Work

DAC: Code for Query-Based Ontology Alignment and Three-Way Calibration for Zero-Shot Event Detection

Year: 2026
Type: software
Abstract: Code and execution records for the manuscript “Query-Based Ontology Alignment and Three-Way Calibration for Zero-Shot Event Detection” by Xianglong Fang and Lamei Liu (revised version submitted to Scientific Reports). The record contains all releases of the DAC (Discover–Align–Calibrate) code, the changes we applied to the official DiCoRe code, the execution records of every run reported in the paper, and the scripts that regenerate all tables of the paper and of the Supplementary Information from these records. README.md describes every file. Files. DAC-code-v2.0.0.zip … DAC-code-v2.0.8.zip: all releases of the DAC code. Each archive contains a CHANGELOG.md. dicore_patches.zip: the eight local changes applied to the official DiCoRe repository (commit b6637021e2b5a4c6aa72b3bebfb8fed0c6260a01) to run it in our environment (Appendix B of the paper). records_*.zip (9 files): execution records of all DAC runs and of the re-executed DiCoRe runs (Appendix D of the paper), without the input texts. Unzip all nine files into the same folder; together they form the folder records/. analysis_scripts.zip: scripts that regenerate every table of the paper and of the Supplementary Information, and Figure 1, from the records. Code releases. The reported runs were produced with v2.0.2, v2.0.3, v2.0.4, v2.0.6 and v2.0.8. The run_manifest.json of every run lists the SHA-256 hash of each source file, which identifies the release. Prompts, decoding settings, parsing and scoring of each reported condition are the same in all releases that ran it. v2.0.5 and v2.0.7 add backends that were not used for any reported run; v2.0.0 and v2.0.1 precede the reported runs. Execution records. There are 298 DAC run records. Each run directory contains the run manifest (start and end times, command, host and GPUs, package versions, hashes of the source, model and data files, all settings including the seed, and hashes of all outputs), the parsed and raw outputs of each stage, the ontology order shown for every input, every Calibrate decision with its raw response, and the output of the official scorer; API runs also log every request. The prompt files were removed because they contain the input texts of the evaluation sets; their SHA-256 hashes remain in each run_manifest.json, and the prompts can be regenerated with the code and the evaluation data. Model outputs occasionally quote short spans of the input; these are kept. Relation to the previous version. The previous version (source-snapshot-2026-09-17, doi:10.5281/zenodo.22807157) archived seven source files of an earlier working copy (“2.28”). That snapshot did not contain the query-dependent ordering of the ontology described in the manuscript, and its final verification stage used the shared sampling temperature, a 50-token limit and substring matching of responses. The results of the revised manuscript were produced with the releases in this record; results reported in earlier versions of the manuscript were not. Data. The evaluation subsets (MAVEN, FewEvent, ACE, GENIA, SPEED and CASIE; 1,300 input windows) are those released with DiCoRe (https://github.com/PlusLabNLP/DiCoRe). They are not included here; access and use are subject to the terms of the original datasets. No model weights or credentials are included. Attribution. DAC builds on DiCoRe (Parekh et al., EMNLP 2025). dac/evaluate_llm_preds.py and dac/scorer.py are unchanged copies of the TextEE evaluation code used by DiCoRe (Huang et al., Findings of ACL 2024). Upstream authorship is retained.DiCoRe paper: https://aclanthology.org/2025.emnlp-main.1038/DiCoRe source: https://github.com/PlusLabNLP/DiCoReTextEE: https://github.com/ej0cl6/TextEE
Language: English
Open Access status: green
OpenAlex ID: W7213507361
Cites:
Cited by:
Related to: 10