Cascade Neural Ensemble for Identifying Scientifically Sound Articles
This work addresses a critical bottleneck in biomedical literature review by improving filtering accuracy, though it is incremental as it builds on existing pre-trained models and ensemble techniques.
The paper tackled the problem of efficiently identifying scientifically sound articles for systematic reviews, which is highly imbalanced with less than 1% relevant articles, by proposing a cascade ensemble of SciBERT models that achieved an F-measure of 0.7639 and reduced error rates by up to 49.1% compared to previous methods.
Background: A significant barrier to conducting systematic reviews and meta-analysis is efficiently finding scientifically sound relevant articles. Typically, less than 1% of articles match this requirement which leads to a highly imbalanced task. Although feature-engineered and early neural networks models were studied for this task, there is an opportunity to improve the results. Methods: We framed the problem of filtering articles as a classification task, and trained and tested several ensemble architectures of SciBERT, a variant of BERT pre-trained on scientific articles, on a manually annotated dataset of about 50K articles from MEDLINE. Since scientifically sound articles are identified through a multi-step process we proposed a novel cascade ensemble analogous to the selection process. We compared the performance of the cascade ensemble with a single integrated model and other types of ensembles as well as with results from previous studies. Results: The cascade ensemble architecture achieved 0.7505 F measure, an impressive 49.1% error rate reduction, compared to a CNN model that was previously proposed and evaluated on a selected subset of the 50K articles. On the full dataset, the cascade ensemble achieved 0.7639 F measure, resulting in an error rate reduction of 19.7% compared to the best performance reported in a previous study that used the full dataset. Conclusion: Pre-trained contextual encoder neural networks (e.g. SciBERT) perform better than the models studied previously and manually created search filters in filtering for scientifically sound relevant articles. The superior performance achieved by the cascade ensemble is a significant result that generalizes beyond this task and the dataset, and is analogous to query optimization in IR and databases.