Abstract:
"Traditional sentiment analysis techniques often fail in analysing social media text reviews
due to the unstructured, context-rich nature of the reviews. Therefore, this research aims to explore the challenges involved in sentiment analysis of social media text reviews. Since the unstructured format and rich contextual depth of social media content cause complex problems, the project aims to significantly improve sentiment analysis accuracy, particularly for unstructured, rich contextual social media text reviews.
The research proposes a novel approach to tackle challenges related to sentiment analysis.
It combines the Lexicon-Enhanced BERT model with Long Short-Term Memory (LSTM)
networks. This is achieved by splitting the input text into sections using n-grams, and the section that contains the word conveying the sentiment is identified using a sentiment lexicon. This is then merged with the BERT embedding layer and passed into the LSTM network to construct and train the model. This novel approach aims to overcome the limitations found in current sentiment analysis models and emphasises the need for exploring different neural network architectures like LSTM. The primary objective of this approach is to enhance the accuracy and scope of text reviewsby focusing on the sequential and temporal aspects of text sentiment.
The study’s tests are promising. Metrics such as f1 score have demonstrated significant
improvements in the LSTM-enhanced LeBERT model through iterative training, testing and
evaluation. It returned a maximum f1 score of 89.48% which outperformed all the baseline model and the LeBERT with CNN model. These enhancements imply that this model has a greater capacity to identify and interpret the subtle emotions contained in text reviews, which is a significant development in the field of sentiment analysis."
