Core Concepts
PE introduces a novel method using hyperbolic spaces to model feature interactions efficiently, demonstrating effectiveness in generating hierarchical explanations.
Abstract
Abstract:
Deep learning models in NLP lack interpretability.
Hierarchical Attribution (HA) is crucial for modeling feature interactions.
PE introduces a method using hyperbolic spaces for efficient feature interaction modeling.
Introduction:
Interpretability studies have emerged due to the opaqueness of deep learning models.
HA categorizes words into clusters to build a hierarchical tree.
Methodology:
PE projects word embeddings into hyperbolic spaces and estimates contributions using game theory.
The algorithm conceptualizes clustering as a super additivity game.
Experiments:
PE outperforms baselines in text classification datasets.
PE demonstrates computational efficiency compared to other methods.
Conclusion:
PE effectively models feature interactions using hyperbolic geometry.
Stats
Inspired by Poincaré model, we propose a framework to project the embeddings into hyperbolic spaces, which exhibit better inductive biases for syntax and semantic hierarchical structures.
We evaluate the proposed method on three classification datasets with BERT, and the results demonstrate effectiveness.
Our code is available at https://github.com/qq31415926/PE.