CrossAligner

CrossAligner, a two-stage LLM-based methodology for discovering directed semantic bridges across three food-domain ontologies.

Pipeline overview

1. Stage 1 — candidate selection: an LLM performs binary screening and retains concept pairs that are semantically related.
2. Stage 2 — semantic-bridge generation: an LLM reassesses the retained pairs using richer ontology context and produces a directed natural-language relationship.

The included experiments cover:

• MeSH → ONS
• OccO → MeSH
• OccO → ONS

Stage 2 supports source-to-target and target-to-source directions, each with zero-shot and few-shot prompts.

Repository structure

CrossAligner/
├── assets/food-onto/       # MeSH, OccO, and ONS ontology modules
├── bash/                   # Slurm scripts for Stage 1 and Stage 2
├── data/ground_truth/      # Expert reference mappings
├── ontoaligner/            # Pipeline implementation
├── results/
│   ├── Stage_1/            # Candidate-selection outputs
│   ├── Stage_2/            # Semantic-bridge outputs
│   └── CrossAligner_Evaluation_results/
│       ├── figures/        # Generated evaluation figures
│       └── tables/         # Generated evaluation tables
├── evaluation/
│   └── CrossAligner_Evaluation_3RQs_EXECUTED.ipynb
├── evaluation_core.py      # Evaluation functions and metrics
├── requirements.txt
└── pyproject.toml

Requirements

The experimental configuration was developed for:

• Python 3.10 or 3.11
• an NVIDIA CUDA GPU
• Java 11 for DeepOnto/OWLAPI
• a Hugging Face account with access to the selected gated model

Install the required packages: pip install -r requirements.txt

Check Java and the CrossAligner imports:

java -version from ontoaligner.pipeline import OntoAlignerPipeline from ontoaligner.encoder import ConceptCandidateLLMEncoder, ConceptLLMEncoder

## Running Stage 1

Submit the script for the required ontology pair:

```bash
sbatch bash/Qwen3-32B-raw_s_f_DeepOnto_mesh_ons_stage1.bash
sbatch bash/Qwen3-32B-raw_s_f_DeepOnto_occo_mesh_stage1.bash
sbatch bash/Qwen3-32B-raw_s_f_DeepOnto_occo_ons_stage1.bash

Stage 1 writes JSON and CSV outputs to:

results/Stage_1/<model>/<ontology_pair>/

Before starting Stage 2, confirm that the selected candidate JSON exists in that directory.

Running Stage 2

Each Stage-2 script contains two experiment selectors:

experiment_name = "source_to_target_zero_shot"
stage1_run = "qwen"

Available experiment_name values are:

• source_to_target_zero_shot
• target_to_source_zero_shot
• source_to_target_few_shot
• target_to_source_few_shot

Available stage1_run values are qwen and llama.

Submit the selected ontology pair:

sbatch bash/Qwen3-32B-raw_s_f_DeepOnto_stages2_mesh_ons.bash
sbatch bash/Qwen3-32B-raw_s_f_DeepOnto_stages2_occo_mesh.bash
sbatch bash/Qwen3-32B-raw_s_f_DeepOnto_stages2_occo_ons.bash

Stage-2 outputs are stored below:

results/Stage_2/<ontology_pair>/

Running the evaluation:

The evaluation reads the Stage 1 and Stage 2 outputs directly from:

results/Stage_1/
results/Stage_2/

Run the notebook evaluation/CrossAligner_Evaluation_3RQs_EXECUTED.ipynb. Generated evaluation tables and figures are written to:

results/CrossAligner_Evaluation_results/
├── figures/
└── tables/

The generated filename records the Stage-1 model, Stage-2 model, ontology pair, prompt mode, and direction.

Acknowledgement and license

CrossAligner builds on OntoAligner by Babaei Giglou et al. The original copyright notices are retained in inherited files.

The project is distributed under the Apache License 2.0.