Sentiment Classification

Overview : This study model is constructed as feed forward neural network using Tensorflow/Keras framework on Twitter dataset. Model composed of 3 dense layers with relu activation function. As final layer a dense layer with softmax activation function. Performance of the model can be seen in evaluation section. Prediction file, model ,dictionaries ,model weights were dockerized. Docker file can be pulled from docker hub.

Data

Twitter dataset on kaggle : https://www.kaggle.com/datasets/jp797498e/twitter-entity-sentiment-analysis

File Description :

• LE.pkl : pickle file for Label encoder
• tv_layer.pkl : pickle file for Tensorflow text vectorization
• model.py : nueral network model
• prediction_fast.py : prediction file for deployment purpose
• sentiment_analysis.py : training file
• sentiment_model.h5 : model weights
• requirements.txt : dependencies

Docker

Docker pull command:

docker pull maliphy/sentiment_class

Docker run for prediction:

docker run -t -i maliphy/sentiment_class:v1

Evaluation

Twitter sentiment analysis

                   precision recall  f1-score   support

  Irrelevant       0.97      0.98      0.97       172
    Negative       0.97      0.99      0.98       266
     Neutral       0.99      0.96      0.98       285
    Positive       0.98      0.97      0.97       277

    accuracy                           0.97      1000
   macro avg       0.97      0.98      0.97      1000
weighted avg       0.98      0.97      0.97      1000