One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations
summary
The gist
One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations proposes an end-to-end retrieval framework that integrates ad generation, discrimination…
In short
The framework integrates ad generation, discrimination, and ranking into a single Large Language Model (LLM) to replace traditional multi-stage systems. It uses hierarchical text representations—commercial intent for overview and ad abstracts for detail—to create efficient retrieval. This end-to-end approach improves performance in real-time sponsored search by leveraging the LLM's full capabilities.
Key concepts
- Multi-stage Cascading Architectures (MCA)
- Traditional systems use multiple, independently optimized modules in a chain. This causes inconsistencies between stages and can lead to the premature elimination of good ads before they are fully assessed, resulting in inefficient retrieval.
- Hierarchical Semantic Text
- This is an intermediate text representation created by the LLM. It consists of two parts: 'commercial intent,' which gives a high-level summary of the ad's purpose, and 'ad abstract,' which provides fine-grained details about the specific advertisement.
- GDR-LLM Training Losses
- The model is trained using three main loss functions: Lgen (cross-entropy) to learn generation accuracy, Ldis (binary cross-entropy) to learn a relevance score from text, and Ldpo (Direct Preference Optimization) to teach the model the relative ranking of ads.
- Ad Indexing Module
- This module builds an inverted index. It maps pairs of intent-abstract text (int-abs pairs) directly to their corresponding advertisements, allowing for precise and efficient retrieval based on these structured semantic representations.
Terminology used across episodes
This episode discusses
- One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations · Paper Radio
- GPT-4 Technical Report
- Hit ratio: An Evaluation Metric for Hashtag Recommendation
- Qwen Technical Report
- OneSearch: A Preliminary Exploration of the Unified End-to-End Generative Framework for E-commerce Search
- UniSearch: Rethinking Search System with a Unified Generative Architecture
- OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment
- Hierarchical Multi-field Representations for Two-Stage E-commerce Retrieval
- FORGE: Forming Semantic Identifiers for Generative Retrieval in Industrial Datasets
- Towards Large-scale Generative Ranking
- Multiview Identifiers Enhanced Generative Retrieval
- Real-time Ad retrieval via LLM-generative Commercial Intention for Sponsored Search Advertising
- OneRec-Think: In-Text Reasoning for Generative Recommendation
- GPR: Towards a Generative Pre-trained One-Model Paradigm for Large-Scale Advertising Recommendation
- OneMall: One Architecture, More Scenarios -- End-to-End Generative Recommender Family at Kuaishou E-Commerce
- Rethinking the Role of Pre-ranking in Large-scale E-Commerce Searching System
- OpenOneRec Technical Report
The paper
One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations · Read on arXiv
Tongtong Liu, Renyu Zhang, Jiayu Ding, Hongchao Guo, Xintao Yang, He Wei, Zhaoyu Li, Haiyang Wu
Tencent Inc.
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations".
Jane: Here's the gist of the paper. One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations proposes an end-to-end retrieval framework that integrates ad generation, discrimination, and ranking into a single LLM to address the inconsistencies inherent in traditional multi-stage cascading architectures The primary innovation of this work lies in utilizing hierarchical semantic text as an intermediate representation and designing a GDR-LLM to enable end-to-end retrieval from users to ads Pruning redundant reasoning tokens Traditional retrieval systems typically use multi-stage cascading architectures (MCA), where each module is optimized independently, leading to inconsistent objectives and the premature elimination of high-potential candidates The one-to-one mapping between SIDs and advertisements leads to inefficient decoding Moreover, these methods rely on a small reward model (e.g.
Tom: pctr) for relevance and ranking, limiting the LLM’s ability to fully assess ads’ commercial value How the training signal is built The GDR-LLM relies primarily on three loss functions for training: Lgen, Ldis, and Ldpo 1.
Jane: Learning to Generate involves the cross-entropy loss function (Lgen) to measure the discrepancy between predicted probability distribution and ground-truth labels 2.
