Domain Adaptation Of Sentiment Classification

As the blogs, product reviews spring up, sentiment classification has become achallenging problem. Sentiment classification, which aims to identify ones’sentimental polarities, is playing an increasing important role in E-commerce andpublic opinion analysis.However, the way of expressing sentiment varies a lot and the datadistributions differ in multiple domains. So sentiment classification tends to beinfluenced by different domains. To solve a sentiment classification problem of anew domain, traditional machine learning methods need to label new training data,which costs a lot of manpower and material resource. Thus, we propose twomethods, LLR based feature selection method and confidence probability basedensemble method, to implement sentiment domain adaptation from aspects offeature space and ensemble strategy.A novel feature selection method, named LTF, is proposed. This methodcreates a common feature space for both source domain and target domain byselecting features, which have sentiment polarities in source domain and importantinfluence in target domain. LTF reduces the data distribution gap between sourcedomain and target domain and prompts knowledge transfer across differentdomains.In terms of ensemble strategy, we propose a confidence probability basedmulti-domain ensemble method, named CEC (Confident Ensemble Classifier). Thismethod utilizes the thought of self-learning and co-training to pre-label the data intarget domain and then ensembles the base classifiers from different domain. Theexperiment results show that CEC actually improves the accuracy of sentimentclassification in target domain.