README.md

---
language:
- en
license: mit
task_categories:
- feature-extraction
- text-classification
tags:
- anomaly-detection
- nlp
- benchmark
---

# NLP-ADBench: NLP Anomaly Detection Benchmark

This repository contains **NLP-ADBench**, the most comprehensive NLP anomaly detection (NLP-AD) benchmark to date. It is a comprehensive benchmarking tool designed for Anomaly Detection in Natural Language Processing (NLP), establishing a benchmark and introducing 8 curated and transformed datasets derived from existing NLP classification datasets. These datasets are specifically tailored for NLP anomaly detection tasks and presented in a unified standard format to support and advance research in this domain.

The benchmark includes results from 19 algorithms applied to the 8 NLPAD datasets, categorized into two groups:
- 3 end-to-end algorithms that directly process raw text data to produce anomaly detection outcomes.
- 16 embedding-based algorithms, created by applying 8 traditional anomaly detection methods to text embeddings generated using two models: BERT's `bert-base-uncased` (**BERT**) and OpenAI’s `text-embedding-3-large` (**OpenAI**).

Paper: [NLP-ADBench: NLP Anomaly Detection Benchmark](https://huggingface.co/papers/2412.04784)
Code: https://github.com/USC-FORTIS/NLP-ADBench

![Performance comparison of 19 Algorithms on 8 NLPAD datasets using AUROC](https://github.com/USC-FORTIS/NLP-ADBench/blob/main/figs/benchmark.png)

## NLPAD Datasets

The datasets required for this project can be downloaded from the following Hugging Face links:

1.  **NLPAD Datasets**: These are the datasets mentioned in the NLP-ADBench paper. You can download them from:
    -   [NLP-AD Datasets](https://huggingface.co/datasets/kendx/NLP-ADBench/tree/main/datasets)

2.  **Pre-Extracted Embeddings**: For embedding-based algorithms, pre-extracted embeddings are provided. If you want to use them directly, you can download them from:
    -   [Pre-Extracted Embeddings](https://huggingface.co/datasets/kendx/NLP-ADBench/tree/main/embeddings)

## Sample Usage

To run the benchmark, first set up the environment and import the pre-extracted embeddings:

### Environment Setup Instructions

1.  **Install Anaconda or Miniconda**:
    Download and install Anaconda or Miniconda from [here](https://docs.conda.io/en/latest/miniconda.html).

2.  **Create the Environment**:
    Using the terminal, navigate to the directory containing the `environment.yml` file in the GitHub repository and run:
    ```bash
    conda env create -f environment.yml
    ```
3.  **Activate the Environment**:
    Activate the newly created environment using:
    ```bash
    conda activate nlpad
    ```

### Import data

Get `Pre-Extracted Embeddings` data from the [Hugging Face link](https://huggingface.co/datasets/kendx/NLP-ADBench/tree/main/embeddings) and put it in the `data` folder.

Place all downloaded embeddings data into the `feature` folder in the `./benchmark` directory of this project.

### Run the code
Run the following commands from the `./benchmark` directory of the project:

#### BERT
If you want to run a benchmark using data embedded with BERT's `bert-base-uncased` model, use this command:
````bash
python [algorithm_name]_benchmark.py bert
````

#### OpenAI
If you want to run a benchmark using data embedded with OpenAI's `text-embedding-3-large` model, use this command:
````bash
python [algorithm_name]_benchmark.py gpt
````

## Citation

If you find this work useful, please cite our paper:

```bibtex
@article{li2025nlp,
  title={Nlp-adbench: Nlp anomaly detection benchmark},
  author={Li, Yuangang and Li, Jiaqi and Xiao, Zhuo and Yang, Tiankai and Nian, Yi and Hu, Xiyang and Zhao, Yue},
  journal={Findings of the Association for Computational Linguistics: EMNLP 2025},
  year={2025}
}
```