Selective Capability Unlearning in End-to-End Spoken Language Understanding

2026-06-22

AI Summary

This work addresses the problem of "capability persistence" in end-to-end spoken language understanding (SLU) models, where slot-filling behavior can be unexpectedly recovered via external intent prefixes even after the corresponding intent has been removed, indicating that the model fails to truly forget deleted functionality. To tackle this issue, we formally define capability persistence and propose the Binding Subspace Unlearning (BSU) framework. BSU analyzes the representation space of autoregressive SLU models to identify implicit binding directions between intents and slots, then employs subspace projection and directional attenuation to selectively disentangle these bindings. Experiments across multiple SLU benchmarks demonstrate that BSU substantially reduces forced prefix recovery rates while preserving the performance of retained functionalities, enabling precise and controllable capability unlearning.

Abstract

Modern spoken language understanding (SLU) systems are increasingly deployed in real-world settings, where specific functionalities may need to be removed due to policy or safety constraints. In SLU, a functionality corresponds to an intent and its associated slot-generation behavior. However, in autoregressive models, suppressing a target intent does not eliminate the conditional mapping that generates slots conditioned on that intent. When the intent prefix is externally supplied, the model can reconstruct the original intent-slot structure. We identify this structural failure as \textbf{\emph{capability persistence}}. We propose \textit{\underline{B}inding \underline{S}ubspace (BSU)}, a representation-level framework that isolates and attenuates intent-conditioned directions underlying this mapping. Across SLU benchmarks, BSU substantially reduces forced-prefix recoverability while preserving retained performance.

Problem

• spoken language understanding
• capability unlearning
• intent-slot structure
• capability persistence
• autoregressive models

Innovation

• capability unlearning
• spoken language understanding
• binding subspace
• intent-slot mapping
• representation-level attenuation