Harvesting Mature Relation Extraction Models from Limited Seed Knowledge: A Self-Development Framework for DS Rule Expansion

Anonymous

Abstract: Distantly-supervised relation extraction (DSRE) is an effective method to scale relation extraction (RE) to large unlabeled corpora with the utilization of knowledge bases (KBs), but suffers from the scale of KBs and the introduced noise.To alleviate the above two problems, we propose a novel framework called Self-development Rule Expansion (SOUP), which starts from limited amount of labeled data and continuously produces low-noise labels on large-scaled unlabeled data by a growing learnable logical rules set.Specifically, SOUP achieves a mutual enhancement of RE model and logical rules set, first a RE model is trained on the labeled data to summarize the knowledge, then the knowledge is utilized to explore candidate rules from unlabeled data, finally high-quality candidates are selected in a graph-based ranking manner to extend the logical rules set and new rule-labeled data are provided for better RE model training.Experiments on wiki20 dataset demonstrate that, with limited seed knowledge from small-scaled manually labeled data, SOUP achieves significant improvement compared to baselines by producing continuous growth of both logical rules and the RE model, and that labeling noise of SOUP is much less than DS. Furthermore, RE model enhanced by SOUP with 1.6k logical rules learned from prior knowledge could produce an equivalent performance to the model trained on data labeled in DS manner by 72k relational facts of KBs.

Paper Type: long