This the repository for the paper "GPT-RE: In-context Learning for Relation Extraction using Large Language Models"
This repository now supports GPT-Random, GPT-SimCSE and the reasoning logic in the paper, GPT-RE_FT method is not supported in the current version, we will soon update.
bash run_relation_ace.sh
In the example script file run_relation_ace.sh, there are multiple options:
--task: name of the task, eg. semeval
--model: name of the model used in gpt3, eg. text-davinci-003
--num_test: number of examples in the test subset, eg. 100, this should be smaller than the length of the test dataset
--example_dataset: the file to choose demonstrations
--test_dataset: the path to the test data, note that it should be changed to "test.json" if you want to work on the whole test data
--fixed_example: if 1, then the demonstration examples will be fixed during the test; if 0, then for each test example, the demonstrations will be retrieved again. In kNN setup, it should be 0
--fixed_test: if 1, then the test dataset will be fixed (this option can be ignored as you can directly change test data by --test_dataset
--num_per_rel: the number of examples per relation type chosed to be the demonstrations, this should be 0 in kNN setup
--num_na: the number of NA examples chosed to be the demonstrations, this should be 0 in both w/o NA setup and kNN setup
--num_run: keep 1
--seed: random seed
--random_label: if 1, the model will change the gold lables to random labels in the demonstration
--reasoning: if 1, the model will add reasoning to the demonstrations
--use_knn: if 1, use kNN in demonstrations
--k: top k in kNN
--var: you can igore it, keep 0
--reverse: reverse the order of demonstrations, if 0, the default order is that more similar demonstrations will be placed at the top.
--verbalize: you can ignore it, keep 0
--entity_info: entity-aware sentence similarity in our paper
--structure: a trial to use structured prompt, you can ignore it
--use_ft: the fine-tuned representation, this option is not yet supported in the current version
--self_error: you can ignore it ,keep 0
--use_dev, store_error_reason, discriminator: keep 0