Exploiting classification correlations for the extraction of evidence-based practice information

Crucial study data in research articles, such as patient details, study design and results, need to be extracted and presented explicitly for the ease of applicability and validity judgment in evidence-based practice. To perform this extraction, we propose to use two soft classifications, one at the sentence level and the other at the word level, and exploit the correlations between them for better accuracy. Our evaluation results show that propagating the results from the first classification to second improves performance of the second and vice versa. Moreover, the two classifications may benefit each other and help improve performance through joint inference algorithms. Another key finding of our work is that irrelevant sentences in the training data need to be properly filtered out; otherwise they compromise system accuracy and make joint inference models less scalable and more expensive to train.