@inproceedings{c-etal-2026-spycomet,
title = "{S}py{C}omet at {S}em{E}val-2026 Task 11: When Adversarial Debiasing Backfires - A Comparison of Data-Level and Model-Level Debiasing",
author = "C, Sai Aravind and
Saumya, Sunil and
Reddy, C Pothan",
editor = "Kochmar, Ekaterina and
Ghosh, Debanjan and
North, Kai and
Komachi, Mamoru",
booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
month = jul,
year = "2026",
address = "San Diego, California, USA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.semeval-1.317/",
pages = "2512--2519",
ISBN = "979-8-89176-414-9",
abstract = "We describe MLA-CI (Multi-Layer Adversarial for Content Invariance), a DeBERTa-v3-base system for SemEval-2026 Task 11 Subtask 1 on content-invariant syllogistic reasoning. MLA-CI combines multi-layer feature extraction, gradient-reversal adversarial training, structure-preserving template augmentation, implausible-class oversampling, and focal loss. Our principal contribution is a systematic ablation study, confirmed across three random seeds, showing that adversarial training at standard strength is counterproductive: removing gradient reversal improves the mean validation score from 26.41 {\ensuremath{\pm}} 0.99 to 38.15 {\ensuremath{\pm}} 5.32. Per-condition analysis reveals that gradient reversal over-suppresses plausibility-correlated features, creating an inverted content effect that disproportionately harms plausible-valid accuracy. A sweep over seven adversarial pressure values reveal that only very light adversarial pressure value ({\ensuremath{\leq}} 0.1) preserves accuracy, while the submitted adversarial pressure value (1.0 or above) cause severe degradation. We conclude that data-level debiasing through structure-preserving augmentation is more effective and robust than model-level adversarial debiasing for this task."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="c-etal-2026-spycomet">
<titleInfo>
<title>SpyComet at SemEval-2026 Task 11: When Adversarial Debiasing Backfires - A Comparison of Data-Level and Model-Level Debiasing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sai</namePart>
<namePart type="given">Aravind</namePart>
<namePart type="family">C</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunil</namePart>
<namePart type="family">Saumya</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">C</namePart>
<namePart type="given">Pothan</namePart>
<namePart type="family">Reddy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 20th International Workshop on Semantic Evaluation (2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ekaterina</namePart>
<namePart type="family">Kochmar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Debanjan</namePart>
<namePart type="family">Ghosh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kai</namePart>
<namePart type="family">North</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mamoru</namePart>
<namePart type="family">Komachi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, California, USA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-414-9</identifier>
</relatedItem>
<abstract>We describe MLA-CI (Multi-Layer Adversarial for Content Invariance), a DeBERTa-v3-base system for SemEval-2026 Task 11 Subtask 1 on content-invariant syllogistic reasoning. MLA-CI combines multi-layer feature extraction, gradient-reversal adversarial training, structure-preserving template augmentation, implausible-class oversampling, and focal loss. Our principal contribution is a systematic ablation study, confirmed across three random seeds, showing that adversarial training at standard strength is counterproductive: removing gradient reversal improves the mean validation score from 26.41 \ensuremath\pm 0.99 to 38.15 \ensuremath\pm 5.32. Per-condition analysis reveals that gradient reversal over-suppresses plausibility-correlated features, creating an inverted content effect that disproportionately harms plausible-valid accuracy. A sweep over seven adversarial pressure values reveal that only very light adversarial pressure value (\ensuremathłeq 0.1) preserves accuracy, while the submitted adversarial pressure value (1.0 or above) cause severe degradation. We conclude that data-level debiasing through structure-preserving augmentation is more effective and robust than model-level adversarial debiasing for this task.</abstract>
<identifier type="citekey">c-etal-2026-spycomet</identifier>
<location>
<url>https://aclanthology.org/2026.semeval-1.317/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>2512</start>
<end>2519</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T SpyComet at SemEval-2026 Task 11: When Adversarial Debiasing Backfires - A Comparison of Data-Level and Model-Level Debiasing
%A C, Sai Aravind
%A Saumya, Sunil
%A Reddy, C. Pothan
%Y Kochmar, Ekaterina
%Y Ghosh, Debanjan
%Y North, Kai
%Y Komachi, Mamoru
%S Proceedings of the 20th International Workshop on Semantic Evaluation (2026)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, USA
%@ 979-8-89176-414-9
%F c-etal-2026-spycomet
%X We describe MLA-CI (Multi-Layer Adversarial for Content Invariance), a DeBERTa-v3-base system for SemEval-2026 Task 11 Subtask 1 on content-invariant syllogistic reasoning. MLA-CI combines multi-layer feature extraction, gradient-reversal adversarial training, structure-preserving template augmentation, implausible-class oversampling, and focal loss. Our principal contribution is a systematic ablation study, confirmed across three random seeds, showing that adversarial training at standard strength is counterproductive: removing gradient reversal improves the mean validation score from 26.41 \ensuremath\pm 0.99 to 38.15 \ensuremath\pm 5.32. Per-condition analysis reveals that gradient reversal over-suppresses plausibility-correlated features, creating an inverted content effect that disproportionately harms plausible-valid accuracy. A sweep over seven adversarial pressure values reveal that only very light adversarial pressure value (\ensuremathłeq 0.1) preserves accuracy, while the submitted adversarial pressure value (1.0 or above) cause severe degradation. We conclude that data-level debiasing through structure-preserving augmentation is more effective and robust than model-level adversarial debiasing for this task.
%U https://aclanthology.org/2026.semeval-1.317/
%P 2512-2519
Markdown (Informal)
[SpyComet at SemEval-2026 Task 11: When Adversarial Debiasing Backfires - A Comparison of Data-Level and Model-Level Debiasing](https://aclanthology.org/2026.semeval-1.317/) (C et al., SemEval 2026)
ACL