Dynamic task balancing for joint information extraction

Joint Information Extraction (IE) aims for joint extraction of various semantic structures such as entities and relations. Most recent joint IE works use static weighting methods by combining task losses with predefined and fixed weights. In this paper, we identify the limitations of the static weighting methods with empirical analysis. We then study the feasibility of applying several dynamic weighting methods for the joint IE problem and evaluate the methods on three benchmark IE datasets in terms of their performance. We find that existing dynamic weighting methods can achieve reasonably good results in a single run, demonstrating their effectiveness and advantages over the static weighting methods. Further, we propose a hybrid dynamic weighting method, Adaptive Weighting for Joint IE (AWIE), based on gradient dynamic task weighting. Experimental results show that our proposed method obtains competitive performance results across datasets cost-effectively with task preference accommodation.