<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="qiu-etal-2026-learning">
    <titleInfo>
        <title>Learning to Recover Bridge Path for SQL Schema Linking via Tentative Answer Profiling</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Wenzhong</namePart>
        <namePart type="family">Qiu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Shixin</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yu</namePart>
        <namePart type="family">Hong</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>Schema linking is a task of investigating relevant SQL schema (items and paths) for a specific query language. Constructing a reliable linker is crucial for Text2SQL auto-coding since it provides informative clues for directing SQL coders. Current studies have obtained considerable performance gains when independently evaluating linkers, which allow F1-score to be higher than 90% at best. However, our observation illustrates that the existing linkers still suffers from a high missing rate of Bridge Path, i.e., omitting one or a series of SQL keywords that link across SQL tables. To address the issue, we propose to enhance linkers by Bridge Path Recovering (abbr., BPR) via Tentative Answer Profiling (TAP). On the one hand, BPR is implemented by learning to rank up the better bridged schemata during finetuning. On the other hand, TAP tentatively conjectures possible answers for the query language, and generates their profiles (types and CoT steps) to guide BPR. This is inspired by the experience that answer conjecturing contributes to reversely deducing the multi-hop routing path over SQL tables. Technologically, we train LLaMA-3.1 for TAP, where teacher-student distillation is conducted via Deepseek-3. To ensure sufficient ranking data used for training BPR, we employ a multilevel retrieval technique to align profiles with diverse schemata. We experiment on the benchmark datasets BIRD and Spider. The experimental results show that our approach (BPR-TAP) consistently yields improvements for F1-score upon all benchmarks, and surpasses multiple competitive schema linkers. All source codes are available at https://anonymous.4open.science/r/BPR-TAP-Text2SQL.</abstract>
    <identifier type="citekey">qiu-etal-2026-learning</identifier>
    <identifier type="doi">10.18653/2026.findings-emnlp.965</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.965/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>20413</start>
            <end>20429</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Learning to Recover Bridge Path for SQL Schema Linking via Tentative Answer Profiling
%A Qiu, Wenzhong
%A Liu, Shixin
%A Hong, Yu
%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 qiu-etal-2026-learning
%X Schema linking is a task of investigating relevant SQL schema (items and paths) for a specific query language. Constructing a reliable linker is crucial for Text2SQL auto-coding since it provides informative clues for directing SQL coders. Current studies have obtained considerable performance gains when independently evaluating linkers, which allow F1-score to be higher than 90% at best. However, our observation illustrates that the existing linkers still suffers from a high missing rate of Bridge Path, i.e., omitting one or a series of SQL keywords that link across SQL tables. To address the issue, we propose to enhance linkers by Bridge Path Recovering (abbr., BPR) via Tentative Answer Profiling (TAP). On the one hand, BPR is implemented by learning to rank up the better bridged schemata during finetuning. On the other hand, TAP tentatively conjectures possible answers for the query language, and generates their profiles (types and CoT steps) to guide BPR. This is inspired by the experience that answer conjecturing contributes to reversely deducing the multi-hop routing path over SQL tables. Technologically, we train LLaMA-3.1 for TAP, where teacher-student distillation is conducted via Deepseek-3. To ensure sufficient ranking data used for training BPR, we employ a multilevel retrieval technique to align profiles with diverse schemata. We experiment on the benchmark datasets BIRD and Spider. The experimental results show that our approach (BPR-TAP) consistently yields improvements for F1-score upon all benchmarks, and surpasses multiple competitive schema linkers. All source codes are available at https://anonymous.4open.science/r/BPR-TAP-Text2SQL.
%R 10.18653/2026.findings-emnlp.965
%U https://aclanthology.org/2026.findings-emnlp.965/
%U https://doi.org/10.18653/2026.findings-emnlp.965
%P 20413-20429

Markdown (Informal)

[Learning to Recover Bridge Path for SQL Schema Linking via Tentative Answer Profiling](https://aclanthology.org/2026.findings-emnlp.965/) (Qiu et al., Findings 2026)

• Learning to Recover Bridge Path for SQL Schema Linking via Tentative Answer Profiling (Qiu et al., Findings 2026)

ACL

• Wenzhong Qiu, Shixin Liu, and Yu Hong. 2026. Learning to Recover Bridge Path for SQL Schema Linking via Tentative Answer Profiling. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 20413–20429, Budapest, Hungary. Association for Computational Linguistics.