Enrich-on-Graph

[Paper][EMNLP 2025] Enrich-on-Graph: Query-Graph Alignment for Complex Reasoning with LLM Enriching

🎨 Overview

🚀 Update

• Our paper is accepted by EMNLP 2025 🎉🎉🎉
• Our code will be released before September 30th 2025 ✨
• Our paper will be published soon ⏳

🛠️ Environment

conda create -n enrich python=3.10
conda activate enrich
pip install -r requirements.txt

📌 Enrich-on-Graph Workflow

NOTE: Please set your DATASET_PATH (e.g. datasets/RoG-cwq), OPENAI_API_KEY, BASE_URL in the following script. (src/workflow/)

DATASET_PATH=DATASET_PATH_HERE.parquet
OPENAI_API_KEY=YOUR_API_KEY_HERE
BASE_URL=API_BASE_URL_HERE

Step 1: Parsing

sh src/workflow/parsing.sh

This action will output question_decompose_datasets containing user_queries (sub-queries of complex question) field in preprocess_datasets/ folder.

Step 2: Pruning

sh src/workflow/pruning.sh

This action will output llm_pruning_three_channels_datasets containing pruned_graph field; as well as triple_trans_datasets containing triple_unit_queries (graph queries) field.

We also provide vanilla pruning script, which will output vanilla_pruning_datasets containing pruned_graph field.

python run.py --task vanilla_pruning -d {webqsp, cwq} --embedding_model sentence-transformers  --embedding_model_path {YOUR EMBEDDING MODEL PATH}

Step 3: Enriching

sh src/workflow/enriching.sh

This action will output filter_triple_datasets containing filter_triples (triples needed to be enriched) field; as well as structural_enrich_datasets containing structural_enrich_triples field; also feature_enrich_datasets containing feature_enrich_triples field.

🏆 Enriched KG

We provide our preprocessed datasets (cwq, webqsp) containing enriched KGs in here (uploading, coming soon).

💡 Reasoning

Here choose to reason with enriched KGs or vanilla KGs.

sh src/workflow/reasoning.sh

Type 1 represents Reasoning on vanilla KGs, 2 represents Reasoning with our EoG. 3 represents exit Reasoning.

This action will output qa_datasets containing predictions field.

📈 Evaluation

Use the following script to eval reasoning results.

python run.py --task qa_evaluate -d {QA_DATASET_PATH} --cal_f1

💖 Citations

@inproceedings{li-etal-2025-enrich,
    title = "Enrich-on-Graph: Query-Graph Alignment for Complex Reasoning with {LLM} Enriching",
    author = "Li, Songze  and
    Liu, Zhiqiang  and
    Gui, Zhengke  and
    Chen, Huajun  and
    Zhang, Wen",
    editor = "Christodoulopoulos, Christos  and
    Chakraborty, Tanmoy  and
    Rose, Carolyn  and
    Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.390/",
    pages = "7683--7703",
    ISBN = "979-8-89176-332-6"
}