Bias in Intent Detection: A Dynamical Systems Perspective

Eduardo Sanchez-Karhunen, Jose F. Quesada-Moreno, Miguel A. Gutiérrez-Naranjo
Proceedings of Fourth European Workshop on Algorithmic Fairness, PMLR 294:396-402, 2025.

Abstract

Intent detection is a critical task in natural language processing (NLP), powering applications such as chatbots and dialogue systems. Although deep learning models have greatly improved intent classification, their internal mechanisms remain poorly understood, raising concerns about transparency and fairness. Recent studies have applied dynamical systems theory to analyze RNNs by examining their internal state dynamics. We propose a novel bias evaluation framework that examines sentence trajectories within the model’s state space. By analyzing the dynamic evolution of hidden states and their final alignment with decision-making layers, we identify key mechanisms for defining new bias metrics: trajectory sparsity, final state clustering, and readout vector alignment. This interpretable framework offers a principled approach to diagnosing and mitigating bias.

Cite this Paper

BibTeX
@InProceedings{pmlr-v294-eduardo-sanchez-karhunen25a, title = {Bias in Intent Detection: A Dynamical Systems Perspective}, author = {Eduardo Sanchez-Karhunen, Jose F. Quesada-Moreno, Miguel A. Guti\'errez-Naranjo}, booktitle = {Proceedings of Fourth European Workshop on Algorithmic Fairness}, pages = {396--402}, year = {2025}, editor = {Weerts, Hilde and Pechenizkiy, Mykola and Allhutter, Doris and Corrêa, Ana Maria and Grote, Thomas and Liem, Cynthia}, volume = {294}, series = {Proceedings of Machine Learning Research}, month = {30 Jun--02 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v294/main/assets/eduardo-sanchez-karhunen25a/eduardo-sanchez-karhunen25a.pdf}, url = {https://proceedings.mlr.press/v294/eduardo-sanchez-karhunen25a.html}, abstract = {Intent detection is a critical task in natural language processing (NLP), powering applications such as chatbots and dialogue systems. Although deep learning models have greatly improved intent classification, their internal mechanisms remain poorly understood, raising concerns about transparency and fairness. Recent studies have applied dynamical systems theory to analyze RNNs by examining their internal state dynamics. We propose a novel bias evaluation framework that examines sentence trajectories within the model’s state space. By analyzing the dynamic evolution of hidden states and their final alignment with decision-making layers, we identify key mechanisms for defining new bias metrics: trajectory sparsity, final state clustering, and readout vector alignment. This interpretable framework offers a principled approach to diagnosing and mitigating bias.} }
Endnote
%0 Conference Paper %T Bias in Intent Detection: A Dynamical Systems Perspective %A Eduardo Sanchez-Karhunen %A Jose F. Quesada-Moreno %A Miguel A. Gutiérrez-Naranjo %B Proceedings of Fourth European Workshop on Algorithmic Fairness %C Proceedings of Machine Learning Research %D 2025 %E Hilde Weerts %E Mykola Pechenizkiy %E Doris Allhutter %E Ana Maria Corrêa %E Thomas Grote %E Cynthia Liem %F pmlr-v294-eduardo-sanchez-karhunen25a %I PMLR %P 396--402 %U https://proceedings.mlr.press/v294/eduardo-sanchez-karhunen25a.html %V 294 %X Intent detection is a critical task in natural language processing (NLP), powering applications such as chatbots and dialogue systems. Although deep learning models have greatly improved intent classification, their internal mechanisms remain poorly understood, raising concerns about transparency and fairness. Recent studies have applied dynamical systems theory to analyze RNNs by examining their internal state dynamics. We propose a novel bias evaluation framework that examines sentence trajectories within the model’s state space. By analyzing the dynamic evolution of hidden states and their final alignment with decision-making layers, we identify key mechanisms for defining new bias metrics: trajectory sparsity, final state clustering, and readout vector alignment. This interpretable framework offers a principled approach to diagnosing and mitigating bias.
APA
Sanchez-Karhunen, E., Quesada-Moreno, J.F. & Gutiérrez-Naranjo, M.A.. (2025). Bias in Intent Detection: A Dynamical Systems Perspective. Proceedings of Fourth European Workshop on Algorithmic Fairness, in Proceedings of Machine Learning Research 294:396-402 Available from https://proceedings.mlr.press/v294/eduardo-sanchez-karhunen25a.html.

Related Material