A Novel Sentiment Classification Method on Medical Reviews

Abstract

Although sentiment analysis has been studied extensively in other domains, there are only a limited number of works in medical sector. Traditional word embedding models such as word2vec and GloVE are acted as foundation of the feature extraction process in medical sentiment analysis systems. However, traditional word embedding models cannot capture sentiment and domain-specific properties which are important in medical SA. The primary objective of this research is to investigate the importance of sentiment and medical domain knowledge in word embedding models for medical reviews sentiment classification. The main research question is whether the incorporation of sentiment and medical knowledges to the pre-trained word embedding vectors improve the performance of sentiment classification on medical reviews. As the aim of this, a new word embedding model, WE-iMKVec, integrated with sentiment and medical knowledge vectors is proposed. The model is based on original pre-trained word embedding models and these models are enriched with the information from the existing sentiment lexicons, Adverse Drug Reaction (ADR) word lists, and Unified Medical Knowledge System (UMLS) knowledgebases. Convolutional Neural Network (CNN) architecture is used in for sentence classification and the proposed word embedding model is acted as a novel feature extractor in CNN for sentiment classification. The evaluations of the proposed method have been done on five different patient-generated medical reviews dataset collected from the online medical forums and social media. For empirical study, pre-trained word2vec models (Google-News, PubMed-PMC, and Drug Reviews) and one GloVE model (GloVe-Twitter) are used to conduct the experiments. According to the empirical results, our proposed model presents the superior performance in sentiment classification on medical reviews compared to original word embedding models.

Although sentiment analysis has been studied extensively in other domains, there are only a limited number of works in medical sector. Traditional word embedding models such as word2vec and GloVE are acted as foundation of the feature extraction process in medical sentiment analysis systems. However, traditional word embedding models cannot capture sentiment and domain-specific properties which are important in medical SA. The primary objective of this research is to investigate the importance of sentiment and medical domain knowledge in word embedding models for medical reviews sentiment classification. The main research question is whether the incorporation of sentiment and medical knowledges to the pre-trained word embedding vectors improve the performance of sentiment classification on medical reviews. As the aim of this, a new word embedding model, WE-iMKVec, integrated with sentiment and medical knowledge vectors is proposed. The model is based on original pre-trained word embedding models and these models are enriched with the information from the existing sentiment lexicons, Adverse Drug Reaction (ADR) word lists, and Unified Medical Knowledge System (UMLS) knowledgebases. Convolutional Neural Network (CNN) architecture is used in for sentence classification and the proposed word embedding model is acted as a novel feature extractor in CNN for sentiment classification. The evaluations of the proposed method have been done on five different patient-generated medical reviews dataset collected from the online medical forums and social media. For empirical study, pre-trained word2vec models (Google-News, PubMed-PMC, and Drug Reviews) and one GloVE model (GloVe-Twitter) are used to conduct the experiments. According to the empirical results, our proposed model presents the superior performance in sentiment classification on medical reviews compared to original word embedding models.