DS4A: Deep Search System for Algorithms from Full-Text Scholarly Big Data

While information retrieval systems have shown tremendous improvements in searching for relevant scientific literature, there is still a gap to cater users' ever demanding need to search for specific metadata-related information from full-text publications. In this paper, we present a deep learning-based system that enhances the capability of search mechanisms by classifying algorithm-specific metadata, such as accuracy, precision and recall, and further details such as the datasets that they operate on and the time complexity - from full-text publications. Specifically, in contrast to traditional term frequency-inverse document frequency (TF-IDF) based approach that uses frequent terms as in 'bag of words' models, we first generated a synopsis of the full-text document and then enriched it with sentences that classify as algorithm-specific metadata from full-text to improve algorithmic-specific searching capabilities. These sentences were classified from deep learning based bi-directional long short-term memory network (LSTM) model. Our bi-directional LSTM model outperformed Support Vector Machine (SVM) by 9.46 % with 0.81 F-measure in classifying 37,000 algorithm-specific metadata lines, annotated by four human experts. Finally, we present a case study on 21,940 full-text publications downloaded from the full-text repository of the ACL (https://aclweb.org/) to show the advantages of a deep learning-based advanced searching system over conventional TF-IDF-based (Lucene) text-retrieval systems.