Published October 4, 2026 | Version v1

CHASTE: Constrained Holistic Annotator System with onTological Entities

Description

Entity and Relationship extraction remain key challenges for information extraction in creating knowledge bases from textual resources. This challenge is especially pronounced in domain-specific applications, where systems need to be fine-tuned and tailored to specific needs. This paper, therefore, aims to enhance zero-shot or simple training approaches for knowledge extraction within the GutbrainIE Task 2026 in the Conference and Labs of the Evaluation Forum (CLEF). Hence, we propose a generative grammar-based constrained generation method informed by ontological and natural-language preprocessing, Constrained Holistic Annotator System with onTological Entities (CHASTE), based on Language Models (LMs). The extraction system is evaluated in both fine-tuned and zero-shot settings for entity and relation extraction subtasks. Furthermore, a modified beam search enables the retrieval of joint probabilities over a sequence of tokens, theoretically enhancing grammar-constrained output through a holistic probabilistic generative approach. The results of this evaluation are compared, both within our submission, including a naive simple baseline, and to the official baseline results. A merged run that intersects our and Graphwise teams’ results extends these results by leveraging our precise pipeline and securing a top-precision spot in subtask 6.1.2 (Named Entity Recognition and Disambiguation (NERD)). These comparisons show that, while the approach is underperforming, it offers potential for low-training pipelines with easy adaptation, requiring no fine-tuning of a large LM, only our specialized generative pipeline.

Files

paper21.pdf

Files (1.3 MB)

Name Size
md5:75f6e2cf72c83b4d37c3c4f10fddd369
1.3 MB Preview Download

Additional details

Funding

European Commission
HEREDITARY - HetERogeneous sEmantic Data integratIon for the guT-bRain interplaY 101137074