Skip to content

Latest commit

 

History

History
101 lines (75 loc) · 6.14 KB

README.md

File metadata and controls

101 lines (75 loc) · 6.14 KB
Raw

Extract, Define, Canonicalize: An LLM-based Framework for Knowledge Graph Construction

This reposistory contains the implementation of the knowledge graph construction framework EDC proposed in the paper Extract, Define, Canonicalize: An LLM-based Framework for Knowledge Graph Construction.

@article{zhang2024extract,
  title={Extract, Define, Canonicalize: An LLM-based Framework for Knowledge Graph Construction},
  author={Zhang, Bowen and Soh, Harold},
  journal={arXiv preprint arXiv:2404.03868},
  year={2024}
}

Please consider citing our work, if you found the provided resources useful.

If you have any question, please leave an issue or directly contact Bowen.

High-level Idea of EDC

EDC uses a three-phase framework to approach the task of knowledge graph construction (KGC) from input text. It first applies open information extraction to extract open knowledge graphs and then performs post-hoc canonicalization using the definition of the schema induced by the open KG generated by LLMs. This approach is able to scale to large realistic schemas compared to previous approaches while still requiring no parameter tuning. Currently, EDC only canonicalizes relations, we plan to extend EDC to other schema components, e.g. entity types, in the future. It is also to be noted that EDC focuses on the relation extraction problem in KGC, for a more comprehensive tool, it is advisable to combine EDC with other tools, including coreference resolution, chunking, entity mention canonicalization, etc.

Here is an illustration of how EDC works:

Our experiments with both automatic and human evaluation demonstrate that EDC achieves state-of-the-art performance on the tasks of knowledge graph construction and scales to a previously impossible size of schema.

Iterative Refinement with Schema Retriever

EDC can be iteratively refined with Schema Retriever, which is trained in the same fashion as information retriever. It is able to retrieve schema components relevant to the input text. It is able to extract those more obscure schema components harder to identify on the surface. The retrived content, together with entities extracted with Entity Extraction and the entities and relations extracted from last run, together form a hint to enhance the performance of open information extraction. To train the Schema Retriever, download the TEKGEN dataset from link and run

python collect_schema_retrieval_data.py \
    --tekgen_path /path/to/tekgen
    --relation_definition_csv_path /output/path/to/tekgen/relation/definitions
    --dataset_size N
    --output_path /output/dataset/path

to prepare a dataset and refer to this repository for how to finetune the model.

Dependencies

Install requirements using a conda environment:

conda env create -f environment.yml
conda activate edc

Usage Instructions

Run EDC by executing python run.py --options, where the possible arguments to tweak EDC and their usages are explained in run.py.

To replicate the Target Alignment experiments in the paper, run the following command:

python run.py \
    --oie_llm {oie_llm} \
    --oie_few_shot_example_file_path ./few_shot_examples/{dataset}/oie_few_shot_examples.txt \
    --sd_llm gpt-3.5-turbo \
    --sd_few_shot_example_file_path ./few_shot_examples/{dataset}/sd_few_shot_examples.txt \
    --sc_llm gpt-3.5-turbo \
    --sc_embedder intfloat/e5-mistral-7b-instruct \
    --input_text_file_path ./datasets/{dataset}.txt \
    --target_schema_path ./schemas/{dataset}_schema.csv \
    --output_dir ./output/{dataset}_target_alignment

To replicate the Self Canonicalization experiments in the paper, run the following command:

python run.py \
    --oie_llm {oie_llm} \
    --oie_few_shot_example_file_path ./few_shot_examples/{dataset}/oie_few_shot_examples.txt \
    --sd_llm gpt-3.5-turbo \
    --sd_few_shot_example_file_path ./few_shot_examples/{dataset}/sd_few_shot_examples.txt \
    --sc_llm gpt-3.5-turbo \
    --sc_embedder intfloat/e5-mistral-7b-instruct \
    --input_text_file_path ./datasets/{dataset}.txt \
    --enrich_schema \
    --output_dir ./output/{dataset}_self_canonicalization

To apply iterative refinement, first train the schema retriever as aforementioned and pass the following extra arguments:

--sr_adapter_path /path/to/trained/adapter (OR --sr_embedder /your/own/embedder)
--oie_refine_few_shot_example_file_path ./few_shot_examples/{dataset}/oie_few_shot_refine_examples.txt \
--ee_llm gpt-3.5-turbo \
--ee_few_shot_example_file_path ./few_shot_examples/{dataset}/ee_few_shot_examples.txt \
--refinement_iterations N \

where oie_llm can take value from gpt-3.5-turbo, gpt-4 and mistralai/Mistral-7B-Instruct-v0.2 (you may use other OpenAI models or other transformers models, but they were not tested) and dataset can take value from webnlg, rebel and wiki-nre. sc_embedder is by default intfloat/e5-mistral-7b-instruct. You may use any sentence transformer of you choice (please refer to https://sbert.net/) but they are not tested. Similarly, for refinement, you can pass the path to the finetuned model using sr_adapter_path or any sentence transformer using sr_embedder. It is to be noted the default embedders used for schema canonicalization and refinement demand significant compute (7B), you may consider using smaller models or even finetune on top of them if your compute is limited (and the LLMs for prompting as well).

You may use EDC on customized input and target schema by following the formats used in datasets and schemas. You may also tweak the prompt templates and few-shot examples used by changing prompt_templates and few_shot_examples. To be noted, if you would like to use OpenAI models, please set the environment variable OPENAI_KEY to your own API key. An example command to run EDC on an example dataset is given in run.sh.

Evaluation

Please refer to evaluate folder and its README.