xaiembeddings 0.1.2

A Python package for explainable AI embeddings analysis with perturbation-based methods.

XAI Embeddings

A Python package for explainable AI (XAI) embeddings analysis, using perturbation based explanation methods for transformer-based models. Explanations show how tokens influence position of each other.

Installation

pip install xai-embeddings

Requirements for POS-PFI

• Spacy Model: path to spacy model e.g.: en_core_web_trf
• Part-of-speech Dictionary: dictionary with list of words for permutations.

Available Explainers

• AttentionExplainer: Uses attention weights to explain token importance
• LOOExplainer: Leave-one-out analysis
• POSPermutationExplainer: Part-of-speech based permutation analysis
• SubsetExplainer: Subset-based explanation generation

Models

• BERTModel: BERT bert-base-uncased with embedding extraction
• Model: Abstract base class for custom model implementations

Examples:

Sample usage was provided in examples.ipynb

License

This project is licensed under the MIT License.