An enhanced LSTM-CRF based approach for Knowledge Integration based Sentiment Analysis

Sentiment analysis is a key element for expert reviews but is at a loss in terms of detection of sarcasm and irony, which distort the actual sentiments. This paper proposes an approach that will fuse syntactic and contextual knowledge to respond to these requirements. Steps for preprocessing include tokenization, stop word removal, lemmatization, followed by feature extraction using BERT embedding. In addressing high-dimensional output, the Relief algorithm is then used to extract the relevant features. The classification model utilizes a combination of Deep Learning for sequence LSTM networks and Maximum entropy methods, i.e., Conditional Random Fields (CRF), and enables sequential dependency learning to give better accuracy. This application of data pertaining to time and context can effectively address subtle nuances such as sarcasm and irony and classify the text as positive, negative, or neutral. It brings several advancements with respect to traditional models, thus improving the understanding of emotion in text.