README.md

metadata

language:
  - en
license: mit
size_categories:
  - 1K<n<10K
task_categories:
  - text-classification
pretty_name: MCA^2 Data & Embeddings
tags:
  - anomaly-detection
  - multi-view
  - embeddings
  - representation-learning
  - contrastive-learning
arxiv: 2601.17786
dataset_info:
  features:
    - name: data
      dtype: file
    - name: embeddings
      dtype: file

MCA^2 Data & Embeddings

Paper | GitHub

This repository provides the raw data (data/) and the corresponding precomputed multi-view embeddings (embeddings/) for MCA^2, a two-stage multi-view text anomaly detection (TAD) framework.

MCA^2 exploits embeddings from multiple pretrained language models (views) and integrates them via a multi-view reconstruction model, contrastive collaboration, and adaptive allocation to identify anomalies. This dataset release facilitates reproduction by providing pre-extracted vectors, avoiding the need for expensive re-computation across various encoders (e.g., BERT, Stella, Qwen, and OpenAI).

Content

• data/: Dataset files including train/test splits (e.g., .npz and .jsonl files).
• embeddings/: Pre-extracted vectors grouped by dataset and split. Multiple embedding files correspond to different "views" or encoders.

Sample Usage

To reproduce the results for a specific dataset (such as OLID) using the MCA^2 framework, you can follow the instructions from the official repository:

# 1. Setup environment
conda create -n MCA2 python=3.9
conda activate MCA2
pip install torch sentence-transformers numpy transformers scikit-learn pandas tqdm pyod accelerate

# 2. Clone the repository and navigate to the evaluation directory
git clone https://github.com/yankehan/MCA2
cd MCA2/multiview_two_stage/eval

# 3. Run the evaluation script (ensure data and embeddings are placed in the project directory)
python ourmethod_eval.py --dataset olid --seeds 41,42,43,44,45

Notes

• Embeddings can be large; it is recommended to start with a smaller dataset like TAD-OLID first.
• If downloads are slow, you may try using a Hugging Face mirror (e.g., https://hf-mirror.com).

Citation

If you use this dataset or the MCA^2 framework in your research, please cite:

@article{liu2026beyond,
  title={Beyond a Single Perspective: Text Anomaly Detection with Multi-View Language Representations},
  author={Yixin Liu, Kehan Yan, Shiyuan Li and others},
  journal={arXiv preprint arXiv:2601.17786},
  year={2026}
}

License

This dataset is released under the MIT License.