Sentiment Classification with Classical and Neural Models

What I Explored

I explored sentiment classification using the IMDB dataset through multiple modeling approaches.

The experiments included a scikit-learn baseline, PyTorch EmbeddingBag, and a BiLSTM using GloVe word representations.

This allowed me to compare different generations of NLP techniques rather than studying a single model in isolation.

What Confused Me

The central challenge was understanding how models operating on fundamentally different representations could solve the same task.

Traditional models can work surprisingly well with sparse representations, while neural networks learn dense representations that capture different kinds of relationships between words.

What Clicked

I realized that representation is one of the most important decisions in NLP.

The classifier itself is only part of the system. How language is converted into features determines what information the classifier can use.

My Mental Model Now

I think of text classification as:

Raw Text → Tokenization → Representation → Context Aggregation → Classifier → Sentiment

Different NLP approaches primarily change how the middle stages are implemented.

Classical ML relies heavily on engineered statistical representations, while neural approaches learn increasingly rich representations directly from data.

Implementation / Example

I compared sentiment-classification approaches using:

• scikit-learn
• PyTorch EmbeddingBag
• pretrained GloVe embeddings
• bidirectional LSTM networks

Repository: Sentiment Classification with Classical and Neural Models

What I Want to Explore Next

I want to compare these approaches with Transformer encoders, contextual embeddings, attention-based classification, and instruction-tuned language models.