This project brings together two trends in NLP. First, with large language models achieving increasingly impressive results on standard benchmarks, the NLP community pays increasing attention to evaluations that go beyond standard benchmark sets. One line of research has returned to challenge sets (e.g. King and Falkedal 1990, Lehmann et al. 1996, Ribeiro et al. 2020). Another point of attention is this year's special theme track of ACL, reality check, raising questions of what happens when models are used in the real world.
Second, as a fast-developing machine learning technique, reinforcement learning (RL) targets sequential decision-making tasks and is capable of adaptively interacting with the users to learn a model with target behavior (Sutton and Barto 2018). Since it is particularly powerful by taking the feedback from users into account and updating the strategy adaptively according to actual needs, in recent years, RL has been widely applied and shown as a well-suited solution in various NLP tasks (Uc-Cetina et al. 2022).
In this project, we aim to start such a research line that will bring these two research directions together and will explore possibilities of using RL to target specific behavior of models. This could be training specifically to avoid a social undesirable bias, but also to exhibit robust behavior on a specific phenomenon (e.g. negation, specific syntactic structures, temporal information interacting with truth values).
Your duties
- Conduct research within the scope of the project culminating in a successful dissertation
- Writing academic articles and presenting your work on conferences
- Taking part in Hybrid Intelligence project activities