@inproceedings{duan-etal-2026-vib,
title = "{VIB}-Simp: {IB}-Inspired Output-Logit Regularization for {G}erman Text Simplification",
author = "Duan, Mofan and
Yang, Yanyan and
Liu, Xiufeng",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.710/",
pages = "16483--16495",
ISBN = "979-8-89176-459-0",
abstract = "German text simplification requires reducing lexical and structural complexity while preserving meaning, yet standard fine-tuning optimizes reference imitation without directly controlling the predictive distribution where lexical choices are made. We propose VIB-Simp, a lightweight output-logit regularization framework inspired by the Variational Information Bottleneck (VIB) literature. Instead of estimating mutual information or adding latent variables, VIB-Simp regularizes the mean and variance of output logits using a closed-form Gaussian Kullback{--}Leibler (KL) surrogate, providing a lightweight way to stabilize lexical choices during fine-tuning. An optional unlikelihood-style lexical penalty further suppresses probability mass assigned to difficult lexical candidates.\par Experiments on TextComplexityDE and 20Min show consistent gains across LLaMA-3.2-3B, Qwen-2.5-3B, mBART, GerPT-2, and mGPT, with consistent SARI improvements. Additional baselines, length-controlled analysis, ablations, expert ratings, and logit diagnostics show that the gains are not explained by shortening or dictionary-only effects. The results support an IB-style view of simplification: preserve meaning-critical content while suppressing unnecessary lexical complexity."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="duan-etal-2026-vib">
<titleInfo>
<title>VIB-Simp: IB-Inspired Output-Logit Regularization for German Text Simplification</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mofan</namePart>
<namePart type="family">Duan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yanyan</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiufeng</namePart>
<namePart type="family">Liu</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>German text simplification requires reducing lexical and structural complexity while preserving meaning, yet standard fine-tuning optimizes reference imitation without directly controlling the predictive distribution where lexical choices are made. We propose VIB-Simp, a lightweight output-logit regularization framework inspired by the Variational Information Bottleneck (VIB) literature. Instead of estimating mutual information or adding latent variables, VIB-Simp regularizes the mean and variance of output logits using a closed-form Gaussian Kullback–Leibler (KL) surrogate, providing a lightweight way to stabilize lexical choices during fine-tuning. An optional unlikelihood-style lexical penalty further suppresses probability mass assigned to difficult lexical candidates.\par Experiments on TextComplexityDE and 20Min show consistent gains across LLaMA-3.2-3B, Qwen-2.5-3B, mBART, GerPT-2, and mGPT, with consistent SARI improvements. Additional baselines, length-controlled analysis, ablations, expert ratings, and logit diagnostics show that the gains are not explained by shortening or dictionary-only effects. The results support an IB-style view of simplification: preserve meaning-critical content while suppressing unnecessary lexical complexity.</abstract>
<identifier type="citekey">duan-etal-2026-vib</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.710/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>16483</start>
<end>16495</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T VIB-Simp: IB-Inspired Output-Logit Regularization for German Text Simplification
%A Duan, Mofan
%A Yang, Yanyan
%A Liu, Xiufeng
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F duan-etal-2026-vib
%X German text simplification requires reducing lexical and structural complexity while preserving meaning, yet standard fine-tuning optimizes reference imitation without directly controlling the predictive distribution where lexical choices are made. We propose VIB-Simp, a lightweight output-logit regularization framework inspired by the Variational Information Bottleneck (VIB) literature. Instead of estimating mutual information or adding latent variables, VIB-Simp regularizes the mean and variance of output logits using a closed-form Gaussian Kullback–Leibler (KL) surrogate, providing a lightweight way to stabilize lexical choices during fine-tuning. An optional unlikelihood-style lexical penalty further suppresses probability mass assigned to difficult lexical candidates.\par Experiments on TextComplexityDE and 20Min show consistent gains across LLaMA-3.2-3B, Qwen-2.5-3B, mBART, GerPT-2, and mGPT, with consistent SARI improvements. Additional baselines, length-controlled analysis, ablations, expert ratings, and logit diagnostics show that the gains are not explained by shortening or dictionary-only effects. The results support an IB-style view of simplification: preserve meaning-critical content while suppressing unnecessary lexical complexity.
%U https://aclanthology.org/2026.emnlp-main.710/
%P 16483-16495
Markdown (Informal)
[VIB-Simp: IB-Inspired Output-Logit Regularization for German Text Simplification](https://aclanthology.org/2026.emnlp-main.710/) (Duan et al., EMNLP 2026)
ACL