Reproduce “Towards a Unified Multi-Dimensional Evaluator for Text Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ming Zhong, Yang Liu, Da Yin, Yuning Mao, Yizhu Jiao, Pengfei Liu, Chenguang Zhu, Heng Ji and 1 more
A unified multi-dimensional evaluator for text generation, UNIEVAL, re-frames evaluation as a Boolean Question Answering task to provide fine-grained quality assessments across various dimensions. Developed by researchers from UIUC, Microsoft, UCLA, and CMU, it achieves over 23% higher correlation with human judgments than previous state-of-the-art unified evaluators in summarization and over 43% in dialogue generation, demonstrating strong zero-shot transfer capabilities.
Key insights
- Formulating multi-dimensional evaluation as Boolean Question Answering enables a unified approach and facilitates knowledge incorporation from diverse NLP tasks.
- Intermediate multi-task learning significantly enhances the model's ability to capture general knowledge relevant to text evaluation, compensating for limited human-annotated data.
- A continual learning strategy with data replay mitigates negative transfer between different evaluation dimensions, improving comprehensive performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.