🇺🇸 🇷🇺

IRPNet

IRPNet is a model for classifying textual reviews into positive and negative. Completed as a bachelor's diploma work.

IRPNet is an abbrevation for "Is Review Positive?"-determining Network.

Repository structure

There are two folders in repository. One is for working with data and training the model, the second is a user interface for interacting with the trained model:

• data processing:
  • Parsing.ipynb - parsing KinoPoisk using API to extract labeled reviews;
  • Embedding.ipynb - training a Word2Vec model based on all extracted KinoPoisk reviews;
  • Preprocessing.ipynb - forming of qualitative training dataset for a neural network;
  • Training.ipynb - training a neural network on a GPU using Google Colab;
  • utils.py - some features like timer and progress bar printing;
• interface:
  • main.py - CLI implementation;
  • model.py - classification model implementation;
  • utils.py - some features like colored output in CLI;
  • parameters/ - parameters of neural network (embedding, dictionary of tokens, weights and biases).

Model architecture

Quality metrics

Metrics calculated for a dataset of more than 200K KinoPoisk reviews:

Metric	Value
Accuracy	0.9134621659162305
Precision	0.9914457491022066
Recall	0.909656810116406
F1-score	0.9487919240454142

Technologies

• Kinopoisk Api Unofficial [2.0.1] - parsing reviews;
• Pandas [1.4.2] - working with data;
• Matplotlib [3.4.3] - visualization;
• PyTorch [1.11.0] - neural network;
• Gensim [4.2.0] - Word2Vec model;
• Jupyter Notebook - interactive environment for Python;
• Google Colab - GPU training.

Usage

First of all, you need to install PyTorch:

$ pip install torch

Next, you need to download a module consisting of model.py and a folder with model parameters, and then insert these files into your project.

After that you can use IRPNet:

from model import Model


model = Model()

example_reviews = [
    'Все плохо, не советую',   # Negative 99.392%
    'Все отлично, рекомендую', # Positive 99.996%
    'В целом пойдет'           # Positive 64.878%
]

for review in example_reviews:
    pos, neg = model.process_review(review)
    print(review + ':')
    print(' Positive:', pos)
    print(' Negative:', neg)