Sentiment Analysis using LSTM model

An LSTM model for sentiment analysis and API provided for those who want to use LSTM sentiment analysis for their applications.

---

Table of Contents

1. Overview
2. Model Training
3. File Descriptions
4. Sentiment Analysis API

Overview:

This is an LSTM model that allows users to do fundamental sentiment analysis.

• Provides the users the ability to prepare data for the ML model
• It provides users the ability to create complex mini-batches of large data.
• Allows users to train model, plot graphs and saves checkpoint for early stopping.

Model Training:

The LSTM model is trained on the dataset provided by sentiment140. The dataset was divided as follows:

• Training set : 954177 tweets (~60%)
• Validation set : 317964 tweets (~20%)
• Testing set : 318072 tweets (~20%)

The data provided where tweets that are considered to have positive sentiment or a negative sentiment. The model was trained on 300 epochs with batches of 128 and a learning rate of 2e-5. The training is shown in the graph below:

As we can observe, there is a slight overfit after epoch ~50. The maximum training accuracy the model reached is ~80%. Even with 80% accuracy, the model is showing promising results with sentiment analysis.

File Descriptions:

The whole project is divided into 5 different files for ease of readability and modifications.

• model.py: The main file that contains the LSTM model.
• prepare.py: This files reads the data from the file specified, divides the files into datasets and prepare them to be used by the model.
• batching.py: This files creates batches of the dataset during the training process.
• train.py: This files contains functions that trains the model, finds it's accuracy and save checkpoints for early stopping purposes.
• loader.py: This file is used to load a checkpoint, and find sentiment on a piece of text.

Sentiment Analysis API:

You can access the REST API created using the LSTM model in this repo. The location for REST API server is:

www.sentiment-analysis-api.site

The REST API sends 2 responses to a request made:

• 0 : If the sentiment of a sentence is negative
• 1 : If sentiment of a sentence is positive

You can access the API as follows:

Single Sentence:

If you want to parse a single sentence, to determine it's sentiment, we can send a basic JSON object using curl as follows:

>> curl -d '{"data":"this is awesome"}' -H "Content-Type: application/json" -X POST https://www.sentiment-analysis-api.site/api/one
1

>> curl -d '{"data":"I need a hug"}' -H "Content-Type: application/json" -X POST https://www.sentiment-analysis-api.site/api/one
0

>> curl -d '{"data":"i wish i was happy"}' -H "Content-Type: application/json" -X POST https://www.sentiment-analysis-api.site/api/one
0

Bulk Sentences

If you want to parse bulk of sentences in order to determine it's sentiment, you can send a basic JSON object using curl as follows:

>> curl -d '{"data":["This is awesome", "I need a hug", "i wish i was happy","My name is Beant"]}' -H "Content-Type: application/json" -X POST https://www.sentiment-analysis-api.site/api/bulk
{
  "I need a hug": 0, 
  "My name is Beant": 1, 
  "This is awesome": 1, 
  "avg": 0.5, 
  "i wish i was happy": 0
}

where avg is average score of the sentences.