@inproceedings{tian-yuan-2026-grip,
title = "{GRIP}: Gated Routing of Instance-level Preferences for Noisy Long-Tailed Text Classification",
author = "Tian, Ziqing and
Yuan, Bo",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1901/",
doi = "10.18653/2026.findings-emnlp.1901",
pages = "40508--40535",
ISBN = "979-8-89176-462-0",
abstract = "Noisy long-tailed text classification is particularly challenging because mislabeled examples and clean tail-class examples often exhibit similarly high losses, despite requiring fundamentally different treatments. Consequently, conventional sample-selection methods may mistakenly discard informative minority-class instances, while uniform multi-teacher supervision can propagate conflicting learning signals. We propose GRIP (Gated Routing of Instance-level Preferences), a dual-teacher dual-gate framework that performs instance-level selective knowledge distillation. A noise-specialized teacher provides robust supervision for unreliable samples, whereas a tail-specialized teacher strengthens learning for under-represented classes. Two lightweight and independent gates use sample-level reliability, uncertainty, teacher-disagreement, and class-rarity signals to determine whether each teacher should supervise a given instance. The gates are optimized with reward-shaped weighted binary cross-entropy objectives, and the student combines the standard classification loss with selectively activated distillation losses. Experiments on eight text-classification benchmarks covering real-world and synthetic label noise, three noise types, and varying imbalance factors show that GRIP achieves the best performance across all reported synthetic noisy long-tailed settings and substantial gains on real-world noisy datasets, including a 9.76-point accuracy improvement on TREC over the strongest baseline. Ablation and routing analyses further show that the two teachers and gates learn complementary, non-redundant roles. Only the student is retained at inference time, preserving the model size and latency of a standard BERT classifier."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tian-yuan-2026-grip">
<titleInfo>
<title>GRIP: Gated Routing of Instance-level Preferences for Noisy Long-Tailed Text Classification</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ziqing</namePart>
<namePart type="family">Tian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bo</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Noisy long-tailed text classification is particularly challenging because mislabeled examples and clean tail-class examples often exhibit similarly high losses, despite requiring fundamentally different treatments. Consequently, conventional sample-selection methods may mistakenly discard informative minority-class instances, while uniform multi-teacher supervision can propagate conflicting learning signals. We propose GRIP (Gated Routing of Instance-level Preferences), a dual-teacher dual-gate framework that performs instance-level selective knowledge distillation. A noise-specialized teacher provides robust supervision for unreliable samples, whereas a tail-specialized teacher strengthens learning for under-represented classes. Two lightweight and independent gates use sample-level reliability, uncertainty, teacher-disagreement, and class-rarity signals to determine whether each teacher should supervise a given instance. The gates are optimized with reward-shaped weighted binary cross-entropy objectives, and the student combines the standard classification loss with selectively activated distillation losses. Experiments on eight text-classification benchmarks covering real-world and synthetic label noise, three noise types, and varying imbalance factors show that GRIP achieves the best performance across all reported synthetic noisy long-tailed settings and substantial gains on real-world noisy datasets, including a 9.76-point accuracy improvement on TREC over the strongest baseline. Ablation and routing analyses further show that the two teachers and gates learn complementary, non-redundant roles. Only the student is retained at inference time, preserving the model size and latency of a standard BERT classifier.</abstract>
<identifier type="citekey">tian-yuan-2026-grip</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1901</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1901/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>40508</start>
<end>40535</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T GRIP: Gated Routing of Instance-level Preferences for Noisy Long-Tailed Text Classification
%A Tian, Ziqing
%A Yuan, Bo
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F tian-yuan-2026-grip
%X Noisy long-tailed text classification is particularly challenging because mislabeled examples and clean tail-class examples often exhibit similarly high losses, despite requiring fundamentally different treatments. Consequently, conventional sample-selection methods may mistakenly discard informative minority-class instances, while uniform multi-teacher supervision can propagate conflicting learning signals. We propose GRIP (Gated Routing of Instance-level Preferences), a dual-teacher dual-gate framework that performs instance-level selective knowledge distillation. A noise-specialized teacher provides robust supervision for unreliable samples, whereas a tail-specialized teacher strengthens learning for under-represented classes. Two lightweight and independent gates use sample-level reliability, uncertainty, teacher-disagreement, and class-rarity signals to determine whether each teacher should supervise a given instance. The gates are optimized with reward-shaped weighted binary cross-entropy objectives, and the student combines the standard classification loss with selectively activated distillation losses. Experiments on eight text-classification benchmarks covering real-world and synthetic label noise, three noise types, and varying imbalance factors show that GRIP achieves the best performance across all reported synthetic noisy long-tailed settings and substantial gains on real-world noisy datasets, including a 9.76-point accuracy improvement on TREC over the strongest baseline. Ablation and routing analyses further show that the two teachers and gates learn complementary, non-redundant roles. Only the student is retained at inference time, preserving the model size and latency of a standard BERT classifier.
%R 10.18653/2026.findings-emnlp.1901
%U https://aclanthology.org/2026.findings-emnlp.1901/
%U https://doi.org/10.18653/2026.findings-emnlp.1901
%P 40508-40535
Markdown (Informal)
[GRIP: Gated Routing of Instance-level Preferences for Noisy Long-Tailed Text Classification](https://aclanthology.org/2026.findings-emnlp.1901/) (Tian & Yuan, Findings 2026)
ACL