GPT-2 Review Generation with N-Shot Prompting

Project Overview

This project demonstrates how to generate customer product reviews using GPT-2 with N-shot prompting to guide the model towards more organic and contextually appropriate reviews. The script leverages the Hugging Face transformers library to fine-tune the model's output based on user input.

The model is fed a series of N-shot examples of customer reviews with varying sentiments (positive, negative, and neutral) before generating its own review based on the specified product, category, and sentiment.

Features

• N-shot Prompting: Provides the model with pre-defined examples to influence its output, resulting in more coherent and contextually relevant reviews.
• Sentiment Control: The generated review can be tailored to a specific sentiment — positive, negative, or neutral.
• Custom Product & Category: Users can specify the product and category for which they want a review.
• Randomized Output: Each run produces a unique review, even with the same input, thanks to the GPT-2 model’s generative capabilities.

Requirements

To run this project, you need the following libraries installed:

pip install transformers

The model is based on the Hugging Face transformers library and utilizes GPT-2 for text generation.

How It Works

1. The model takes a series of N-shot examples (i.e., examples of product reviews for different sentiments) as input.
2. The user provides the product name, category, and desired sentiment (positive, negative, or neutral).
3. The script generates a natural-sounding review based on the user’s input while leveraging the given examples to guide the model’s output.
4. The final output excludes the N-shot examples from the generated review, ensuring only the newly generated text is displayed.

Code Structure

• Review Generation Pipeline: The model pipeline is built using GPT-2 (gpt2 model) from Hugging Face. It generates the review based on the constructed prompt and N-shot examples.
• N-shot Prompting: Before generating the review, several examples (positive, negative, and neutral reviews) are passed to the model as a reference.
• Sentiment Selection: The user selects a desired sentiment, and the model tailors its review output accordingly.

Example Input and Output

1. Input:

   • Product: Samsung Phone
   • Category: Smartphone
   • Sentiment: Negative

2. Generated Review:

   I bought the Samsung Phone in the Smartphone category and it was quite disappointing. I was not happy because the phone’s battery drained quickly, and it lagged while using basic apps. The overall performance was below expectations.
   

How to Use

1. Clone the repository or download the script.
2. Install the required dependencies using the command: pip install transformers.
3. Run the script. It will prompt you to provide:
   • Product Name
   • Category
   • Sentiment (Positive, Negative, Neutral)
4. The model will generate a review based on the provided inputs.

Running the Script

python review_generation.py

After running the script, it will ask for the product details and the desired sentiment. Based on the inputs, it will generate a review influenced by the N-shot examples and display the result.

Customization

You can customize the N-shot examples by modifying the n_shot_examples list in the script. For instance, you can add or remove examples to better suit your domain or use case.

Example for adding a new positive review to the list:

n_shot_examples = [
    "I bought the headphones and they were absolutely amazing! The sound quality was crystal clear and the comfort was top-notch. I would highly recommend them to anyone. (Positive)",
    "I purchased the phone charger and it was disappointing. It broke within a week and didn’t work well at all. I wouldn’t recommend this to others. (Negative)",
    "I got the blender, and it was okay. It worked as expected, but there was nothing special about it. It does the job, but I don’t feel strongly either way. (Neutral)",
    "I bought the new tablet and it exceeded my expectations! The screen resolution is stunning and the performance is flawless. (Positive)"  # New example added
]

License

This project is open-source and available under the MIT License.