Aspect-Based Sentiment Analysis Using Transformers

This repository contains an implementation of an Aspect-Based Sentiment Analysis (ABSA) model using Transformer architectures. The model aims to determine the sentiment polarity (positive, negative, neutral) of specific aspects within a given text.

Table of Contents

• Introduction
• Model Architecture
• Dataset Preparation
• Installation
• Usage
• Training
• Evaluation
• Results
• Contributing
• License

Introduction

Aspect-Based Sentiment Analysis is a fine-grained sentiment analysis technique that not only determines the overall sentiment of a text but also identifies the sentiment toward specific aspects or features mentioned in the text.

For example, in the sentence:

"The battery life of this camera is amazing, but the lens quality is disappointing."

The model should identify:

• Aspect: "battery life" - Sentiment: Positive
• Aspect: "lens quality" - Sentiment: Negative

Model Architecture

Our implementation leverages pre-trained Transformer models (like BERT) to encode textual data and applies a classification layer to predict the sentiment polarity of each aspect.

High-Level Architecture

graph TD
    A[Input Sentence] --> B[Tokenization]
    B --> C[Transformer Encoder]
    C --> D[Aspect Extraction]
    D --> E[Aspect-Specific Representation]
    E --> F[Classification Layer]
    F --> G[Sentiment Prediction]

Loading

Detailed Workflow

1. Tokenization: The input sentence is tokenized using the tokenizer associated with the pre-trained Transformer model.

2. Transformer Encoding: The tokenized input is passed through the Transformer encoder to obtain contextualized word embeddings.

3. Aspect Extraction: Aspects are identified using a sequence labeling approach or provided externally.

4. Aspect-Specific Representation: For each aspect, a representation is generated by pooling the embeddings of tokens corresponding to the aspect.

5. Classification: The aspect-specific representation is fed into a classification layer to predict the sentiment polarity.

Dataset Preparation

The model expects the dataset in the following format:

• Sentence: The input text containing one or more aspects.
• Aspects: A list of aspects present in the sentence.
• Sentiments: Corresponding sentiment labels for each aspect.

Example

Sentence	Aspects	Sentiments
"Great food but the service was dreadful and slow."	["food", "service"]	["Positive", "Negative"]
"The screen quality is superb, but the battery drains fast."	["screen quality", "battery"]	["Positive", "Negative"]

Installation

Clone the repository and install the required packages:

git clone https://github.com/SurajDonthi/AspectBasedSentimentAnalysis.git
cd AspectBasedSentimentAnalysis
pip install -r requirements.txt

Usage

Inference

You can use the pre-trained model for inference on new sentences:

from model import AspectSentimentAnalyzer

model = AspectSentimentAnalyzer.load_pretrained('path_to_model')
sentence = "The camera resolution is fantastic, but it's quite bulky."
aspects = ["camera resolution", "size"]
predictions = model.predict(sentence, aspects)
print(predictions)

Training

To train the model on your dataset:

python train.py --data_path data/dataset.csv --epochs 10 --batch_size 32

Training Pipeline

graph LR
    A[Input Data] --> B[Data Preprocessing]
    B --> C[Model Initialization]
    C --> D[Training Loop]
    D --> E[Validation]
    E --> D
    D --> F[Trained Model]

Loading

Evaluation

Evaluate the model's performance on a test set:

python evaluate.py --model_path saved_model --test_data data/test_dataset.csv

Metrics reported include Accuracy, Precision, Recall, and F1-Score for each sentiment class.

Results

The model achieves the following performance on the benchmark dataset:

Metric	Positive	Negative	Neutral
Precision	85%	83%	78%
Recall	82%	80%	75%
F1-Score	83.5%	81.5%	76.5%

Contributing

Contributions are welcome! Please open an issue or submit a pull request for any improvements.

License

This project is licensed under the MIT License.