Prompting, Decoding, Embedding: Leveraging Pretrained Language Models for High-Quality and Diverse Open Rule Induction

Wangtao Sun, Shizhu He, Jun Zhao, Kang LiuIEEE Transactions on Audio, Speech and Language Processing 2025 · OpenReview

Abstract

Open rule induction (OpenRI) is devoted to obtaining reasoning rules expressed in natural languages. Compared with traditional rules with predefined logical symbols and domain predicates, open rules are more expressive and easier to use in real-world applications. Owing to the rich knowledge content and powerful text generation ability of pretrained language models (PLMs), researchers have begun adapting PLMs to induce open rules in recent years. However, the previous OpenRI method suffered from the bias of PLMs, thus usually generating tedious and redundant open rules. To alleviate the above issues, this paper proposes a novel framework, named Quadori (high-Quality and diverse open rule induction), which leverages the power of PLMs by prompting, decoding, and embedding to increase the quality and diversity of the induced open rules. Specifically, entity-type-based prompts and sampling-supported beam search are proposed for obtaining support instances according to the given rule head. A determinantal point process is then employed to generate rule bodies from the assembly support instances. Adequate experiments on a public open rule induction dataset and six relation extraction datasets show that Quadori is superior to the existing methods in both quality and diversity
BibTeX
@article{sun2025prompting,
  author = {Wangtao Sun and Shizhu He and Jun Zhao and Kang Liu},
  title = {{Prompting, Decoding, Embedding: Leveraging Pretrained Language Models for High-Quality and Diverse Open Rule Induction}},
  journal = {IEEE Transactions on Audio, Speech and Language Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=WhIsqDWjhi},
}

Loading related content...

Figure