🧬 LLMs4OL 2025 – Task C1 (OBI)

Hybrid Embedding–Clustering–LLM Pipeline for Taxonomy Discovery

This repository contains our solution for Task C1 – Ontology for Biomedical Investigations (OBI) in the LLMs4OL 2025: Large Language Models for Ontology Learning Challenge.

The goal is to automatically discover taxonomic (is-a) relationships between biomedical investigation types.

Given only a list of ontology types, the system predicts parent → child hierarchical relations and outputs them in the required _pairs.json format.

---

🧠 Methodology Overview

Our approach is a three-stage hybrid pipeline:

1. Sentence Embeddings → semantic representation
2. Clustering → local ontology grouping
3. Large Language Model (LLM) → hierarchical reasoning

This design reduces noise, improves scalability, and enables directional is-a prediction.

---

🖼️ Methodology Diagram

Replace methodology.png with your actual diagram filename

---

⚙️ Pipeline

Stage 1 — Sentence Embeddings

• Model: sentence-transformers/all-MiniLM-L6-v2
• Converts each ontology type into a dense vector
• Captures semantic similarity between biomedical terms

Stage 2 — Clustering

• Algorithm: MiniBatch K-Means
• Cluster size ≈ 50 terms per cluster
• Purpose:
  • Reduce search space
  • Group semantically related types
  • Enable local hierarchy discovery

Stage 3 — LLM-based is-a Prediction

• Model: unsloth/Qwen3-1.7B-unsloth-bnb-4bit
• Quantized (4-bit) for memory efficiency
• Candidate parent–child pairs generated using Top-K nearest neighbors
• LLM performs Yes/No classification for is-a relations
• Batch inference used for faster processing

---

🧪 Implemented Approaches

1️⃣ Cosine Similarity Baseline

File: cosine-similarity.ipynb

Pipeline:

• Sentence-BERT embeddings
• Pairwise cosine similarity
• Threshold = 0.95 → predict is-a

Limitation: captures similarity but cannot model hierarchy direction

---

2️⃣ K-Means Clustering + Few-Shot LLM

File: kmeansclustuering-llm.ipynb

Pipeline:

• Embeddings → clustering
• Top-K candidate pairs
• Few-shot prompting with biomedical examples
• LLM classifies is-a relations

---

3️⃣ Embeddings + Clustering + Zero-Shot LLM (Final)

File: sentence-embeddings-llm.ipynb

Pipeline:

• Embeddings
• MiniBatch K-Means clustering
• Nearest neighbor candidate generation
• Zero-shot LLM classification
• JSON output in required format

This is the final and most scalable system.

---

📂 Repository Structure

. ├── cosine-similarity.ipynb ├── kmeansclustuering-llm.ipynb ├── sentence-embeddings-llm.ipynb ├── methodology.png └── README.md

Data

The dataset is publicly available from the official challenge repository:

🔗 https://github.com/sciknoworg/LLMs4OL-Challenge/tree/main/2025/TaskC-TaxonomyDiscovery/OBI

File	Split	Description
obitrainpairs.json	Train	8,249 labeled is-a pairs
obitraintypes.txt	Train	4,237 unique ontology types
obitesttypes.txt	Test	2,821 ontology types

Do not re-upload the data to this repo. Download directly from the source above.

How to Run

1. Clone the official challenge repo or download the OBI files directly
2. On Kaggle: upload the files as a dataset named data-ontology
3. Open any notebook and run all cells
4. Output saved to outputs/submission_pairs.json [ {"parent": "distillation", "child": "simple distillation"} ]