Skip to content

Latest commit

 

History

History
121 lines (109 loc) · 5.88 KB

README.md

File metadata and controls

121 lines (109 loc) · 5.88 KB
Raw

Knowledge Graph Triplet Extraction with LLMs

A pipeline for extracting knoweldge triplets from text using LLMs and Knowledge Bases. Refer to Zero- and Few-Shots Knowledge Graph Triplet Extraction with Large Language Models for the details. If you find this useful please consider citing us:

@inproceedings{papaluca-etal-2024-zero,
    title = "Zero- and Few-Shots Knowledge Graph Triplet Extraction with Large Language Models",
    author = "Papaluca, Andrea  and
      Krefl, Daniel  and
      Rodr{\'\i}guez M{\'e}ndez, Sergio  and
      Lensky, Artem  and
      Suominen, Hanna",
    editor = "Biswas, Russa  and
      Kaffee, Lucie-Aim{\'e}e  and
      Agarwal, Oshin  and
      Minervini, Pasquale  and
      Singh, Sameer  and
      de Melo, Gerard",
    booktitle = "Proceedings of the 1st Workshop on Knowledge Graphs and Large Language Models (KaLLM 2024)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.kallm-1.2",
    pages = "12--23",
    abstract = "In this work, we tested the Triplet Extraction (TE) capabilities of a variety of Large Language Models (LLMs) of different sizes in the Zero- and Few-Shots settings. In detail, we proposed a pipeline that dynamically gathers contextual information from a Knowledge Base (KB), both in the form of context triplets and of (sentence, triplets) pairs as examples, and provides it to the LLM through a prompt. The additional context allowed the LLMs to be competitive with all the older fully trained baselines based on the Bidirectional Long Short-Term Memory (BiLSTM) Network architecture. We further conducted a detailed analysis of the quality of the gathered KB context, finding it to be strongly correlated with the final TE performance of the model. In contrast, the size of the model appeared to only logarithmically improve the TE capabilities of the LLMs. We release the code on GitHub for reproducibility.",
}

Requirements

llama-index
langchain
transformers
bs4
nltk
sklearn
tqdm
matplotlib
lxml
networkx
torch

Datasets

WebNLG

NYT

  • nyt/: the New York Times dataset by Riedel et al

Building the Knowledge Base

The Knowledge Base can be built by running the kb_builder.py script specifying which data to use. For example to build the KB associated to the train and valid splits of the WebNLG dataset, just run:

python kb_builder.py --data webnlg_modified/{train,valid}.json

This will create two indices webnlg_modified/kb_triplets/ and webnlg_modified/kb_sentence-triplets/. The first one stores separately each triplet found in the provided data (0.5-shots):

Node-1
------
    (Egg Harbor Township , New Jersey, isPartOf, Atlantic)
Node-2
------
    (Egg Harbor Township , New Jersey, isPartOf, New Jersey)
Node-3
------
    (Atlantic City International Airport, location, Egg Harbor Township , New Jersey)

the second one stores in each node the original sentence followed by the triplets it contains (few-shots):

Node-1
------
    Atlantic City International Airport in Egg Harbor Township , New Jersey is in the U.S.A . The airport has a runway that is 1,873 long .
    (Egg Harbor Township , New Jersey, isPartOf, Atlantic)
	(Egg Harbor Township , New Jersey, isPartOf, New Jersey)
	(Atlantic City International Airport, location, Egg Harbor Township , New Jersey)

It's recommended to pass also the --normalize argument to the script in order to normalize sentences and triplets (e.g. removing unwanted punctuation, moving everything to the lower case, etc...).

Running the Benchmark

The benchmark.py script, instead, runs the triplet extraction pipeline on the specified data with the specified LLM and prompt. For example to test the gpt2 model on the WebNLG dataset with the base prompt:

python benchmark.py \
    --data webnlg_modified/test.json \
	--conf model_conf/gpt2.conf \
	--prompt prompts/prompt_base.json

The configurations of the different LLMs can be found under model_conf/. Note that for the gpt3-turbo and gpt4 models, an OpenAI key is required, please place it in the openai_key.txt file as the only line. Upon completion, the extracted triplets are going to be saved under webnlg_modified/gpt2/prompt_base/generated_triples_gpt2_temp-0.1.xml. By specifying the --kb argument it is possible to make the LLM rely on the Knowledge Base built above, remember to also set the appropriate prompt, e.g. prompt_base_kb.json:

python benchmark.py \
    --data webnlg_modified/test.json \
	--conf model_conf/gpt2.conf \
	--kb webnlg_modified/kb_triplets/ \
	--prompt prompts/prompt_base_kb.json \
	--top_k 5

Additionally the number of nodes of the KB index to retrieve can be specified with the --top_k argument.

Performance Evaluation

To evaluate the extracted triplets use the Evaluation_script.py, that is adapted from the evaluation script that was provided originally for the WebNLG challenge https://github.com/WebNLG/WebNLG-Text-to-triples. For example, to evaluate the triplets generated by the gpt2 model shown above, first, generate the groundtruth (webnlg_modified/groundtruth_triples.xml) by simply running:

python benchmark.py \
    --data webnlg_modified/test.json \
	--groundtruth

and then evaluate the micro/macro averaged performance:

python Evaluation_script.py \
    --predictions webnlg_modified/gpt2/prompt_base/generated_triples_gpt2_temp-0.1.xml \
	--groundtruth webnlg_modified/groundtruth_triples.xml \
	--avg micro