@inproceedings{jiang-etal-2026-small,
title = "Small Model, Big Retrieval: Unified Heterogeneous Retrieval via Multi-Label-Aware Contrastive Fine-Tuning",
author = "Jiang, Shaobai and
Liu, Dantong and
Gonzalez, Federico and
Herrero-Vidal, Pedro and
Chen, You-Lin and
Chu, Xiaoyu and
Sen, Prithviraj",
editor = "Atanasova, Pepa and
Haffari, Gholamreza and
Wang, Minghan",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-industry.140/",
pages = "2355--2368",
ISBN = "979-8-89176-460-6",
abstract = "Contrastive fine-tuning of a single embedding model across heterogeneous retrieval tasks is challenging when the tasks differ in label structure. In product-compliance retrieval, some tasks have multi-label supervision {---} every item carries a set of compliance categories, and many in-batch items share at least one category with the query, producing frequent in-batch false negatives. Other tasks (e.g., cross-catalog product matching) have no category labels: each query has one canonical match. Existing multi-label contrastive methods require every batch item to carry comparable labels, so they cannot serve a single embedding model trained on both label-bearing and label-free tasks. We propose label-aware false-negative exclusion: a zero-cost binary mask on the InfoNCE denominator that drops same-label candidates from the negative set rather than including them to positives. The mask has no effect on label-free items, so the same loss handles both task types. A 0.6 B-parameter embedding model fine-tuned this way handles four retrieval tasks simultaneously, lifting average Recall by 49{\%} (0.489{~}$\to${~}0.729) and matching the $13\times$ larger off-the-shelf Qwen3-Embedding-8B at $32\times$ lower embedding dimensionality. Fine-tuning further enables \textbf{learned query fusion}: concatenating regulatory text with an example product, which hurts retrieval before fine-tuning, becomes the strongest query type after, beating either alone by 6-11{\%}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="jiang-etal-2026-small">
<titleInfo>
<title>Small Model, Big Retrieval: Unified Heterogeneous Retrieval via Multi-Label-Aware Contrastive Fine-Tuning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shaobai</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dantong</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Federico</namePart>
<namePart type="family">Gonzalez</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pedro</namePart>
<namePart type="family">Herrero-Vidal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">You-Lin</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoyu</namePart>
<namePart type="family">Chu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Prithviraj</namePart>
<namePart type="family">Sen</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 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pepa</namePart>
<namePart type="family">Atanasova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gholamreza</namePart>
<namePart type="family">Haffari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minghan</namePart>
<namePart type="family">Wang</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">979-8-89176-460-6</identifier>
</relatedItem>
<abstract>Contrastive fine-tuning of a single embedding model across heterogeneous retrieval tasks is challenging when the tasks differ in label structure. In product-compliance retrieval, some tasks have multi-label supervision — every item carries a set of compliance categories, and many in-batch items share at least one category with the query, producing frequent in-batch false negatives. Other tasks (e.g., cross-catalog product matching) have no category labels: each query has one canonical match. Existing multi-label contrastive methods require every batch item to carry comparable labels, so they cannot serve a single embedding model trained on both label-bearing and label-free tasks. We propose label-aware false-negative exclusion: a zero-cost binary mask on the InfoNCE denominator that drops same-label candidates from the negative set rather than including them to positives. The mask has no effect on label-free items, so the same loss handles both task types. A 0.6 B-parameter embedding model fine-tuned this way handles four retrieval tasks simultaneously, lifting average Recall by 49% (0.489 0.729) and matching the 13\times larger off-the-shelf Qwen3-Embedding-8B at 32\times lower embedding dimensionality. Fine-tuning further enables learned query fusion: concatenating regulatory text with an example product, which hurts retrieval before fine-tuning, becomes the strongest query type after, beating either alone by 6-11%.</abstract>
<identifier type="citekey">jiang-etal-2026-small</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-industry.140/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>2355</start>
<end>2368</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Small Model, Big Retrieval: Unified Heterogeneous Retrieval via Multi-Label-Aware Contrastive Fine-Tuning
%A Jiang, Shaobai
%A Liu, Dantong
%A Gonzalez, Federico
%A Herrero-Vidal, Pedro
%A Chen, You-Lin
%A Chu, Xiaoyu
%A Sen, Prithviraj
%Y Atanasova, Pepa
%Y Haffari, Gholamreza
%Y Wang, Minghan
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-460-6
%F jiang-etal-2026-small
%X Contrastive fine-tuning of a single embedding model across heterogeneous retrieval tasks is challenging when the tasks differ in label structure. In product-compliance retrieval, some tasks have multi-label supervision — every item carries a set of compliance categories, and many in-batch items share at least one category with the query, producing frequent in-batch false negatives. Other tasks (e.g., cross-catalog product matching) have no category labels: each query has one canonical match. Existing multi-label contrastive methods require every batch item to carry comparable labels, so they cannot serve a single embedding model trained on both label-bearing and label-free tasks. We propose label-aware false-negative exclusion: a zero-cost binary mask on the InfoNCE denominator that drops same-label candidates from the negative set rather than including them to positives. The mask has no effect on label-free items, so the same loss handles both task types. A 0.6 B-parameter embedding model fine-tuned this way handles four retrieval tasks simultaneously, lifting average Recall by 49% (0.489 0.729) and matching the 13\times larger off-the-shelf Qwen3-Embedding-8B at 32\times lower embedding dimensionality. Fine-tuning further enables learned query fusion: concatenating regulatory text with an example product, which hurts retrieval before fine-tuning, becomes the strongest query type after, beating either alone by 6-11%.
%U https://aclanthology.org/2026.emnlp-industry.140/
%P 2355-2368
Markdown (Informal)
[Small Model, Big Retrieval: Unified Heterogeneous Retrieval via Multi-Label-Aware Contrastive Fine-Tuning](https://aclanthology.org/2026.emnlp-industry.140/) (Jiang et al., EMNLP 2026)
ACL
- Shaobai Jiang, Dantong Liu, Federico Gonzalez, Pedro Herrero-Vidal, You-Lin Chen, Xiaoyu Chu, and Prithviraj Sen. 2026. Small Model, Big Retrieval: Unified Heterogeneous Retrieval via Multi-Label-Aware Contrastive Fine-Tuning. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 2355–2368, Budapest, Hungary. Association for Computational Linguistics.