This project aims to perform sentiment analysis on Amazon reviews ranging from May 1996 to October 2018. By analyzing customer reviews in the Electronics and Home & Kitchen categories, the project predicts whether a review is positive or negative.

Input : “Do not buy this.”

• Outputs : "92.21% probability of being a negative review.”

Input : “I love this!”

• Output : “95.25% probability of being a positive review.”

• Running system example:
  • WSL2 / Ubuntu 22.04
  • Python 3
  • TensorFlow 2.14.0
  • CUDA : 11.8
  • cuDNN : 8.7
  • Hardware:
    • GPU : RTX 4060TI 6GB
    • Memory : 32 GB

First, update your local repository:

{

"$ git pull",

}

Next, download the dataset with DVC:

{

$ dvc pull

}

Run the setup script to prepare your environment automatically:

{

$ python3 setting_up.py

}

To predict whether a given input text is negative or positive, run the prediction script:

{

$ python3 prediction.py

}

Input any text when prompted to receive a sentiment prediction.

Source: Amazon Reviews (May 1996 - Oct 2018)

Format: JSON

Datasets for the prediction model : Electronics_5.json, Home_and_Kitchen_5.json

Example data:

{ "sort_timestamp": 1634275259292,

"rating": 3.0,

"helpful_votes": 0,

"title": "Meh",

"text": "These were lightweight and soft but much too small for my liking. I would have preferred two of these together to make one loc. For that reason I will not be repurchasing.",

"images": [{ "small_image_url": "https://m.media-amazon.com/images/I/81FN4c0VHzL.SL256.jpg", "medium_image_url": "https://m.media-amazon.com/images/I/81FN4c0VHzL.SL800.jpg", "large_image_url": "https://m.media-amazon.com/images/I/81FN4c0VHzL.SL1600.jpg", "attachment_type": "IMAGE" }],

"asin": "B088SZDGXG",

"verified_purchase": true,

"parent_asin": "B08BBQ29N5",

"user_id": "AEYORY2AVPMCPDV57CE337YU5LXA" }

Total number of reviews : 13,638,545

Preprocessing steps include converting JSON files to dataframes, sampling, and transforming into TFRecord format for efficient training.

• Convert JSON files to data frames with columns: reviewText, overall, and vote.
• Apply custom preprocessing functions.
• Sampling the data
  • Sample data due to physical memory limitations (sample size = 100,000 for each dataframe)
  • Assign weight values based on vote counts for reliability
  • Drop the vote column.

• Create a positivity column where ratings over 3 are labeled as 1 (positive), otherwise 0 (negative)
• Drop the overall column.

• The data is then transformed into TFRecord format, requiring serialization methods.

The preprocessed data is loaded, processed for tokenization and padding, and prepared for model training.

• Load and convert data into a dataframe.
• Set X_df as reviewText and y_df as positivity, then split.
• Process Tokenization and Padding for the reviewText data
  • Use tensorflow.keras
• Save the tokenizer in JSON format for the prediction model.
• Returning X_train, X_test, y_train, y_test, and the tokenizer.

A TensorFlow model is constructed to classify texts based on sentiment.

• Use the TensorBoard callback to monitor the training process.
• Adjust for WSL2 GPU memory usage.
• Obtain X_train, X_test, y_train, y_test, and the tokenizer via the process_data function from training.py.
• Adjust Hyperparameters:
  • Tune the model with specific hyperparameters including Embedding Dimension, Neuron Units, Vocabulary Size, Batch Size, and Number of Epochs.
• Construct the Model:
  • Add a Dense layer with a sigmoid activation function suitable for binary classification.
• Implement EarlyStopping and ModelCheckpoint callbacks for improved training and memory efficiency.
• Load and save the model data in HDF5 format for efficient storage and access.
• Prediction:
  • Assess test accuracy to validate the model’s predictive capability.

• Apply preprocessing to the user-provided string using the tokenizer from training.py.
• Utilize the trained model to predict whether the input string conveys a negative or positive sentiment.