Source-linked AI summary
Every Article Deserves a Video: Contextual Video Matching for Digital Publishers
Arnaud Corone, Brice Pierre de la Briere, Gladys Roch, Samuel Leonardo Gracio, Yassine Bouher, Parvati Chauchaix
TL;DR
Digital publishers need relevant video in articles, but manual matching across large article and video catalogs is impractical. The paper presents Contextual Video Matching, an end-to-end system using LLMs, textual grounding, embeddings, and retrieval to automate this task. It reports improved relevance and engagement in production while identifying data and multimodal scope limitations.
Problem
Manual matching of large article and video catalogs is impractical, while existing research largely targets short text rather than full-length articles.
Method
Contextual Video Matching uses LLMs and vector embeddings to represent full-length articles, retrieve relevant videos, and serve matches at publisher scale.
Results
Contextual Video Matching improves relevance over baselines, raising score ≥4 matches from 15.5% to 19.9% versus basic HTML parsing and achieving the highest Bradley-Terry rating of 1666.
Takeaways & Limitations
The deployed system increases user engagement, enriches user experience, and provides value to publishers while extending Dailymotion player usage and revenue opportunities.
Takeaways & Limitations
The system currently relies only on textual information, with future work needed on video audio, visual features, transcripts, and learnable ranking.
Abstract
from arXiv · showhide
As digital publishers face the challenge of managing massive content catalogs, the ability to effectively embed relevant video within text-based articles has become critical for both monetization and user retention. However, manual selection is impractical for large scale publishers, especially when navigating their own extensive video libraries or the entire global Dailymotion catalog. In this paper, we present the "Contextual Video Matching" system, a solution that automatically matches relevant videos with text-heavy web pages and articles. By leveraging Large Language Models (LLMs) and textual embeddings, we provide a scalable solution for publishers to efficiently combine video content with their articles. We discuss in detail the motivations, architecture, evaluations, and deployment of this system within Dailymotion's production environment. Since its launch, the system has been adopted by hundreds of publishers, significantly increasing user engagement and enriching user experiences with highly relevant video content.
1 Introduction
Digital publishers need relevant video in articles for monetization, SEO, and asset reuse, but manual matching across large catalogs does not scale. The paper introduces Contextual Video Matching, a scalable system using LLMs and vector embeddings, and reports production deployment and online impact analysis.
- Motivation: Video integration can improve monetization, dwell time, and reuse of social media assets across proprietary platforms.The motivation identifies higher in-stream video CPMs, SEO optimization through increased dwell time, and continuous media ecosystems as drivers.
- Motivation: Manual editorial pairing cannot scale across daily articles, videos, millions of potential matches, and publishers’ own archives.Automated contextual intelligence is presented as necessary for operationally scalable matching.
- Research gap: Existing Video-Text Matching research mainly aligns videos with short snippets or queries rather than full-length articles.The paper states that this long-form article matching task is not explicitly studied in the known information-retrieval or recommendation literature.
- Contributions: The paper introduces an end-to-end system that uses LLMs and vector embeddings to bridge long-form text and video.Its contributions also cover industrial deployment for hundreds of publishers and online A/B testing of page quality and engagement.
2 Contextual Video Matching: System Design
Contextual Video Matching combines real-time webpage processing with catalog indexing and vector retrieval. Its design extracts and grounds article meaning, aligns it with multilingual video embeddings, and adapts filtering and ranking to each publisher.
- Architecture: The system is an end-to-end, multi-stage architecture for automated contextual media selection at scale.The pipeline includes structural text extraction, semantic synthesis, and vector-based candidate retrieval.
- Article processing: The online pipeline uses an LLM to extract meaningful article content from noisy HTML and explicitly identify the article language.Language filtering helps ensure that retrieved videos match the article language across a multilingual publisher network.
- Semantic representation: The system generates hypothetical video metadata with grounded entities, embeds it with MUSE, and retrieves actual videos through cosine similarity in a vector database.Grounding helps represent recent persons, places, and events that the frozen 2019 MUSE model may not know.
