Principal investigator (ÚFAL): 
Project Manager (ÚFAL): 
Grant id: 

Controllable NLG

This project will expand the capabilities and increase the output quality of systems for natural language generation
(NLG) by developing neural NLG models for controlling text attributes (such as style, sentiment etc.). Current
neural NLG models can generate high-quality text, but lack the naturalness and ability to adapt to the user that
would be needed for practical deployment in areas such as persona-based dialogue systems, audience-appropriate
story generation, report generation highlighting key findings, etc. Our goal is to focus on controlling the neural
generation process to adhere to specific attributes. This project addresses this objective in two ways: (1) We will
introduce new methods for automatic attribute transfer using an adversarial generation technique (i.e., using a
target-style discriminator classifier) to make the output match the desired style. (2) We plan to develop a second
model based on an attribute and content conditioned language model using pre-trained BERT or GPT. Bothapproaches will subsequently be combined, either in a unified model, or in an ensemble, to further improve accuracy. A secondary goal of the project is to explore ways of improving automatic evaluation of controllable NLG system outputs in order to accelerate NLG research in this area. Finally, we will evaluate the controllable generator developed in the project in real-world applications, such as controllable dialogue systems, story and
report generation.