@inproceedings{kanner-etal-2026-towards,
title = "Towards More Reliable Web Metadata: Do Register Classifiers Learn Register or Genre Features?",
author = "Kanner, Antti and
Lundberg, Tuomas and
Henriksson, Erik and
Laippala, Veronika",
editor = "Ortiz Suarez, Pedro and
Burchell, Laurie and
Ljube{\v{s}}i{\'c}, Nikola and
Scherrer, Yves and
Laippala, Veronika and
Dini{\'c}, Vuk and
Vaughan, Thom",
booktitle = "Proceedings of the 13th Web-as-Corpus Workshop",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.wac-1.14/",
pages = "147--160",
ISBN = "979-8-89176-481-1",
abstract = "Transformer-based web register classifiers are widely used to provide metadata for web crawls and to focus web-data sampling on specific register classes, such as news articles or advertisements. However, it remains unclear what these classifiers actually learn: do they detect pervasive linguistic features characteristic of registers, or do they rely on structural cues more indicative of genre? We examine this question through two hypotheses: the \textit{text-macrostructure hypothesis} proposes that classifiers rely on structural features more typical of genres, whereas the \textit{feature-density hypothesis} predicts that they identify typically register-associated features. Using English, Finnish, and French data from HPLT 3.0, we segment documents into thirds and compare classifier predictions for full documents and segments through five complementary analyses: main-label agreement, cosine similarity, confusion profiles, SACX keyword-weight correlations, and SACX keyword stability. As a more direct control, we compare predictions for a sample of untruncated documents with predictions for the same documents after their sentence order has been randomised. Top-label assignments and prediction vectors remain highly similar after shuffling, although the positional gradients are attenuated, indicating that sentence order contributes to classifier behaviour without determining its broader prediction profile. Overall, the results are more consistent with the feature-density hypothesis, although both hypotheses retain explanatory value."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kanner-etal-2026-towards">
<titleInfo>
<title>Towards More Reliable Web Metadata: Do Register Classifiers Learn Register or Genre Features?</title>
</titleInfo>
<name type="personal">
<namePart type="given">Antti</namePart>
<namePart type="family">Kanner</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tuomas</namePart>
<namePart type="family">Lundberg</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Erik</namePart>
<namePart type="family">Henriksson</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Veronika</namePart>
<namePart type="family">Laippala</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 13th Web-as-Corpus Workshop</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pedro</namePart>
<namePart type="family">Ortiz Suarez</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Laurie</namePart>
<namePart type="family">Burchell</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nikola</namePart>
<namePart type="family">Ljubešić</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yves</namePart>
<namePart type="family">Scherrer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Veronika</namePart>
<namePart type="family">Laippala</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vuk</namePart>
<namePart type="family">Dinić</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thom</namePart>
<namePart type="family">Vaughan</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">979-8-89176-481-1</identifier>
</relatedItem>
<abstract>Transformer-based web register classifiers are widely used to provide metadata for web crawls and to focus web-data sampling on specific register classes, such as news articles or advertisements. However, it remains unclear what these classifiers actually learn: do they detect pervasive linguistic features characteristic of registers, or do they rely on structural cues more indicative of genre? We examine this question through two hypotheses: the text-macrostructure hypothesis proposes that classifiers rely on structural features more typical of genres, whereas the feature-density hypothesis predicts that they identify typically register-associated features. Using English, Finnish, and French data from HPLT 3.0, we segment documents into thirds and compare classifier predictions for full documents and segments through five complementary analyses: main-label agreement, cosine similarity, confusion profiles, SACX keyword-weight correlations, and SACX keyword stability. As a more direct control, we compare predictions for a sample of untruncated documents with predictions for the same documents after their sentence order has been randomised. Top-label assignments and prediction vectors remain highly similar after shuffling, although the positional gradients are attenuated, indicating that sentence order contributes to classifier behaviour without determining its broader prediction profile. Overall, the results are more consistent with the feature-density hypothesis, although both hypotheses retain explanatory value.</abstract>
<identifier type="citekey">kanner-etal-2026-towards</identifier>
<location>
<url>https://aclanthology.org/2026.wac-1.14/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>147</start>
<end>160</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Towards More Reliable Web Metadata: Do Register Classifiers Learn Register or Genre Features?
%A Kanner, Antti
%A Lundberg, Tuomas
%A Henriksson, Erik
%A Laippala, Veronika
%Y Ortiz Suarez, Pedro
%Y Burchell, Laurie
%Y Ljubešić, Nikola
%Y Scherrer, Yves
%Y Laippala, Veronika
%Y Dinić, Vuk
%Y Vaughan, Thom
%S Proceedings of the 13th Web-as-Corpus Workshop
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-481-1
%F kanner-etal-2026-towards
%X Transformer-based web register classifiers are widely used to provide metadata for web crawls and to focus web-data sampling on specific register classes, such as news articles or advertisements. However, it remains unclear what these classifiers actually learn: do they detect pervasive linguistic features characteristic of registers, or do they rely on structural cues more indicative of genre? We examine this question through two hypotheses: the text-macrostructure hypothesis proposes that classifiers rely on structural features more typical of genres, whereas the feature-density hypothesis predicts that they identify typically register-associated features. Using English, Finnish, and French data from HPLT 3.0, we segment documents into thirds and compare classifier predictions for full documents and segments through five complementary analyses: main-label agreement, cosine similarity, confusion profiles, SACX keyword-weight correlations, and SACX keyword stability. As a more direct control, we compare predictions for a sample of untruncated documents with predictions for the same documents after their sentence order has been randomised. Top-label assignments and prediction vectors remain highly similar after shuffling, although the positional gradients are attenuated, indicating that sentence order contributes to classifier behaviour without determining its broader prediction profile. Overall, the results are more consistent with the feature-density hypothesis, although both hypotheses retain explanatory value.
%U https://aclanthology.org/2026.wac-1.14/
%P 147-160
Markdown (Informal)
[Towards More Reliable Web Metadata: Do Register Classifiers Learn Register or Genre Features?](https://aclanthology.org/2026.wac-1.14/) (Kanner et al., WAC 2026)
ACL