On the use of Silver Standard Data for Zero-shot Classification Tasks in Information Extraction

Source code for LREC-COLING 2024 paper titled "On the use of Silver Standard Data for Zero-shot Classification Tasks in Information Extraction".

Figure 1: The diagram shows the procedure of Clean-LaVe in the zero-shot relation extraction task

• We propose Clean-LaVe to first detect a small amount of clean data which are later used to finetune the pre-trained model. We then use the finetuned model to infer the categories on the test data.
• We propose a clean data detection module that enhances the selection process through Iteratively Weighted Negative Learning and Class Aware Data Selector.
• The experimental results demonstrate that our method can outperform the baseline by a large margin on various zero-shot classification tasks

Requirements

Install the necessary packages with:

$ pip install -r requirements.txt

Usage

• There three infomation extraction task (RE, EAC and multi-lingual RE).
  (For more detailed information, please consult the paper.)

• Correspondence between directory and tasks:

  RE task: ./Clean_LaVe4RE
  EAC task: ./Clean_LaVe4EAC
  Multi-lingual RE: ./Clean_LaVe4REMul
  

• For conducting Clean-LaVe on these three information extraction task, pls refer to the readme.md file under the corresponding directory.

Dataset

• We have released a pre-processed dataset on https://huggingface.co/datasets/wjwilliam/Clean_LaVe to ensure reproducibility.