EMOTIBERT: AN ENHANCEMENT OF SENTIMENT ANALYSIS FOR USER REVIEWS

he explosive growth of e-commerce platforms has led to a massive surge in
user-generated product reviews, making manual evaluation increasingly
impractical for consumers. Many existing sentiment analysis methods struggle to
interpret complex linguistic phenomena such as negation, sarcasm, blended
opinions, and subtle emotional cues that commonly appear in user-writtenfeedback. This paper presents EmotiBERT, an improved review analysis
framework that extends the BERT (Bidirectional Encoder Representations from
Transformers) model by integrating Syntax-Enhanced BERT (SynBERT) with
affective emotion embeddings derived from domain-oriented lexicons.
EmotiBERT applies dependency parsing to model syntactic relationships
between words and incorporates emotion-aware token representations to
strengthen semantic understanding. These enriched embeddings are combined
using attention driven fusion, enabling the model to dynamically focus on tokens
that are syntactically and emotionally significant during sentiment prediction. The
proposed framework is implemented as a secure web-based application where
users can search for products and obtain automated purchase recommendations
generated from aggregated review sentiment. Experimental results on benchmark
datasets- Amazon Product Reviews, IMDb, and Yelp-show that EmotiBERT
achieves classification accuracy above 92%, consistently surpassing standard
BERT, BILSTM, and traditional lexicon-based approaches.