|
Distantly supervised relation extraction has been widely used to extract semantic relations from text. However, it suffers from wrong labeling problems and hinders the performance of a model trained on such noisy data. To deal with this problem, previous neural network model assumed at least one instance is true and only selected the most likely one instance in a bag for training, which missed rich information by discarding all other true positive instances. Instead of using the at-least-one assumption, we assume that most of the distantly labeled instances are true positive instances and true positive instances often share the same feature patterns but false positive instances do not. We argue that all instances should be given to the model and leave the model to decide the contribution of different instances. Therefore we propose two kinds of methods to dynamically determine the weights of all the instances and make the neural network model more tolerant to noise. Experiments show that our approach is effective, and it outperforms several competitive baseline methods. |
|
Keywords:Natural Language Processing, Relation Extraction, Distant Supervision |
|