(2/n)
What is generalizable classification here? We think there are three key elements
1. New data domains - from short informal text to long passages.
2. New moral and value dimensions.
3. New frameworks - e.g. moral foundations, Schwartz human values, and many more!
(3/n)
Our new methodology insight: "all@once LLM prompting strategy" outperforms fine-tuned models across multiple domains and frameworks. Why does it work? It uses inter-label dependencies resembling a classifier chain approach in ML.
(4/n)
MoVa provides resources defining this generalizable classification -- 16 labeled datasets and benchmarking results across four major, theoretically-grounded frameworks: Moral Foundations Theory (MFT), Human Values, Common Morality, and Morality-as-Cooperation (MAC)
(5/n)
MoVa also offers a new application for evaluating psychological surveys:
By using MoVa to score the relevance of moral dimensions for each survey item, we can detect potentially multi-loaded items in instruments like MFQ, MAQ, and PVQ, helping researchers rethink questionnaire design.
So what? This technology could help us:
(1) Analyze Public Discourse: Understand the core values driving large-scale conversations on social and political issues.
(2) Build Better AI: Ensure that artificial intelligence systems communicate in a way that's aligned with basic human ethics #AIalignment
(6/n)
Future work? Generalizable classification across cultures and languages, and investigating generalisable prompting methodology on other subjective text classification tasks.
(7/n)
Led by CMlab PhD student Ziyu Chen, with @ml4x.bsky.social Junfei Sun, Chenxi Li at UChicago, @joshnguyen.bsky.social at UPenn, and Jing Yao, Xiaoyuan Yi and Xing Xie at Microsoft Research Asia
(8/n)
Read the paper here arxiv.org/abs/2509.24216
Explore the MoVa resources, data, and code supporting this work here: 👉 github.com/ZiyuChen0410...
#NLP #ResearchTools #DataScience
Identifying human morals and values in language is crucial for analysing lots of human- and AI-generated text.
We introduce "MoVa: Towards Generalizable Classification of Human Morals and Values" - to be presented at @emnlpmeeting.bsky.social oral session next Thu #CompSocialScience #LLMs
🧵 (1/n)
😂
We'll focus on complex information needs with a purpose: on climate change and beyond.
We are open to new algorithms, paradigms for human-AI collaboration, innovations with LLM
2x Research Fellows positions on NLP and ML working w wonderful Jing Jiang, Thang Bui and myself + partners in Data61.
jobs.anu.edu.au/jobs/researc...
Apply by 31 Dec or help spread the words!
(more details below)
