A Natural Language Processing and Deep Learning project that performs sentiment classification on 50,000 IMDB movie reviews.

This project compares three different text representation approaches:

The goal is to understand how different text representation techniques affect classification accuracy, model performance, computational cost, and error patterns.

The primary objective is to build and compare multiple sentiment analysis models using the IMDB Movie Reviews dataset.

The project focuses on:

Text preprocessing

Feature extraction

Word embeddings

Transformer-based NLP

Sentiment classification

Model evaluation

Performance comparison

Error analysis

The IMDB Movie Reviews Dataset contains:

Binary sentiment classification

Positive and negative reviews

Label Sentiment 0 Negative 1 Positive

The dataset is loaded using the Hugging Face datasets library.

The reviews are preprocessed before feature extraction.

Convert text to lowercase

Remove HTML tags

Remove unnecessary punctuation

Normalize whitespace

Tokenize text where required

For BERT, minimal preprocessing is used because BERT's tokenizer is designed to process natural language text.

Term Frequency-Inverse Document Frequency represents each review as a numerical vector based on the importance of words within the document collection.

Logistic Regression is used as the classification algorithm.

Word2Vec learns dense vector representations of words based on their contextual relationships.

A Word2Vec model is trained using only the training data to prevent data leakage.

Each review is represented by the average of its word vectors.

The resulting document vectors are used to train a classifier.

BERT (Bidirectional Encoder Representations from Transformers) is a Transformer-based language model that captures contextual relationships between words.

Unlike TF-IDF and basic Word2Vec representations, BERT considers the context in which words appear.

A pre-trained BERT model is used for sentiment classification.

Example:

The model is fine-tuned on the IMDB training data.

BERT uses its own tokenizer:

Minimal preprocessing is used before BERT tokenization to preserve important linguistic information.

Each model is evaluated on the held-out test dataset.

The following metrics are calculated:

Accuracy

Precision

Recall

F1-Score

The performance of the three approaches is summarized below.

Model Accuracy Precision Recall F1-Score TF-IDF + Logistic Regression 0.8954 TBD TBD TBD Word2Vec + Logistic Regression 0.8482 TBD TBD TBD BERT 0.9078 TBD TBD 0.9101

Based on the completed experiments:

TF-IDF Accuracy: 89.54%

Word2Vec Accuracy: 84.82%

BERT Accuracy: 90.78%

BERT F1-Score: 91.01%

BERT achieved the strongest overall performance among the tested approaches.

BERT achieved the highest classification performance in this project.

Its contextual representation allows the model to understand relationships between words more effectively than traditional TF-IDF and averaged Word2Vec representations.

Accuracy: 90.78%

F1-Score: 91.01%

BERT achieved the best overall performance, demonstrating the advantage of contextual Transformer-based representations.

TF-IDF performed strongly with approximately 89.54% accuracy despite using a relatively simple representation and classifier.

Word2Vec achieved approximately 84.82% accuracy, which was lower than both TF-IDF and BERT.

Averaging Word2Vec embeddings can lose important word-order and contextual information.

TF-IDF is computationally simpler than BERT and can provide strong performance for traditional sentiment classification.

BERT requires significantly more computational resources and training time than TF-IDF or Word2Vec.

BERT can better handle contextual meaning, negation, and relationships between words.

TF-IDF may struggle when sentiment depends heavily on word context rather than individual words or n-grams.

Word2Vec provides dense semantic representations but averaging word vectors can remove important sentence-level information.

For applications where computational resources are limited, TF-IDF can provide an excellent accuracy-to-cost trade-off.

Common sentiment classification errors can occur in reviews containing:

Sarcasm

Mixed positive and negative opinions

Complex sentence structures

Negation

Ambiguous language

Long reviews with multiple opinions

For example, a review may contain many positive words while the overall sentiment is negative because of sarcasm or context.

BERT is generally better suited to these cases because it can model contextual relationships between words.

Approach Representation Computational Cost Context Understanding TF-IDF Sparse vectors Low Limited Word2Vec Dense embeddings Medium Moderate BERT Contextual embeddings High Strong

TF-IDF: Best when speed, simplicity, and low computational cost are important.

Word2Vec: Useful for learning semantic word representations and traditional NLP applications.

BERT: Best when maximum classification performance and contextual understanding are priorities.

Fixed random seeds are used where applicable to improve reproducibility.

For machine learning models:

Python

Pandas

NumPy

Scikit-learn

Gensim

TensorFlow / PyTorch

Hugging Face Transformers

Hugging Face Datasets

Matplotlib

Seaborn

Google Colab

The complete implementation, preprocessing steps, model training, evaluation metrics, comparison table, and final conclusion are available in the Google Colab notebook:

Make sure the sharing permission is:

This project demonstrates practical knowledge of:

Natural Language Processing

Text preprocessing

TF-IDF

Word2Vec

Word embeddings

BERT

Transformer models

Sentiment classification

Logistic Regression

Model evaluation

Accuracy

Precision

Recall

F1-score

Error analysis

Model comparison

Reproducible Machine Learning

This project compared three different approaches for sentiment analysis on the IMDB movie review dataset: TF-IDF, Word2Vec, and BERT. TF-IDF with Logistic Regression achieved strong performance while requiring relatively low computational resources. Word2Vec provided dense semantic representations but performed less effectively when document representations were created by simply averaging word vectors. BERT achieved the highest performance, demonstrating the advantage of contextual Transformer-based representations for understanding complex natural language. The experiments also highlighted an important trade-off between model accuracy and computational cost. While BERT provides stronger contextual understanding and better overall classification performance, TF-IDF remains an attractive solution for applications where simplicity, speed, and resource efficiency are important. Overall, the project demonstrates that the choice of text representation has a significant impact on sentiment classification performance.

MSc in CSE — Data Science Machine Learning | Deep Learning | NLP | Artificial Intelligence | Computer Vision

⭐ If you find this project useful, consider giving the repository a star.