Aspect-Based Sentiment Analysis for Mobile Phone Reviews Using AI Techniques

(7PAM2002: Data Science Master’s Project)

This project aims to perform aspect-based sentiment analysis on customer reviews of mobile phones from Flipkart. The primary objective is to classify sentiments and extract insights using advanced AI techniques.

---

Owned by:

Name: Vullinkala Ratan Srivasthava Reddy

Student ID: 22031426

UNIVERSITY OF HERTFORDSHIRE

Dataset

• Source: Kaggle Flipkart Cell Phone Reviews
• Content: Reviews and product details of various mobile phones from Flipkart.

---

Input and Output

Input:

• Customer reviews of mobile phones from Flipkart in textual format.
• Corresponding product metadata such as brand, model, and ratings.
• Pre-processed text after cleaning, tokenization, and normalization.

Output:

• Predicted sentiment labels (Positive, Negative) for each aspect of the review.
• Visualization of sentiment trends using bar plots and word clouds.
• Performance metrics (Loss, Accuracy, F1 Score) for different models.
• Comparison of model effectiveness in sentiment classification.

---

Methodology

Data Processing

1. Cleaning:

   • Removed HTML tags, decoded emoticons, and handled special characters to ensure clean and standardized text data.

2. Tokenization:

   • Split reviews into sentences and words using word tokenization, preparing the data for model training.

3. Text Normalization:

   • Removed stop words and performed stemming and lemmatization to reduce words to their root form for better model learning.

4. Balancing:

   • Applied antonym augmentation to balance the dataset, specifically targeting negative reviews to ensure a more representative training set.

Exploratory Data Analysis (EDA)

• Analyzed review lengths, active reviewers, and common terms using bar plots and word clouds to understand the dataset better.
• Conducted target distribution and n-grams analysis for deeper insights into the nature of reviews and sentiments.

Modeling

Three models were trained and evaluated for aspect-based sentiment analysis:

1. LSTM (Long Short-Term Memory)
2. BiLSTM (Bidirectional LSTM)
3. Enhanced BiLSTM (Proposed)

These models were trained using different hyperparameter configurations, as detailed below.

---

Hyperparameter Details

Hyperparameters for LSTM, BiLSTM, and Enhanced BiLSTM:

1. LSTM_hparam1 (Default Hyperparameters)

   • Embedding Dimension: 50
   • LSTM Units: 128
   • Dropout Rate: 0.3
   • Learning Rate: 0.001
   • Epochs: 30
   • Batch Size: 64

2. LSTM_hparam2

   • Embedding Dimension: 64
   • LSTM Units: 128
   • Dropout Rate: 0.4
   • Learning Rate: 0.0005
   • Epochs: 30
   • Batch Size: 64

3. LSTM_hparam3

   • Embedding Dimension: 128
   • LSTM Units: 128
   • Dropout Rate: 0.5
   • Learning Rate: 0.0001
   • Epochs: 30
   • Batch Size: 32

4. BiLSTM_hparam1, BiLSTM_hparam2, BiLSTM_hparam3
   These hyperparameters mirror the LSTM model configurations but with bidirectional LSTM layers for enhanced learning of sequence patterns.

5. Enhanced BiLSTM_hparam1, Enhanced BiLSTM_hparam2, Enhanced BiLSTM_hparam3
   Enhanced models include dropout and L2 regularization techniques to prevent overfitting and improve generalization.

---

Results

Model	Accuracy (%)	Precision	Recall	F1 Score	ROC-AUC
LSTM (H_param1)	91.0	0.91	0.94	0.94	0.95
BiLSTM (H_param1)	89.0	0.89	0.93	0.92	0.94
Enhanced BiLSTM (H_param1)	89.0	0.89	0.94	0.92	0.94
LSTM (H_param2)	91.0	0.91	0.94	0.94	0.95
BiLSTM (H_param2)	89.0	0.89	0.93	0.92	0.94
Enhanced BiLSTM (H_param2)	90.0	0.90	0.93	0.93	0.94
LSTM (H_param3)	91.0	0.91	0.94	0.94	0.95
BiLSTM (H_param3)	89.0	0.89	0.94	0.93	0.94
Enhanced BiLSTM (H_param3)	89.0	0.89	0.95	0.93	0.94

Table 4.1: Model Performance Comparison
This table summarizes the performance of LSTM, BiLSTM, and Enhanced BiLSTM models across different hyperparameter configurations. LSTM consistently achieved the highest accuracy, precision, recall, and F1 scores, with an overall strong ROC-AUC of 0.95.

---

Key Features

• Antonym Augmentation: Improved dataset balance by increasing negative reviews, contributing to better model robustness.
• Dropout and Learning Rate Scheduler: These techniques helped prevent overfitting and ensured stable model convergence during training.
• Comparative Model Analysis: The results from different models with varying hyperparameters show trade-offs between performance and computational efficiency.

---

Conclusion

The project demonstrated the effective use of advanced AI techniques for aspect-based sentiment analysis on mobile phone reviews. LSTM outperformed other models in terms of overall performance across various metrics. However, Enhanced BiLSTM also showed promising results in terms of recall for the positive class. The findings suggest that LSTM is the best model for sentiment analysis in this case, and it can be further fine-tuned for even better performance. This approach not only aids in understanding customer sentiments but can also be leveraged for making data-driven decisions in the mobile phone e-commerce sector.