Abstract

Sentiment analysis has been quite a critical task in natural language processing or NLP which is used in the classification of opinions or sentiments in texts as positive, neutral or negative. But there are few challenges in understanding sarcastic, ambiguous and implicit texts, which usually limit the performance of traditional models like Naïve Bayes, SVM, Machine Learning, LSTM. Here we have taken social media and news into consideration, to show how different news or posts on twitter on Facebook can be analysed to extract information about the sentiments or opinion of the user even if it is sarcastic or ambiguous in nature, which can be a very challenging task. This paper proposes a novel framework using Hidden Markov Model with Contextual Embedding Layers using BERT (HMM-BERT), to address these said challenges. The unique combination of deep contextual understanding provided by pre-trained contextual embeddings like BERT with the sequential pattern learning of Hidden Markov Models enhances sentiment prediction by integrating sarcasm detection and word sense disambiguation techniques to manage complex contextual relationships within texts, and achieves an accuracy of 89-95% which is almost 10-20% better than the traditional and base models Thus, this method marks a considerable improvement in the sentiment analysis of ambiguous and sarcastic texts.

Authors

Saberi Goswami, Supratim Bhattacharya, Jayanta Poray
Techno India University, India

Keywords

Hidden Markov Model, BERT, Contextual Embeddings, Sentiment Analysis, Sarcasm Detection