Sentiment Analysis on IMDB Movie Reviews

Introduction

I explored the IMDB Dataset of 50K Movie Reviews to apply knowledge of word vectors, contextualized embeddings, and sentence transformers.

Dataset

The dataset has 50,000 balanced data points with two columns: review (text) and sentiment (positive/negative).

Exploratory Data Analysis

• No null values.
• Reviews are variable-length texts.
• Sentiments are either positive or negative.

Text Cleaning

• Removed unwanted characters using regex.
• Tokenized text using NLTK's word_tokenize and sent_tokenize.

Label Encoding

• Converted sentiments to binary labels (1 for positive, 0 for negative).

Top Algorithms for Vectorization and Word Embeddings

CountVectorizer

-CountVectorizer is a straightforward algorithm that converts textual data into numerical vectors by counting the occurrence of each word in the text.

• The MLPClassifier model, when trained on the word vectors generated by the CountVectorizer algorithm, achieved an impressive accuracy score of 90%. In comparison, the Logistic regression model scored 89%, and the XGBClassifier model performed at 85%.

TF-IDF

-This technique is a statistical measure to evaluate the importance of a word relative to a collection of documents(corpus).

• The MLPClassifier model worked best with the vectors produced by the TF-IDF algorithm and got an outstanding accuracy of 90%. The Logistic regression model and the XGBClassifier models also performed well, achieving 89% and 85% accuracy, respectively.

Word2Vec-- Skip-gram Model

-The Skip-gram model is a type of neural network model used for learning word embeddings, which are dense vector representations of words in a continuous vector space.

• The Skip-gram algorithm of the Word2Vec model could not perform well yet achieved good scores. The MLPClassifier accuracy was 79%, the logistic regression model got 73%, and the XGBClassofier could achieve 71%.

Sentence Transformers

all-MiniLM-L6-v2

-This sentence transformer model was trained on 1 billion pairs of input/output. Its embeddings have 384 dimensions. It takes a maximum input of sequence length 256.

bert-base-nli-mean-tokens

-This model generates sentence embeddings by taking the average of token embeddings.

multi-qa-mpnet-base-dot-v1

-It is widely used for semantic search. Its embeddings have 768 dimensions. It takes a maximum input of sequence length 512.

roberta-large-nli-stsb-mean-tokens

-The model is fine-tuned on NLI and STSB. It produces a mean pooling of token embeddings.

Model Comparisons

Model	Accuracy (%)
CountVectorizer - MLPClassifier	90.00
CountVectorizer - Logistic Regression	89.00
CountVectorizer - XGBClassifier	85.00
TF-IDF - MLPClassifier	90.00
TF-IDF - Logistic Regression	89.00
TF-IDF - XGBClassifier	85.00
Word2Vec (Skip-gram) - MLPClassifier	79.00
Word2Vec (Skip-gram) - Logistic Regression	73.00
Word2Vec (Skip-gram) - XGBClassifier	71.00
Sentence Transformers (multi-qa-mpnet-base-dot-v1)	89.52
all-MiniLM-L6-v2	81.73
bert-base-nli-mean-tokens	83.81
roberta-large-nli-stsb-mean-tokens	84.42

Conclusion

After extensive experimentation, the best combinations for sentiment analysis on the IMDB dataset are:

1. TF-IDF with Logistic Regression
2. TF-IDF with MLPClassifier
3. Sentence Transformer (multi-qa-mpnet-base-dot-v1) with RNN

I conducted comprehensive experiments to determine the optimal solution for the IMDB 50K dataset.