Identifying the Question Similarity of Regulatory Documents in the Pharmaceutical Industry by Using the Recognizing Question Entailment System: Evaluation Study.

backgroundThe regulatory affairs (RA) division in a pharmaceutical establishment is the point of contact between regulatory authorities and pharmaceutical companies. They are delegated the crucial and strenuous task of extracting and summarizing relevant information in the most meticulous manner from various search systems. An artificial intelligence (AI)-based intelligent search system that can significantly bring down the manual efforts in the existing processes of the RA department while maintaining and improving the quality of final outcomes is desirable. We proposed a "frequently asked questions" component and its utility in an AI-based intelligent search system in this paper. The scenario is further complicated by the lack of publicly available relevant data sets in the RA domain to train the machine learning models that can facilitate cognitive search systems for regulatory authorities.

objectiveIn this study, we aimed to use AI-based intelligent computational models to automatically recognize semantically similar question pairs in the RA domain and evaluate the Recognizing Question Entailment-based system.

methodsWe used transfer learning techniques and experimented with transformer-based models pretrained on corpora collected from different resources, such as Bidirectional Encoder Representations from Transformers (BERT), Clinical BERT, BioBERT, and BlueBERT. We used a manually labeled data set that contained 150 question pairs in the pharmaceutical regulatory domain to evaluate the performance of our model.

resultsThe Clinical BERT model performed better than other domain-specific BERT-based models in identifying question similarity from the RA domain. The BERT model had the best ability to learn domain-specific knowledge with transfer learning, which reached the best performance when fine-tuned with sufficient clinical domain question pairs. The top-performing model achieved an accuracy of 90.66% on the test set.

conclusionsThis study demonstrates the possibility of using pretrained language models to recognize question similarity in the pharmaceutical regulatory domain. Transformer-based models that are pretrained on clinical notes perform better than models pretrained on biomedical text in recognizing the question's semantic similarity in this domain. We also discuss the challenges of using data augmentation techniques to address the lack of relevant data in this domain. The results of our experiment indicated that increasing the number of training samples using back translation and entity replacement did not enhance the model's performance. This lack of improvement may be attributed to the intricate and specialized nature of texts in the regulatory domain. Our work provides the foundation for further studies that apply state-of-the-art linguistic models to regulatory documents in the pharmaceutical industry.