@inproceedings{ouyang-etal-2026-definition,
title = "Definition-Only {RAG} for Financial {QA}: Average Gains and a Diagnostic Boundary",
author = "Ouyang, Leyi and
Marrese-Taylor, Edison and
Matsuo, Yutaka",
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.2448/",
pages = "52551--52560",
ISBN = "979-8-89176-462-0",
abstract = "Previous work shows that a model{'}s ability to define domain terms correlates to performance in other tasks related to that domain. For example, in a QA setting, a language model can miss a question in a specialized domain for no reason other than not properly recalling what a term in the question means. We argue that a short glossary definition is the smallest unit of evidence a retrieval system can supply against that failure. Concretely, focusing on the finance NLP domain, we assess the role of definitions in improving performance on XFinBench, and also pairing MMLU-Finance with a 13,609-term Investopedia glossary of financial terms. We experiment with a five-model panel spanning mid-capability open models to a frontier closed model, and show that definition-only retrieval produces a structured gain on these datasets, with the effect concentrating in a mid-capability band rather than scaling with model size. Finally, we connect term-level measurement to individual questions through a human-validated annotation, and find that term knowledge does not locate the questions where a definition changes the answer, with none our implemented internal routing signals approaching the per-question oracle."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ouyang-etal-2026-definition">
<titleInfo>
<title>Definition-Only RAG for Financial QA: Average Gains and a Diagnostic Boundary</title>
</titleInfo>
<name type="personal">
<namePart type="given">Leyi</namePart>
<namePart type="family">Ouyang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Edison</namePart>
<namePart type="family">Marrese-Taylor</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yutaka</namePart>
<namePart type="family">Matsuo</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>Previous work shows that a model’s ability to define domain terms correlates to performance in other tasks related to that domain. For example, in a QA setting, a language model can miss a question in a specialized domain for no reason other than not properly recalling what a term in the question means. We argue that a short glossary definition is the smallest unit of evidence a retrieval system can supply against that failure. Concretely, focusing on the finance NLP domain, we assess the role of definitions in improving performance on XFinBench, and also pairing MMLU-Finance with a 13,609-term Investopedia glossary of financial terms. We experiment with a five-model panel spanning mid-capability open models to a frontier closed model, and show that definition-only retrieval produces a structured gain on these datasets, with the effect concentrating in a mid-capability band rather than scaling with model size. Finally, we connect term-level measurement to individual questions through a human-validated annotation, and find that term knowledge does not locate the questions where a definition changes the answer, with none our implemented internal routing signals approaching the per-question oracle.</abstract>
<identifier type="citekey">ouyang-etal-2026-definition</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2448/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>52551</start>
<end>52560</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Definition-Only RAG for Financial QA: Average Gains and a Diagnostic Boundary
%A Ouyang, Leyi
%A Marrese-Taylor, Edison
%A Matsuo, Yutaka
%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 ouyang-etal-2026-definition
%X Previous work shows that a model’s ability to define domain terms correlates to performance in other tasks related to that domain. For example, in a QA setting, a language model can miss a question in a specialized domain for no reason other than not properly recalling what a term in the question means. We argue that a short glossary definition is the smallest unit of evidence a retrieval system can supply against that failure. Concretely, focusing on the finance NLP domain, we assess the role of definitions in improving performance on XFinBench, and also pairing MMLU-Finance with a 13,609-term Investopedia glossary of financial terms. We experiment with a five-model panel spanning mid-capability open models to a frontier closed model, and show that definition-only retrieval produces a structured gain on these datasets, with the effect concentrating in a mid-capability band rather than scaling with model size. Finally, we connect term-level measurement to individual questions through a human-validated annotation, and find that term knowledge does not locate the questions where a definition changes the answer, with none our implemented internal routing signals approaching the per-question oracle.
%U https://aclanthology.org/2026.findings-emnlp.2448/
%P 52551-52560
Markdown (Informal)
[Definition-Only RAG for Financial QA: Average Gains and a Diagnostic Boundary](https://aclanthology.org/2026.findings-emnlp.2448/) (Ouyang et al., Findings 2026)
ACL