Amazon Book Reviews 3-class sentiment classifier

Note

This repository contains all the needed code to train/test/run a simple flask review classification service, but all the trained models and data files are not included. Check the explotary notebook for the details on the data wrangling/analysis and the training of the traditional classification models reported below.

Introduction

This repository contains the code for replicating results described below and can be used to run a simple flask service delivering one of two classification models (Traditional Voting Ensemble or a Deep Learning Bi-GRU model)

Service API Documentation

1. POST/GET /classify

Request

The classify endpoint requires a single argument utterance that contains the text to classify, an example of a valid GET request is as follows : SERVICE_URL/classify?utterance=I loved it

Response

A request to the method classify returns a json object that contains the predicted class as in the example below :

{
    "detected_class": "neutral"
}

The detected_class key can take one of three values : {negative, neutral, positive}

Experiments Results

Below are the results of the top scoring classifiers and their tuned hyper parameters values

# model	# hyper-parameters	accuracy	macro F1	Micro F1
MultiNomial Naive Bayes	alpha: 0.2 - Uniform Prior	85%	0.69	.87
Logistic Regression	L1 Regularization - balancing class weights - C: 1 - one-vs-rest	89%	.73	.90
Linear Kernel Support Vector	L2 Regularization -balancing class weights - C: .2 - one-vs-rest	90%	.73	.90
Soft Voting Ensemble (MNB + LR + Calibrated Linear SV)	same hyper-params for each consituting classifier as mentioned above	90%	.74	.90
Bi-LSTM GRU + Twitter-Glove-200d embeddings	hidden-size of GRU cells : 100 - learning-rate: .001 - Early Stopping	86%	.70	.87

Usage

Requirements

• Python 3.6 with pip
• Installing the dependencies in the requirements.txt file

Training the traditional ML models

• Run the explotary_notebook residing in the explotary_analysis directory

Training the Bi-GRU

• Download a gensim compatible word2vec/glove embeddings
• Build the data by running python build_data_files.py with the appropriate arguments
• Train the model by running python training.py with the appropriate arguments
• Test the model by running python testing.py with the appropriate arguments

Running Flask Classification Service

To run the flask service using a trained model :

• Run python sentiment.py with the appropriate arguments