I got the information a couple of days ago that I got a new grant proposal accepted by the DFG. It’s a Sachbeihilfe, requesting funding for one Ph.D. student and substantial additional support for user studies.
The project’s name is “Interactive Prompt Optimization with the Human in the Loop for Natural Language Understanding Model Development and Intervention” (INPROMPT).
The work will start from the motivation of few-shot or zero-shot settings for the creation of models in algorithmic natural language understanding. A currently modern and popular approach to develop models without too much annotated data is to use pre-trained neural language models and use a prompt to generate a word that describes an instance of text. For example, you can do sentiment polarity classification by entering a text instance such as “The person is very satisfied with the product.” associated with a prompt and check whether the sentence “The product is good” or “The product is bad” results in a higher probability.
Creating such prompts has the advantage that it does not necessarily require technical expertise, but creating good prompts is still not trivial. Existing research has approached the problem from (at least) two perspectives: (1) adapting existing language models using (few) annotated data points and manually generated prompt sets, and (2) using data-driven automatic prompt generation.
We will combine these two directions and start with the typical situation in which a language comprehension task is formulated vaguely, a more precise specification is still missing, and no annotated (but certainly non-annotated) texts are available. Our goal is to develop and analyze systems that automatically guide domain experts without technical training in machine learning to create well-functioning prompts.
To do this, we use optimization methods that change prompts iteratively and estimate their quality with the help of a target function. This estimation is based on automatic predictions on text instances, based on the readability of the prompt, and based on the conclusiveness of an explanation of the decision-making. In our project, the objective function based on these factors is not automatically evaluated, but replaced by a “human in the loop”. However, in order to study the problem of iterative optimization of prompts on a larger scale, we also simulate human decisions using automatic approximations of the human objective function.
We expect that our project will significantly improve the transparency of prompt-based models and contribute to the democratization of the use of machine learning algorithms.
My current plan to is start the project latest in April 2024.