Building automated text classifiers have assumed significant importance since the development of large online information platforms. Several compelling use cases have emerged in the field of artificial intelligence and analytics in recent years. However, building and training text classifiers become problematic in the healthcare context, which deals with a sensitive and limited volume of data. In this paper, we explore the development of a classifier and apply it to a specific case of classifying physician reviews into either clinical and non-clinical reviews. The primary purpose of this paper is to demonstrate the methodology using which the classifier has been developed, including a novel technique in curating datasets.

We leverage unsupervised guided Latent Dirichlet Allocation (LDA) method and supervised methods such as deep neural networks, Long-Short Term Memory (LSTM) networks, and Bi-directional LSTMs. Further, we compare the various models and choose the one with the best classification performance by validating the output results with the ground truth. Our methodology provides insights into making the best use of semi-supervised and supervised algorithms along with grounded data for developing classifiers that can be generalized for other novel contexts where dataset availability is limited.