Sentiment Vectors

A Python reproduction of Learning Word Vectors for Sentiment Analysis by Maas et al. (2011).

This project learns word representations from IMDb movie reviews. The objective is to build word vectors that capture both:

semantic similarity: words used in similar contexts should be close; sentiment orientation: words carrying similar polarity should also be close.

The implementation follows the main ideas of the paper and evaluates the learned representations using a linear SVM classifier.

📁 Project Structure

SentimentAnalysis/
├── aclImdb/
│   ├── train/
│   │   ├── pos/
│   │   ├── neg/
│   │   └── unsup/
│   └── test/
│       ├── pos/
│       └── neg/
├── params/
├── src/
│   └── sentiment_vectors/
│       ├── __init__.py
│       ├── data.py
│       ├── vocabulary.py
│       ├── semantic.py
│       ├── sentiment.py
│       ├── evaluation.py
│       └── utils.py
├── Scripts/
│   ├── step1_load_data.py
│   ├── step2_build_vocab.py
│   ├── step3_train_semantic.py
│   ├── step4_train_full.py
│   └── step5_evaluate.py
├── pyproject.toml
└── README.md

Dataset

The project uses the IMDb dataset introduced by Maas et al. (2011).

aclImdb/
├── train/
│   ├── pos/    positive labeled reviews
│   ├── neg/    negative labeled reviews
│   └── unsup/  unlabeled reviews
└── test/
    ├── pos/    positive test reviews
    └── neg/    negative test reviews

The training data contains: 25,000 labeled reviews 50,000 unlabeled reviews 25,000 test reviews

Example results

Nearest neighbors obtained from the learned word vectors:

Query word	Semantic + Sentiment	Semantic Only
wonderful	perfect, fantastic, deserved, incredible, awesome	perfect, amazing, excellent, superb, fabulous
terrible	awful, horrible, plain, corny, crap	awful, horrible, atrocious, 1/10, badly
boring	terrible, predictable, awful, annoying, honestly	amateur, pointless, mess, 1/10, terrible
romantic	romance, appealing, charm, charming, performer	romance, charming, appealing, roles, charm

Classification results:

Method	Our Result	Paper
Bag of Words (bnc)	87.94%	87.80%
Our Semantic Only	82.84%	87.30%
Our Full	87.52%	87.44%
Our Full + BoW	88.06%	88.89%

Reference

Maas, Andrew L., Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. Learning Word Vectors for Sentiment Analysis. ACL 2011. Dataset: IMDb Large Movie Review Dataset.# Sentiment-Analysis