Generating Word and Document Embeddings for Sentiment Analysis

AbstractSentiments of words can differ from one corpus to another. Inducing general sentiment lexicons for languages and using them cannot, in general, produce meaningful results different domains. In this paper, we combine contextual supervised information with the semantic representations occurring dictionary. Contexts help us capture domain-specific scores are indicative polarities those words. When features extracted their dictionary definitions, observe an increase success rates. We try out combinations contextual, supervised, dictionary-based approaches, generate original vectors. also word2vec approach hand-crafted features. induce sentimental vectors two corpora, which movie domain Twitter datasets Turkish. thereafter document employ support vector machines method utilising vectors, our approaches perform better than baseline studies Turkish a significant margin. evaluated models on English corpora as well these outperformed approach. It shows that cross-domain portable other languages.KeywordsSentiment analysisOpinion miningWord embeddingsMachine learning