IMDB Sentiment Analysis

A repository to explore the Doc2Vec algorithm on the IMDB dataset.

Data Source

The dataset can be downloaded from Kaggle here. It comes already split into Train, Validation and Test sets.

Text Visualization Dashboard

A dashboard to explore the dataset can be found here.

Analysis Pipeline

1 Clean Raw Text

The script reads in the raw Train, Validation and Test sets and writes out cleaned datasets. It uses BeautifulSoup to remove html tags and spaCy library to perform the following tasks:

1. remove stopwords
2. remove punctuation
3. retains alpha characters only

python src/pre-processing.py

2 Build Doc2Vec Model

The script uses gensim library to build a model that represents the entire movie review document as a fixed length vector. Initially, the model is parameterized to represent each review in 100 dimensions.

The model is built using the Train set only. After the Doc2Vec model is trained, it will be used to infer the document vectors for the Validation and Test sets.

python src/doc2vec.py

3 Extract Doc2Vec Features

This script uses the Doc2Vec model produced in step 2 to extract individual document vectors for the Train set and infer the individual document vectors for the Validation and Test portions of the data. The extracted vectors are then saved as a new dataset where each row is a movie review with 100 features learned from the model.

python src/feature_extraction.py

Doc2Vec Model Experimentation

The sentiment analysis task was carried using multiple doc2vec models. The parameters of the different models is documented below.

Version 1: (PV-DM)

This model averages together the paragraph vector along with the learned context window word vectors to perform the prediction task of predicting the next word in the sequence.

• max_epochs = 100
• vec_size = 100
• min_count=2
• alpha = 0.025
• dm=1
• window=10

Version 2: (PV-DM)

The only difference between this model and the one above is that instead of averaging the paragraph vector and context word vectors together before the prediction task, it concatenates the paragraph vector and the window context word vectors together. This model is supposed to improve performance due to the concatentation. However, this model produces a very large model object (344MB + 50MB)

• max_epochs = 100
• vec_size = 100
• min_count=2
• alpha = 0.025
• dm=1
• window=10
• dm_concat = 1

Version 3: (PV-DBOW)

For this model, only the paragraph vectors are trained without the word vectors. The network is trained to predict the words occurring in the context window, given the paragraph vector.

• max_epochs = 100
• vec_size = 100
• min_count=2
• alpha = 0.025
• dm=0
• window=10