Uncertainty-Aware Neural Language Models

In this project, we propose to develop novel models that incorporate the interpretation uncertainty of natural languages into modern neural language models. We expect that these uncertainty-aware language models will provide a much better fit for human languages and communication in general, which will become visible in practical tasks such as question-answering and machine translation.

  • 2022-2024, funded by the Academy of Finland

Research Team

External collaborators:

Project Setup

The project consists of 4 work packages and their connections are illustrated in the figure below:

WP1 focuses on the augmentation of existing state-of-the-art neural language models with uncertainty information. In WP2, we aim at the development of natively uncertainty-aware models based on Bayesian extensions of deep neural network architectures. WP3 and WP4 are concerned with evaluation in order to assess the quality and internal workings of the novel uncertainty-aware models we create throughout the project. In WP3, we will use established NLP benchmarks emphasizing tasks that require advance natural language understanding with strong ambiguities and fuzzy decision boundaries. WP4 tackles the explainability of neural language models and we aim to study the parameters and dynamics of uncertainty-aware language models.

News and activities

  • SHROOM at SemEval 2024 – a Shared task on on Hallucinations and Related Observable Overgeneration Mistakes