SENTIMENT CLASSIFICATION USING PS-POS EMBEDDING WITH BILSTM-CRF AND ATTENTION

Abstract

Sentiment analysis uses natural language processing and machine learning to
interpret and classify emotions in subjective data. Sentiment analysis is often used in business to
detect sentiment in social data, gauge brand reputation, and understand customers. In this paper a
novel target extraction which is performed by a combination of two methods PS-POS (Predecessor
Successor-Parts of Speech) and character based BiLSTM-CRF (Bidirectional Long Short Term
Memory - Conditional Random Field). Initially, a PS-POS word embedding layer is built to integrate
the features of the sentimental words into the word vector during the word vector training and then
BiLSTM-CRF attention layer is dedicated to extract the targets in the text as single target, multi
target and no target. Finally the 2D-CNN method was introduced to classify the targeted words. The
word vector is used characterize the emotional words as input, and the output represents the
different stress level of the information. According to this new model, a text is classified as positive,
very positive, negative, very negative and neutral based on semantic similarity. The simulation result
shows that the proposed method is more effective than the existing method in considering the relation
between features in the text. Compared with state-of-the-art of network methods, the target
classification performance of developed model is greatly enhanced.