@inproceedings{lyu-etal-2026-kode,
title = "{KODE}: Mining Implicit Domain Knowledge from Historical Experience for Text-to-{SQL}",
author = "Lyu, Qingshuo and
Yin, Jiangyi and
Li, Zhao and
Du, Meijie and
Jiang, Zhuojun and
Liu, Qingyun",
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.2032/",
pages = "43245--43266",
ISBN = "979-8-89176-462-0",
abstract = "Text-to-SQL systems built on large language models still struggle in real deployment, where users underspecify queries while databases enforce rigid schemas. The gap is rarely lexical; it lies in implicit knowledge{---}hidden joins, derived formulas, and domain-specific filters{---}that natural language (NL) omits but SQL must encode. We observe that this knowledge is not missing: every historical (NL, SQL) pair already records it, with NL stating the goal and SQL stating the steps. Treating the two as dual views of one program, we propose KODE-SQL, a framework with three agents: KODE-Miner aligns NL and SQL in a shared graph space and distills the unmatched regions into reusable templates; KODE-Solver retrieves these templates to generate schema-grounded SQL under a rule-then-retrieval policy; KODE-Critic verifies each candidate on the same graph and feeds successful cases back to the knowledge base. KODE-SQL improves execution accuracy by up to 11.1 points on the knowledge-heavy ScienceBenchmark, remains competitive on BIRD and Spider, and yields consistent gains (+4{--}7 points) when plugged into existing solvers{---}validating both its standalone strength and its portability as a knowledge module."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lyu-etal-2026-kode">
<titleInfo>
<title>KODE: Mining Implicit Domain Knowledge from Historical Experience for Text-to-SQL</title>
</titleInfo>
<name type="personal">
<namePart type="given">Qingshuo</namePart>
<namePart type="family">Lyu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiangyi</namePart>
<namePart type="family">Yin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhao</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Meijie</namePart>
<namePart type="family">Du</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhuojun</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qingyun</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>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>Text-to-SQL systems built on large language models still struggle in real deployment, where users underspecify queries while databases enforce rigid schemas. The gap is rarely lexical; it lies in implicit knowledge—hidden joins, derived formulas, and domain-specific filters—that natural language (NL) omits but SQL must encode. We observe that this knowledge is not missing: every historical (NL, SQL) pair already records it, with NL stating the goal and SQL stating the steps. Treating the two as dual views of one program, we propose KODE-SQL, a framework with three agents: KODE-Miner aligns NL and SQL in a shared graph space and distills the unmatched regions into reusable templates; KODE-Solver retrieves these templates to generate schema-grounded SQL under a rule-then-retrieval policy; KODE-Critic verifies each candidate on the same graph and feeds successful cases back to the knowledge base. KODE-SQL improves execution accuracy by up to 11.1 points on the knowledge-heavy ScienceBenchmark, remains competitive on BIRD and Spider, and yields consistent gains (+4–7 points) when plugged into existing solvers—validating both its standalone strength and its portability as a knowledge module.</abstract>
<identifier type="citekey">lyu-etal-2026-kode</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2032/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>43245</start>
<end>43266</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T KODE: Mining Implicit Domain Knowledge from Historical Experience for Text-to-SQL
%A Lyu, Qingshuo
%A Yin, Jiangyi
%A Li, Zhao
%A Du, Meijie
%A Jiang, Zhuojun
%A Liu, Qingyun
%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 lyu-etal-2026-kode
%X Text-to-SQL systems built on large language models still struggle in real deployment, where users underspecify queries while databases enforce rigid schemas. The gap is rarely lexical; it lies in implicit knowledge—hidden joins, derived formulas, and domain-specific filters—that natural language (NL) omits but SQL must encode. We observe that this knowledge is not missing: every historical (NL, SQL) pair already records it, with NL stating the goal and SQL stating the steps. Treating the two as dual views of one program, we propose KODE-SQL, a framework with three agents: KODE-Miner aligns NL and SQL in a shared graph space and distills the unmatched regions into reusable templates; KODE-Solver retrieves these templates to generate schema-grounded SQL under a rule-then-retrieval policy; KODE-Critic verifies each candidate on the same graph and feeds successful cases back to the knowledge base. KODE-SQL improves execution accuracy by up to 11.1 points on the knowledge-heavy ScienceBenchmark, remains competitive on BIRD and Spider, and yields consistent gains (+4–7 points) when plugged into existing solvers—validating both its standalone strength and its portability as a knowledge module.
%U https://aclanthology.org/2026.findings-emnlp.2032/
%P 43245-43266
Markdown (Informal)
[KODE: Mining Implicit Domain Knowledge from Historical Experience for Text-to-SQL](https://aclanthology.org/2026.findings-emnlp.2032/) (Lyu et al., Findings 2026)
ACL