@inproceedings{chatzikyriakidis-etal-2026-heptatax,
title = "{H}epta{TAX}: A Neuro-Symbolic Pipeline and Benchmark for Classifying 16th-Century Heptanesian Notarial Acts",
author = "Chatzikyriakidis, Stergios and
Karantzola, Eleni and
Makri, Vasiliki",
editor = "Anastasopoulos, Antonis and
Markantonatou, Stella and
Ralli, Angela and
Zampieri, Marcos and
Bompolas, Stavros and
Stamou, Vivian",
booktitle = "Proceedings of the First Workshop on Dialects in {NLP} {---} A Resource Perspective",
month = may,
year = "2026",
address = "Palma de Mallorca",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.dialres-1.26/",
doi = "10.63317/2sh2japrnd73",
pages = "265--273",
abstract = "This study originates in the investigation of lexical bundles and formulaic language within sixteenth-century Corfiot notarial documents. The observed functional variation across identical formulaic sequences motivated the development of a document classification framework designed to support the structural interpretation of such language. Given that 16th-century Corfiot notarial acts represent a rich, albeit understudied, dialectal resource, their systematic categorization into subgenres is essential for their full exploration. However, this task requires substantial manual work, while NLP tools for this task and dialect do not exist. In this paper, we attempt to take an initial step in this direction. First, we present a corpus of 1,088 notarial acts from 5 notaries spanning 1500-1567, a 3-tier annotation schema (17 core genres, extension subcategories, hybrid cross-cutting tags), and a 40-act benchmark with gold annotations at all three tiers. Then, we evaluate 12 LLMs across 4 architectures, zero-shot, few-shot, full-context and Neuro-Symbolic. For the latter, we introduce a symbolic engine comprising a set of deterministic rules for identifying discriminative legal formulae, whose output is then injected into the neural (LLM) engine. The results show that the NeSy architecture compresses the accuracy gap between stronger and weaker models from 47.5 pp to 12.5 pp, with the smallest model (Llama 3.1 8B) gaining 47.5{\%} and matching frontier models that operate without symbolic support. Three models reach a ceiling of 72.5{\%} on the core tier. However, consistent errors in procedurally dense material reveal the limits of lexical and formulaic cues for identifying legal effect, motivating the use of symbolic signals in the NeSy pipeline. Extension and hybrid classification remain open challenges, with best scores of {\ensuremath{\sim}}63{\%} and {\ensuremath{\sim}}35{\%} respectively."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="chatzikyriakidis-etal-2026-heptatax">
<titleInfo>
<title>HeptaTAX: A Neuro-Symbolic Pipeline and Benchmark for Classifying 16th-Century Heptanesian Notarial Acts</title>
</titleInfo>
<name type="personal">
<namePart type="given">Stergios</namePart>
<namePart type="family">Chatzikyriakidis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Eleni</namePart>
<namePart type="family">Karantzola</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vasiliki</namePart>
<namePart type="family">Makri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the First Workshop on Dialects in NLP — A Resource Perspective</title>
</titleInfo>
<name type="personal">
<namePart type="given">Antonis</namePart>
<namePart type="family">Anastasopoulos</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Stella</namePart>
<namePart type="family">Markantonatou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Angela</namePart>
<namePart type="family">Ralli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marcos</namePart>
<namePart type="family">Zampieri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Stavros</namePart>
<namePart type="family">Bompolas</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vivian</namePart>
