Concept Level Sentiment Analysis of Bengali Text

Overview

This project explores the concept-level sentiment analysis of Bengali text, focusing on restaurant reviews, comments, and blogs. Unlike traditional feature-based sentiment analysis, this research introduces a concept-level approach using custom-built dictionaries to identify and classify sentiments more accurately.

---

Key Features

• Data Collection:

  • Collected 2,700 Bengali text reviews, comments, and blogs from Facebook.
  • Labeled data manually into positive, negative, and neutral sentiment categories.

• Preprocessing Workflow:

  • Punctuation Removal: Removed unnecessary symbols and spaces.
  • Tokenization & Detokenization: Processed text using nltk for splitting and rejoining meaningful words.
  • Stopword Removal: Eliminated uninformative words like "অথচ," "অথবা," and "এটি."
  • Concept Construction: Built custom dictionaries for nouns, adjectives, verbs, and adverbs to extract sentiment-bearing concepts.

• Machine Learning Models:

  • Evaluated algorithms: Naïve Bayes, Support Vector Machine (SVM), Random Forest, Decision Tree, and K-Nearest Neighbor.
  • Compared model performance on feature-based and concept-level datasets.

---

Methodology

1. Concept Extraction:

   • Created self-made dictionaries to identify sentiment-rich concepts.
   • Mapped these concepts to BanglaSenticNet for sentiment matching.

2. Data Processing:

   • Cleaned data and transformed text into concepts using custom algorithms.
   • Generated two datasets: one for feature-based analysis and another for concept-level analysis.

3. Evaluation Metrics:

   • Used metrics such as accuracy, precision, recall, and F1-score for model performance evaluation.
   • Confusion matrix to analyze predictions.

---

Results

1. Feature-Based Approach:

   • SVM achieved the highest accuracy: 84%.
   • Other models' accuracy: Random Forest (73%), Decision Tree (72%), Naïve Bayes (66%).

2. Concept-Level Approach:

   • SVM achieved the highest accuracy: 96%.
   • Other models' accuracy: Random Forest (75%), Decision Tree (71%), Naïve Bayes (68%), K-Nearest Neighbor (68%).

3. Performance Comparison:

   • The concept-level approach outperformed feature-based models across all metrics, showcasing its efficacy in sentiment analysis for Bengali text.

---

Conclusion

• The proposed concept-level approach improves sentiment detection by incorporating domain-specific knowledge from dictionaries.
• SVM consistently outperformed other models, validating its robustness in both feature-based and concept-level approaches.
• The research highlights the challenges of extracting meaningful concepts from Bengali text and the effectiveness of custom-built dictionaries.

---

Future Directions

1. Advanced Concept Extraction:
   • Implement dependency-based semantic parsing for improved results.
2. Expanding Sentiment Labels:
   • Move beyond positive, negative, and neutral to include emotions like joy, anger, and sadness.
3. Larger Datasets:
   • Work with a more extensive and diverse dataset to generalize findings.
4. Multi-language Support:
   • Apply the methodology to other languages using R or Python.

---

References

Key references include:

• BanglaSenticNet for concept matching.
• SentiWordNet for sentiment classification.
• Various research papers on multilingual sentiment analysis and dependency parsing.

For more details, please refer to the research paper attached in this repository.