/pytorch-nlugen

Generative models for natural language understanding

Primary LanguagePythonMIT LicenseMIT

PyTorch Implementation of NLU Generative Models

Features

  • train: train a generative model using a labeled NLU dataset.
  • predict: predict intents and slot labels for an unlabeled NLU dataset.
  • generate: generate utterances and their respective intents and slot labels with the given latent variable sampling distribution (uniform, gaussian, gaussian mixture, etc.)

Run python -m (train|predict|generate) --help to checkout available options

Required Packages

Python >3.6 is required at the least.

  • Install packages listed in requirements.txt.
  • To validate generated sentences using the Universal Sentence Encoder (arxiv:1803:11175), install tensorflow and tensorflow-hub packages.
  • To use tensorboard as the visualization tool, install tensorboardX==1.7