Intent Classification for virtual assistants

This is a project made for the Natural Language processing and Machine Learning seminar.

This project contains 2 pytorch models:

• model - A simple model using pre-trained paraphrase-albert-small-v2 from HuggingFace without any fine-tuning
• model2 - Model based on paraphrase-albert-small-v2 fine tuned with the a part of the Out-of-Scope Intent Classification Dataset

Additionally there's also a simple interactive mode to try out the model.

Requirements

• python >=3.10
• pytorch
• pandas
• scikit-learn
• transformers
• numpy

Run interactive mode

• Download repository

  git clone https://github.com/vabenil/albert_intent_analysis

• Move to the root directory of the project

  cd albert_intent_analysis

• Run

  python -m intent_analysis.interactive

Installation

Requires pip installed.

• Download repository

  git clone https://github.com/vabenil/albert_intent_analysis

• Move to the root directory of the project

  cd albert_intent_analysis

• Install package

  pip install .

Usage

Example usage of model

from intent_analysis.model import Net, run_model
model = Net()

# token representing sentence as (1, 768) vector
Y = model(["Hello!"])
labels = run_model(model, ["Hello!"])

Example usage of model2

from intent_analysis.model2 import Net, run_model

model = Net.pretrained()
# one-hot-encoded vector representing label
Y = model(["Hello!"])
# List of labels. In this case ["GREETING"]
labels = run_model(model, ["Hello!"])

Dataset

Training data is slightly modified version of a part of the Out-of-Scope Intent Classification Dataset

• train data (found in datasets/train_data.json)
• test data (found in datasets/test_data.json)