IMDb Sentiment Analysis — Contextual Word Classification - Shanghai Jiao Tong University/Mines Paris

Authors: Wael Ben Slima & Marko Babic
Notebook: Wael_Marko_IMDB_Sentiment_Analysis.ipynb
Dataset: IMDb Dataset of 50K Movie Reviews

---

Project Overview

This notebook implements a contextual word sentiment classification model using the IMDb movie review dataset.
The primary goal is to classify individual words as positive, negative, or neutral by leveraging sentence-level sentiment labels and the context of surrounding words.

For example:

• “beautiful” → Positive
• “defeat” → Negative

---

Dataset Description

The IMDb dataset contains 50,000 movie reviews, split into:

• 25,000 for training
• 25,000 for testing

Each review is labeled as either positive or negative.

---

Workflow

1. Data Loading & Preprocessing

• Load data using Pandas.
• Clean the text: remove HTML tags & punctuation, lowercase, strip numbers, remove stopwords.
• Tokenize and pad sequences for model input.

2. Model Building

• Utilize TensorFlow / Keras.
• Architecture includes:
  • Embedding layer
  • (Bi)LSTM layer to capture contextual dependencies
  • Dense output layers for classification

3. Training

• Train on sentence-level labels.
• Use callbacks (e.g. ModelCheckpoint) to save the best model.

4. Evaluation

• Plot accuracy and loss curves.
• Compute confusion matrix and classification metrics.

---