Sentiment-Classification

sentiment classification based on multiple deep learning models

This repository provides implementations of various deep learning models for sentiment classification tasks. The models include:

• TextCNN
• LSTM
• BiLSTM
• LSTM with Attention
• BiLSTM with Attention
• BERT
• RoBERTa

Each model is implemented using PyTorch and designed to classify sentiment polarity (positive/negative) from text data.

---

Supported Datasets

The following datasets can be used with the provided code:

1. IMDB Dataset

• Description: The IMDB dataset contains 50,000 movie reviews, labeled as positive or negative.
• Source: The dataset can be loaded directly using torchtext or the datasets library from Hugging Face.
• Where to Access:
  • For torchtext: Automatically downloaded when using the datasets.IMDB.splits method.
  • For datasets: Automatically downloaded using the datasets.load_dataset("imdb") method.

2. Amazon Reviews Dataset

• Description: A large-scale dataset containing customer reviews and ratings from Amazon.
• Source: Available from AWS Open Data Registry.
• How to Download:
  • Use the datasets library: datasets.load_dataset("amazon_polarity").

3. Yelp Reviews Dataset

• Description: Yelp dataset containing customer reviews with sentiment labels (positive/negative).
• Source: Available from the Yelp Dataset Challenge.
• How to Download:
  • Use the datasets library: datasets.load_dataset("yelp_polarity").

4. Twitter Sentiment Analysis Dataset

• Description: Dataset of tweets annotated with sentiment labels.
• Source: Various repositories, such as Kaggle (Twitter US Airline Sentiment).
• How to Download:
  • Manually download from Kaggle or preprocess a Twitter dataset using custom scripts.

---

Model Descriptions and Dataset Usage

TextCNN

• Dataset: IMDB (via torchtext)
• Code Location: textcnn_sentiment.py
• Usage:
  • Automatically loads and preprocesses IMDB data using torchtext.
  • Embeddings are initialized with GloVe vectors.

LSTM and BiLSTM

• Dataset: IMDB (via torchtext)
• Code Location: lstm_sentiment.py, bilstm_sentiment.py
• Usage:
  • Processes data using torchtext and tokenizes using SpaCy.
  • Can be trained on other datasets by replacing the dataset loading part with any binary sentiment dataset compatible with torchtext.

LSTM with Attention and BiLSTM with Attention

• Dataset: IMDB (via torchtext)
• Code Location: lstm_attention_sentiment.py, bilstm_attention_sentiment.py
• Usage:
  • Extends LSTM/BiLSTM models with an attention mechanism.

BERT

• Dataset: IMDB (via datasets)
• Code Location: bert_sentiment.py
• Usage:
  • Loads IMDB dataset using the Hugging Face datasets library.
  • Tokenizes using the BERT tokenizer (BertTokenizer from transformers).
  • Adapts the pre-trained BERT model (BertForSequenceClassification) for sentiment classification.

RoBERTa

• Dataset: IMDB (via datasets)
• Code Location: roberta_sentiment.py
• Usage:
  • Similar to the BERT implementation but uses the RoBERTa tokenizer (RobertaTokenizer) and model (RobertaForSequenceClassification).

---

Prerequisites

Install Required Libraries

• Python 3.8+
• PyTorch 1.10+
• Hugging Face Transformers Library
• torchtext for TextCNN, LSTM, BiLSTM
• datasets library for BERT and RoBERTa
• SpaCy (for tokenization)

Installation Commands

pip install torch torchvision torchtext transformers datasets spacy
python -m spacy download en_core_web_sm

---

How to Run

1. Select a Model:

   • Each model is stored in a separate file (e.g., textcnn_sentiment.py, bert_sentiment.py).
   • Open the file corresponding to the model you want to train/test.

2. Choose a Dataset:

   • By default, all scripts use the IMDB dataset.
   • To use a different dataset, modify the dataset loading and preprocessing sections.

3. Run the Script:

   python <script_name>.py

4. Evaluate the Model:

   • After training, the script will evaluate the model on the test dataset and print the accuracy and loss.

---

Notes on Datasets

• The IMDB dataset is binary-labeled (positive/negative). Ensure other datasets you use are similarly formatted.
• For other datasets, ensure compatibility with the tokenizer and input pipeline of the model.
• Hugging Face datasets library supports many datasets out of the box, and you can replace the dataset name (e.g., "imdb") to experiment with others.