feature space,while features wrongly generated by NLP tools will lead to error propagation.The end-to-end method breaks the limitation from feature engineering.A hierarchical attention network is proposed to deal with the noises from wrong labels of weakly supervised learning.And a multi-label learning method is designed to solve the problem of relation overlapping.The hierarchical attention mechanism eliminate noises by reducing the weights of wrong labeled samples through learning multiple instances.Besides the inter-class interactions are aggregated to boost model performance.A globally optimized sequence labeling model is adopted to jointly extract entiies and their relations.This method is more effective than the traditional way that pipelines the two tasks seperately,moreover the model performance is improved by the joint way.