Sentence Factorization for Opinion Feature Mining

Opinion mining has tremendous potentials in extracting valuable information and experience from individuals on products and services. In particular, product features extraction and sentiment scoring on extracted features are fundamental steps. Opinion knowledge extraction often involves extensive application of natural language processing, manual labeling and machine learning methods.In this paper, we focus on developing fine-grained product feature extractions with minimal tailor build language models and labeling.A threshold-normalized sentence-level word model is proposed for opinion feature mining. The opinion feature extraction is then solved via matrix factorization technique. Evaluation on feature-entropies, sentence-entropies and human evaluation demonstrated the superiority of our approach. Highly relevant and fine-grained opinion features are extracted automatically.