Sentiment Classification via a Response Recalibration Framework

Probabilistic learning models have the ability to be calibrated to improve the performance of tasks such as sentiment classification. In this paper, we introduce a framework for sentiment classification that enables classifier recalibration given the presence of related, context-bearing documents. We investigate the use of probabilistic thresholding and document similarity based recalibration methods to yield classifier improvements. We demonstrate the performance of our proposed recalibration methods on a dataset of online clinical reviews from the patient feedback domain that have adjoining management responses that yield sentiment bearing information. Experimental results show the proposed recalibration methods outperform uncalibrated supervised machine learning models trained for sentiment analysis, and yield significant improvements over a robust baseline.