A Neuro-Symbolic Approach for Question Answering on Scholarly Articles

The number of research articles is increasing exponentially. It has become difficult for researchers to stay updated with the latest development in science with the deluge of papers. Hence, keeping abreast with the current literature is one of the most significant challenges to present-day researchers. However, if one can query a scientific article, they can quickly comprehend it and elicit the required information. Hence, a question-answering (QA) system on scholarly articles would be a helpful assistant for researchers to survey the literature. Recently logic-infused deep networks have been showing good promise for solving several downstream NLP tasks. Here in this paper, we implement a neural network-based symbolic approach for QA on scholarly articles. We incorporate logical boolean functions into the deep network, significantly improving the model’s performance without additional parameters. Further, we reduce the dependency on domain-specific training data by using external knowledge from the ConceptNet. We perform our experiments on the benchmark ScholarlyRead dataset and achieve significant performance improvement (∼ double) over the baseline approach.