Sentiment Analysis using Novel Distributed Word Embedding for Movie Reviews

Sentiment analysis is the recent trend explored to provide the insight of public and consumer's emotions, opinion and attitude to product developers, governments, marketing organizations, political organizations etc. Feature engineering is an essential requirement for Machine Learning (ML) based sentiment analysis. Word embedding is set of techniques to map the row text word as a real-valued vector which can be utilized as feature vector in sentiment analysis. Word2vec is a powerful and efficient word embedding technique to preserve the semantic relationship of words in low-dimensional embedding space. It can effectively handle small text corpus with few millions of unique words (also called as vocabulary). Whilst, real-life applications like social media, search engine, etc. yield voluminous text data consisting of large vocabulary. Word2vec has demonstrated poor performance due to in-memory processing of entire vocabulary and associated word embedding. It also demands the huge computation capability due to numerous mathematical calculations. Scalability and processing complexity make the commodity hardware incapable for word2vec of voluminous text data. This paper presents novel and scalable technique namely MapReduce framework for Word2Vec (MRW2V), amalgamates the computational resources from cluster of commodity machines to accomplish the above essentials. MRW2V explores the Hadoop distributed environment to generate the feature vectors from IMDB movie review dataset. Various ML algorithms are applied over the generated feature vectors to learn the sentiment models. It is observed from experiments that MRW2V is well suited word embedding framework for voluminous data which also sustaining the admissible performance.