Aspect Based Sentiment Analysis Using Self-Attention Based LSTM Model with Word Embedding

Abstract

Sentiment analysis is in advance more attention for research due to incremental convention in affairs of state, online marketing and social networking. Users afford their opinions for exacting object as reviews. Analysts categorize these reviews into prejudiced information. Consequently sentiment pulling out is the development of take out human awareness from amorphous text reviews. Extracted sentiment will assist to recognize in general appropriate polarity of users towards scrupulous object or event. Reviews can be classified by earnings of learning models such as ANN, SVM etc. Deep neural network is a detachment of machine learning. LSTM is artificial recurrent neural network architecture, skillful of acquaintance long term dependencies. LSTMs have received more accomplishment when working with succession of words and paragraphs, normally Natural Language Processing. Current attention methods planned for aspect based sentiment classification for the most part focal point on distinguish the sentiment words, without in apparition of the significance of such words with esteem to the individual aspects in the sentence. To solve this difficulty, paper proposes a new architecture, which will be using self-attention mechanism to trounce weak point of LSTMs and modify LSTM activation function. Word embedding is method in NLP wherever phrases or words from transcript are record to vectors of actual numbers. Suggest model captures the significance of each word of documents using state of the art word embedding and Bidirectional LSTM. This study evaluated the proposed approach on benchmark dataset of SemEval, experimental results make obvious that propose model outperforms on SemEval dataset.