Learning to Identify Subjective Sentences

Subjective sentences describe people’s opinions, points-of-view, interpretations, comparisons, sentiments, judgments, appraisals or feelings toward entities, events and their properties. Identifying subjective sentences is the basis for opinion mining and sentiment analysis and is important in applications like political analysis, social media analytics and product review analytics. We use standard classifiers to build models of SUBJECTIVE vs. NON-SUBJECTIVE sentences and demonstrate that they outperform the approaches reported in the literature on several interesting datasets. We discuss two novel applications of this work: prevalence of subjective sentences in performance appraisal text and scientific papers. We demonstrate that scientific papers also contain a substantial fraction of subjective sentences. We compare the nature of subjective sentences in performance appraisals text and scientific papers, and observe different reasons why some sentences in these domains are subjective. We propose the need to further investigate the linguistic and semantic basis for subjective sentences across different domains.