Controllable and faithful generation of language in time critical low-resource scenarios
Mayank Jobanputra (Ph.D. Student)
Today's neural generation systems produce fluent text, but are prone to generate factual errors. This is particularly problematic for safety-critical settings such as generation systems that assist human experts in making quick decisions for autonomous systems. The thesis project will focus on generating relevant, faithful and reliable messages based on structured input data. Additionally, multimodal challenges such as generating references or messages that aggregate across several entites and make reference to information shown in the visual modality will be addressed. To support this research, we will create the necessary datasets and develop an evaluation metric to assess the reliability of the generated messages. The ultimate goal is to validate the effectiveness of our end-to-end system through a user study focused on autonomous drone flight scenarios.
|Primary Host:||Vera Demberg (Saarland University)|
|Exchange Host:||Raquel Fernández (University of Amsterdam)|
|PhD Duration:||01 September 2023 - 31 December 2026|
|Exchange Duration:||- Ongoing - Ongoing|