<namePart type="family">Stamou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Palma de Mallorca</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>This study originates in the investigation of lexical bundles and formulaic language within sixteenth-century Corfiot notarial documents. The observed functional variation across identical formulaic sequences motivated the development of a document classification framework designed to support the structural interpretation of such language. Given that 16th-century Corfiot notarial acts represent a rich, albeit understudied, dialectal resource, their systematic categorization into subgenres is essential for their full exploration. However, this task requires substantial manual work, while NLP tools for this task and dialect do not exist. In this paper, we attempt to take an initial step in this direction. First, we present a corpus of 1,088 notarial acts from 5 notaries spanning 1500-1567, a 3-tier annotation schema (17 core genres, extension subcategories, hybrid cross-cutting tags), and a 40-act benchmark with gold annotations at all three tiers. Then, we evaluate 12 LLMs across 4 architectures, zero-shot, few-shot, full-context and Neuro-Symbolic. For the latter, we introduce a symbolic engine comprising a set of deterministic rules for identifying discriminative legal formulae, whose output is then injected into the neural (LLM) engine. The results show that the NeSy architecture compresses the accuracy gap between stronger and weaker models from 47.5 pp to 12.5 pp, with the smallest model (Llama 3.1 8B) gaining 47.5% and matching frontier models that operate without symbolic support. Three models reach a ceiling of 72.5% on the core tier. However, consistent errors in procedurally dense material reveal the limits of lexical and formulaic cues for identifying legal effect, motivating the use of symbolic signals in the NeSy pipeline. Extension and hybrid classification remain open challenges, with best scores of \ensuremath\sim63% and \ensuremath\sim35% respectively.</abstract>
<identifier type="citekey">chatzikyriakidis-etal-2026-heptatax</identifier>
<identifier type="doi">10.63317/2sh2japrnd73</identifier>
<location>
<url>https://aclanthology.org/2026.dialres-1.26/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>265</start>
<end>273</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T HeptaTAX: A Neuro-Symbolic Pipeline and Benchmark for Classifying 16th-Century Heptanesian Notarial Acts
%A Chatzikyriakidis, Stergios
%A Karantzola, Eleni
%A Makri, Vasiliki
%Y Anastasopoulos, Antonis
%Y Markantonatou, Stella
%Y Ralli, Angela
%Y Zampieri, Marcos
%Y Bompolas, Stavros
%Y Stamou, Vivian
%S Proceedings of the First Workshop on Dialects in NLP — A Resource Perspective
%D 2026
%8 May
%I Association for Computational Linguistics
%C Palma de Mallorca
%F chatzikyriakidis-etal-2026-heptatax
%X This study originates in the investigation of lexical bundles and formulaic language within sixteenth-century Corfiot notarial documents. The observed functional variation across identical formulaic sequences motivated the development of a document classification framework designed to support the structural interpretation of such language. Given that 16th-century Corfiot notarial acts represent a rich, albeit understudied, dialectal resource, their systematic categorization into subgenres is essential for their full exploration. However, this task requires substantial manual work, while NLP tools for this task and dialect do not exist. In this paper, we attempt to take an initial step in this direction. First, we present a corpus of 1,088 notarial acts from 5 notaries spanning 1500-1567, a 3-tier annotation schema (17 core genres, extension subcategories, hybrid cross-cutting tags), and a 40-act benchmark with gold annotations at all three tiers. Then, we evaluate 12 LLMs across 4 architectures, zero-shot, few-shot, full-context and Neuro-Symbolic. For the latter, we introduce a symbolic engine comprising a set of deterministic rules for identifying discriminative legal formulae, whose output is then injected into the neural (LLM) engine. The results show that the NeSy architecture compresses the accuracy gap between stronger and weaker models from 47.5 pp to 12.5 pp, with the smallest model (Llama 3.1 8B) gaining 47.5% and matching frontier models that operate without symbolic support. Three models reach a ceiling of 72.5% on the core tier. However, consistent errors in procedurally dense material reveal the limits of lexical and formulaic cues for identifying legal effect, motivating the use of symbolic signals in the NeSy pipeline. Extension and hybrid classification remain open challenges, with best scores of \ensuremath\sim63% and \ensuremath\sim35% respectively.
%R 10.63317/2sh2japrnd73
%U https://aclanthology.org/2026.dialres-1.26/
%U https://doi.org/10.63317/2sh2japrnd73
%P 265-273
Markdown (Informal)
[HeptaTAX: A Neuro-Symbolic Pipeline and Benchmark for Classifying 16th-Century Heptanesian Notarial Acts](https://aclanthology.org/2026.dialres-1.26/) (Chatzikyriakidis et al., DialRes 2026)
ACL