- Retrieval and ranking: Publisher-specific thresholds use embedding spread to filter candidates, while additional freshness or performance metrics rank videos that pass the threshold.The threshold is based on the 75th percentile of each publisher’s distribution, and publishers may override it editorially.
- Retrieval and ranking: Figure 3 compares semantic diversity across ten publishers using pairwise cosine similarity on MUSE embeddings of their 200 most recent videos.The distribution differences motivate publisher-specific thresholding rather than one global threshold.
3 Offline Evaluation
The offline evaluation benchmarks Contextual Video Matching against heuristic, parsing, summarization, and HyDE-based baselines using automated LLM judging and Bradley-Terry comparisons. The system consistently outperforms the baselines, although absolute relevance scores remain moderate.
- The evaluation uses 1,032 production web pages spanning news, sports, entertainment, and other thematic domains.
- Because labeled long-form article–video data is unavailable and manual relevance annotation is impractical, the study uses LLMs-As-A-Judge as an automated evaluation proxy.
- The judge assigns 1–5 relevance scores using topical relevance, reader intent, and reasoning-before-rating criteria.
- The tournament evaluation compares candidate videos pairwise and converts wins, losses, and ties into Bradley-Terry ratings with cluster-bootstrap confidence intervals.
- Contextual Video Matching outperforms all baselines: its mean score reaches 2.51, while highly relevant matches rise from 15.5% with basic HTML parsing to 19.9%.
- The system achieves the highest Bradley-Terry rating at 1666, but moderate absolute mean scores motivate dynamic thresholding before publisher delivery.
4 Online Experiments
The online experiment tests whether semantically matched videos improve consumption behavior relative to random selection. Across ten days of controlled production traffic, the system increases both visitor-level and view-level watch time.
- A ten-day A/B test randomly and persistently assigns production-network visitors to control or treatment groups.
- +19% relative Watch Time per Visitor is achieved compared with the random baseline.Watch Time per Visitor measures total video-watching time for a unique visitor.
- +21% relative Average Watch Time per View is achieved when videos are selected by the system.Average Watch Time per View measures the average duration of a single playback.
5 Deployment Challenges and Real-World Impact
Deployment combines publisher-agnostic extraction, cached LLM synthesis, and scalable vector retrieval to serve many publishers efficiently. After rollout, the system reached broad adoption, billions of views, and a measurable share of publisher-site revenue.
- Production deployment addresses latency, cost efficiency, and content extraction while serving hundreds of publishers.
- LLM-based parsing removes the need for domain-specific scrapers and enables automatic onboarding across different HTML structures.
- Caching the LLM-generated hypothetical video proxy limits processing to the first request, reducing latency and decoupling LLM costs from page traffic.
- Caching the document representation preserves dynamic retrieval, allowing newly uploaded videos to be recommended without another LLM invocation.
- Qdrant provides approximate-nearest-neighbor vector retrieval for Dailymotion’s large catalog and can be self-deployed on Kubernetes.
- The system serves over 200 publishers across hundreds of thousands of pages and has generated billions of video views.
- Contextual Video Matching accounts for 2% to 3% of Dailymotion’s revenue from publisher sites.
- The paper characterizes LLM-grounded contextual matching as a driver of global ad-driven revenue streams.
6 Conclusion & Future work
Contextual Video Matching addresses the challenge of connecting full-length editorial articles with relevant videos at scale, increasing engagement and enriching user experience. The authors describe multimodal enrichment and improved ranking as important directions for future work.
- The system bridges the semantic gap between full-length editorial articles and video content for publishers producing large volumes of both.
- It increases user engagement, enriches user experience, and provides value to publishers.
- For Dailymotion, the system creates a new revenue source and expands the contexts in which its player is used, improving brand visibility.
- The current system relies solely on textual information, leaving multimodal video signals such as audio and visual features as improvement opportunities.
- Future work includes incorporating video transcripts into embeddings and exploring learnable ranking functions to refine article–video alignment.