Hybrid Sentiment Analysis for Drug Efficiency Determination: Integrating Lexicon-Based and BERT Models with Sarcasm Detection

1. Introduction Purpose and Importance Focuses on sentiment analysis of drug reviews from healthcare forums using a hybrid deep learning and lexicon-based approach. Highlights the importance of drug safety monitoring after market release. Emphasizes challenges in sentiment analysis: Learning-based models (e.g., BERT) need labeled data. Lexicon-based models may not generalize well to medical reviews. Proposes a hybrid BERT + Lexicon approach with majority voting. Research Objectives Develop a hybrid sentiment classification model combining BERT and lexicons. Use majority voting to determine final sentiment labels. Compare performance of BERT, lexicon-based models, and the hybrid approach. 2. Methods (Methodology) 2.1 Dataset Description Source: Kaggle (UCL drug review dataset). Size: 161,297 drug reviews. Attributes: Drug name Condition (medical issue) Review (text-based opinion) Rating (1-10 scale) Date Useful count (number of helpful votes). Challenge: Reviews are unlabeled, requiring sentiment annotation. 2.2 Preprocessing Steps Prepares text data for deep learning models: Lowercasing – Converts text to lowercase. Tokenization – Splits text into words for BERT processing. Removing punctuation and stopwords – Cleans unimportant words. Stemming/Lemmatization – Reduces words to root forms. Padding and Truncation – Ensures uniform text length for BERT. 2.3 Sentiment Labeling (Lexicon-Based Approach) Since the dataset is unlabeled, sentiment scores are assigned using three lexicon-based methods: TextBlob – Scores words using polarity-based sentiment dictionaries. VADER – Detects sentiment in short texts (e.g., social media). AFINN – Assigns integer scores (+5 for very positive, -5 for very negative). Each word receives a sentiment polarity score: Positive: Score > 0 Negative: Score < 0 Neutral: Score = 0 2.4 Deep Learning Model: BERT Uses pre-trained BERT model for contextual sentiment analysis. Converts logits to probabilities using softmax. Selects the highest probability class (torch.max(probabilities, 1)). 2.5 Hybrid Model with Majority Voting If BERT and Lexicon agree, that sentiment is chosen. If they disagree, BERT’s prediction is used. Evaluated using accuracy, precision, recall, and F1-score. 3. Results and Discussion 3.1 Sentiment Labeling Results AFINN assigns more negative labels due to its dictionary structure. TextBlob provides a balanced distribution. VADER is better for intensity-based sentiment detection. 3.2 Model Performance (Accuracy, Precision, Recall, F1-Score) Model Sentiment Lexicon Accuracy Precision Recall F1-score BERT Model – 92% 0.92 0.91 0.92 Lexicon Model (TextBlob + VADER + AFINN) – 85% 0.84 0.85 0.85 Hybrid Model (BERT + Lexicon Voting) – 95% 0.95 0.94 0.95 Hybrid Model outperforms individual models (BERT: 92%, Lexicon: 85%, Hybrid: 95%). BERT alone struggles with certain lexicon-based sentiment nuances. Majority voting helps correct misclassifications in edge cases. 3.3 Confusion Matrix Analysis Hybrid Model achieves highest TP and TN rates. BERT misclassifies neutral sentiments more often than Hybrid. Lexicon-based models struggle with context-dependent sentiment. 4. Figures and Tables in the Paper Key Figures: Figure 1 – Preprocessing Flowchart (Steps for cleaning text data). Figure 2 – Hybrid Model Architecture (BERT + Lexicon integration). Figure 3 – Accuracy Comparison Graph (Lexicon vs BERT vs Hybrid). Figure 4 – Confusion Matrix Heatmap (Visualizes classification results). Key Tables: Table 1 – Dataset Attributes (Lists dataset features). Table 2 – Lexicon Sentiment Scores (AFINN, VADER, TextBlob scores). Table 3 – Performance Comparison (Accuracy, Precision, Recall). 5. Conclusion and Future Work Key Findings: The hybrid approach (BERT + Lexicon) improves accuracy by 3-10% over individual models. BERT struggles with highly polarized words, which lexicon models assist with. Majority voting fusion leads to more balanced sentiment classification. Limitations & Future Research: Weighted Voting – Give higher weight to more confident models. Sarcasm Detection – Address cases like “Great, another side effect!”. Context-Aware Analysis – Consider LSTMs or fine-tuned BERT models. Final Thoughts Your research successfully integrates BERT and Lexicon-based sentiment analysis using majority voting fusion. The results demonstrate higher accuracy (95%) compared to standalone models, making this hybrid approach highly effective for analyzing drug reviews. 🚀