@inproceedings{khazanchi-etal-2026-lightweight,
title = "Lightweight Temporal Attention for Video Retrieval over Frozen Image and Text Encoders",
author = "Khazanchi, Vidit and
Sharma, Umang and
Komirishetty, Shabhareesh",
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.116/",
pages = "1976--1994",
ISBN = "979-8-89176-460-6",
abstract = "Large-scale visual search platforms maintain production text-image embedding models trained on behaviorally-mined query-image pairs, where the text encoder is implicitly tuned to organisation-specific search vocabulary. Extending such a system to video retrieval is non-trivial: retraining risks disrupting hard-won query understanding, while naive temporal aggregation discards the sequential structure that distinguishes a door opening from a door closing. We add a lightweight temporal attention pooling head over independently encoded frame embeddings, keeping both encoders frozen and training only the pooling head and positional encodings. We train on a large-scale dataset from a commercial creative platform using bidirectional InfoNCE with a two-stage hard negative mining pipeline. Benchmarking against LoRA fine-tuned Gemma 3{~}(12B) and SmolVLM2, neither matches our approach despite substantially larger parameter counts. We also introduce an evaluation benchmark with graded relevance judgments annotated by a VLM judge, where each query is mapped to all relevant videos in a large corpus rather than a single ground-truth clip, avoiding protocols that greatly underestimate model quality. Our model improves nDCG@1 by 10.1{\%} over mean-pooling overall and 25{\%} on complex action queries. With Matryoshka Representation Learning, 256-dim embeddings retain 95.6{\%} of full performance, and int8 quantization costs under 1{\%} in retrieval quality."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="khazanchi-etal-2026-lightweight">
<titleInfo>
<title>Lightweight Temporal Attention for Video Retrieval over Frozen Image and Text Encoders</title>
</titleInfo>
<name type="personal">
<namePart type="given">Vidit</namePart>
<namePart type="family">Khazanchi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Umang</namePart>
<namePart type="family">Sharma</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shabhareesh</namePart>
<namePart type="family">Komirishetty</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>Large-scale visual search platforms maintain production text-image embedding models trained on behaviorally-mined query-image pairs, where the text encoder is implicitly tuned to organisation-specific search vocabulary. Extending such a system to video retrieval is non-trivial: retraining risks disrupting hard-won query understanding, while naive temporal aggregation discards the sequential structure that distinguishes a door opening from a door closing. We add a lightweight temporal attention pooling head over independently encoded frame embeddings, keeping both encoders frozen and training only the pooling head and positional encodings. We train on a large-scale dataset from a commercial creative platform using bidirectional InfoNCE with a two-stage hard negative mining pipeline. Benchmarking against LoRA fine-tuned Gemma 3 (12B) and SmolVLM2, neither matches our approach despite substantially larger parameter counts. We also introduce an evaluation benchmark with graded relevance judgments annotated by a VLM judge, where each query is mapped to all relevant videos in a large corpus rather than a single ground-truth clip, avoiding protocols that greatly underestimate model quality. Our model improves nDCG@1 by 10.1% over mean-pooling overall and 25% on complex action queries. With Matryoshka Representation Learning, 256-dim embeddings retain 95.6% of full performance, and int8 quantization costs under 1% in retrieval quality.</abstract>
<identifier type="citekey">khazanchi-etal-2026-lightweight</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-industry.116/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1976</start>
<end>1994</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Lightweight Temporal Attention for Video Retrieval over Frozen Image and Text Encoders
%A Khazanchi, Vidit
%A Sharma, Umang
%A Komirishetty, Shabhareesh
%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 khazanchi-etal-2026-lightweight
%X Large-scale visual search platforms maintain production text-image embedding models trained on behaviorally-mined query-image pairs, where the text encoder is implicitly tuned to organisation-specific search vocabulary. Extending such a system to video retrieval is non-trivial: retraining risks disrupting hard-won query understanding, while naive temporal aggregation discards the sequential structure that distinguishes a door opening from a door closing. We add a lightweight temporal attention pooling head over independently encoded frame embeddings, keeping both encoders frozen and training only the pooling head and positional encodings. We train on a large-scale dataset from a commercial creative platform using bidirectional InfoNCE with a two-stage hard negative mining pipeline. Benchmarking against LoRA fine-tuned Gemma 3 (12B) and SmolVLM2, neither matches our approach despite substantially larger parameter counts. We also introduce an evaluation benchmark with graded relevance judgments annotated by a VLM judge, where each query is mapped to all relevant videos in a large corpus rather than a single ground-truth clip, avoiding protocols that greatly underestimate model quality. Our model improves nDCG@1 by 10.1% over mean-pooling overall and 25% on complex action queries. With Matryoshka Representation Learning, 256-dim embeddings retain 95.6% of full performance, and int8 quantization costs under 1% in retrieval quality.
%U https://aclanthology.org/2026.emnlp-industry.116/
%P 1976-1994
Markdown (Informal)
[Lightweight Temporal Attention for Video Retrieval over Frozen Image and Text Encoders](https://aclanthology.org/2026.emnlp-industry.116/) (Khazanchi et al., EMNLP 2026)
ACL