A winning (1st place) solution for "AI 4 Humanities: ruGPT-3" track of the AI Journey 2020 competition.
In the work, I use pre-trained ruGPT-3 model for the task of semantically controlled response generation in context of goal-oriented dialog systems. The work is highly influenced by Peng et. al 2020 and details may be found in the presentation (ru).
A model trained on RuCoS dataset:
- Text generation (colab)
- Training of the model (colab)
A model trained on SentiRuEval_2016 dataset:
- Text generation (colab)
- Training of the model (colab)
This repository includes third-party resources:
shell/pretrain_transformers.py
A modified version of the scipt distributed with the ruGPT-3 model and mantained by the SberDevices team.data/datasets/SentiRuEval_2016
The dataset consists of user feedback regarding bank and telecom companies collected on Twitter. It was presented on the International Conference "Dialogue 2016" by Lukashevich and Rubtsova.data/datasets/RuCoS
The dataset automatically generated from CNN/Daily Mail news articles and distibuted within Russian SuperGLUE benchmark.