This is the github repository for the paper to be appeared at NAACL 2024 main conference: Self-Improving for Zero-Shot Named Entity Recognition with Large Language Models.

A self-improving framework for zero-shot named entity recognition (NER) with large language models (LLMs), which utilizes unlabeled data to stimulate the self-learning ability of LLMs via a training-free process.

This work pushes the performance boundary of zero-shot NER with LLMs by proposing a training-free self-improving framework, which utilizes an unlabeled corpus to stimulate the self-learning ability of LLMs. The pipeline of the proposed framework is:

• First, we use the LLM to make predictions on the unlabeled corpus using self-consistency (SC) and obtain a self-annotated dataset.
• Second, we explore various strategies to select reliable annotations to form a reliable self-annotated dataset.
• Finally, for each test input, we retrieve demonstrations from the reliable self-annotated dataset and perform inference via in-context learning.

Please find more details of this work in our paper.

• code: All python code
  • standard: Regular zero-shot code.
  • self_consistent_annotation: Self-improving code.
• data: Data file, including self-annotated demonstrations generated by the LLM itself.
• prompts: Prompts file, generated during running, used to prompt the LLM to perform NER.
• results: Results file, generated during running, including model output and result metrics.
• scripts: Bash script to run the code.

We run our code on Windows. The following are the dependencies: python 3.8, openai 0.27.4, pytorch 2.0.1, pandas, hanlp

We provide processed WikiGold dataset in this repository for quick start, along with its generated prompts, self-annotation results and inference results.

In data folder of wikigold:

• *GPTEmb.npy: Sample embeddings obtained from text-embedding-ada-002.
• train_shuffle_42*: Files related to training set shuffled with seed 42.
• self_consistent_annotate: Directory for the generated self-annotated demonstrations. Here are the meaning. Below are the meanings of different string tags in the file name of the provided demonstration files.
  • std: The default standard prompt method used in this paper. Custom prompt methods can be added by modifying DesignPrompts.py in standard and self_consistent_annotation.
  • c5: Self-consistency with sampling quantity set to 5.
  • all_th_spl_4.0_296: Keep all entities without filtering, then do sample-level threshold filtering with threshold set to 4.0, and the number of filnal retained samples is 296.
  • th_ent_4.0_500: Entity-level threshold filtering with threshold set to 4.0, no sample-level filtering, and the number of retained samples is 500.

We also provide other processed datasets used in our paper at the Google Drive, except ACE05 and Ontonotes 4 for copyright reasons. You can download and unzip the dataset files and put them in the data folder.

Run the below command to generate embeddings with OpenAI API.

python code/standard/GenerateEmbsGPT.py --dataname wikigold --datamode test --emb_model text-embedding-ada-002 --emb_encoding cl100k_base

We provide shell scripts in folder scripts for quick start.The generated prompts will be saved to folder prompts. The response from the LLM and the evaluation results will be saved to folder result.

Run the following commands in the provided order to use our methods. Before you run on ChatGPT, please set your OpenAI API Keys in my_openai_api_keys in const.py.

# --- No-demos ---
sh scripts/wikigold_no_demos.sh

# --- Self-improving ---
# Self-annotating and two-stage majority voting
sh scripts/wikigold_1_self_annotate_TSMV.sh
# Entity-level threshold filtering
sh scripts/wikigold_2_entity_level_sel.sh
# Sample-level threshold filtering
sh scripts/wikigold_2_sample_level_sel.sh
# Inference with Diverse nearest with SC ranking
sh scripts/wikigold_3_test_inference.sh

Self-consistency (SC)

• --consistency. 0 means no SC, 1 means using SC with two-stage majority voting.
• --temperature. 0 for not using SC; default set to 0.7 for SC.
• --query_times. Number of sampled responses in SC. We defaultly set to 5 in our main experiments.

Reliable annotation selection

• --entity_level_selection. all means all entities. th_ent means entity-level SC threshold filtering.
• --sample_level_selection. th_spl means sample-level SC threshold filtering.

Inference

• --demo_retrieval_method. "GPTEmbDvrsKNN" means diverse nearest.
• --diverseKNN_number: the value of K in the diverse nearest.
• --diverseKNN_sampling: random means random sample from K neighbors. Sc means select top-k SC score samples from K neighbors.

@misc{xie2023selfimproving,
      title={Self-Improving for Zero-Shot Named Entity Recognition with Large Language Models}, 
      author={Tingyu Xie and Qi Li and Yan Zhang and Zuozhu Liu and Hongwei Wang},
      year={2023},
      eprint={2311.08921},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}