Affective Processing emerging from Large Language Models?
At the 11th Internatinal Conference on Affective Computing and Intelligent Interaction (ACII) at the MIT Media Lab, the Social AI Group paper “Fine-grained Affective Processing Cpapbilities Emerging from Large Language Models” has received th Honorable Mention Award. The paper was a joint collaboration between Joost Broekens, Suzan Verberne and Aske Plaat from LIACS at Leiden University, Kim Baraka from the Social AI Group at VU and Bernhard Hilpert (Leiden + VU) as well as Patrick Gebhard from the Affective Computing Group at the German Research Center for Artificial Intelligence (DFKI).
In 6 controlled experiments, they tested ChatGPT's capability to process and transform affective stimuli. Specifically, Affect Recognition, Affect Representation and Appraisal capabilities were examined: The team extracted VAD-affect values from situations and words, map those two stimulus sets together based on a numerical and latent affective representation and even implemented a functioning version of the OCC model of appraisal! And even more interestingly the LLM was capable of creating whole new sets of valid stimuli based on merely a description of affective states as stimuli.
Key takeaways are:
- Powerful sequence predictors trained on massive amounts of data show emergent capabilities in the affective domain
- Language data seems to serve as a powerful medium that carries semantic context for affect processing
- LLMs seem to have a latent representation of affective meaning that gets better with more context
That is useful for research and practice:
- It enables systems to symbolically ground information and may therefore help to make classical symbolic models work
- The paper demonstrates that LLMs can be utilized for stimulus validation and creation.
- LLMs could be used as an API/interpreter for human agent interaction by giving agents a structured knowledge base
Read the paper here: https://arxiv.org/pdf/2309.01664.pdf
A special thanks also goes to the Hybrid Intelligence project for supporting and funding this research!