Tom: Learning to Discriminate uses a multilayer perceptron (MLP) function f: Rd → R to obtain a scalar relevance score s, minimizing binary crossentropy loss Ldis 3.
Paper summary: Lu: Thinking about the implications of this work, it seems like the main shift is moving from piece-wise optimization in a pipeline to a single, unified model that handles everything from understanding user intent to ranking the final ad.
Jane: That unification means we stop worrying as much about how well each individual component performs in isolation and start focusing on how the whole system works together dynamically.
Tom: The paper proposes ANGLE as this unified approach, which uses a hierarchical semantic text—intent and abstract pairs—to bridge the gap between understanding what a user wants and finding the most relevant ad.
Meng: From my side, it suggests that if we can build representations that capture both the high-level intent and the fine details simultaneously, we might be able to deploy more robust systems without needing dozens of separate fine-tuned components.
Jane: It really changes how we think about how an AI should process a search query; instead of a series of steps, it's one comprehensive generation and ranking step.
Lu: The limitation the authors themselves flag is that they are still relying on generating these intent-abstract pairs first, which means the quality of the retrieval heavily depends on how accurately those initial representations are formed.
Tom: And I think that points toward future work needing to focus on making sure those initial representations capture enough commercial value so the subsequent ranking step isn't fighting against a poor foundation.
Jane: So for someone just listening, this paper suggests that the future of ad retrieval isn't just better individual components, but a single system where intent and abstract information flow seamlessly from search to display.
Conclusion: Tom: So we’re wrapping up on ANGLE, which is that One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations by their authors at Google Research and a few others.
Jane: It basically means they built one single system that does ad generation, discrimination, and ranking all at once instead of using a bunch of separate tools in a pipeline.
Lu: The core idea is this hierarchical text, which has commercial intent for the big picture and an ad abstract for the fine details.
Meng: From what I can tell about deployment, it’s trying to move away from those older methods where you had to memorize every single ad mapping because that gets really expensive and hard to maintain.
Tom: Yeah, they use this mechanism to make the ranking much more direct by letting the LLM do all three things with dynamic constrained decoding.
Jane: So for a listener, it’s about an AI that understands what you want and finds the best ad in one go without needing separate steps for intent and relevance.
Lu: The training uses three specific loss functions, Lgen for generation, Ldis for discrimination scoring, and DPO for learning the final ranking preference.
Meng: I’m interested in how they handle the scaling of that process; if that hierarchical text gets too big at scale, does it still keep up with real-time requirements?
Tom: Well, they showed results in a real search scenario where they saw an increase of one point eight one percent in consumption and two point one six percent in gross merchandise volume.
Jane: That’s a concrete number showing that this single framework isn't just theoretical; it actually moves the needle on how much people buy.
Lu: Offline evaluations against seven baselines show ANGLE wins on key metrics like HR and ACR, which are measures of relevance and accuracy.
Tom: It’s interesting because they’re not just claiming better performance; they’re showing how this unified approach actually beats the traditional multi-stage architectures on those specific measures.
Jane: But what this really means is that we might stop spending time optimizing each small piece and start focusing on building one comprehensive model for retrieval.
Lu: And if you look at where they left off, there are still some limitations they mentioned regarding the quality of the initial intent-abstract pairs used to build that index.
Meng: That’s a fair point; if those initial representations are weak, the ranking step might just be ranking bad things better.
Jane: Exactly, so while this framework is a big step forward in unification, we still need to figure out how to make sure the foundation of that intent-abstract text is solid enough for real-world traffic.
Tom: And that brings us right up against what I’ve been thinking about next—how much does this unified system actually cost to run when you try to deploy it at scale?
More episodes
- 2505.16254-Human-AI Writing as Distributed Cognition: Reassessing Collaborative Writing Theories in the Age of LLMs
- 2506.10913-Choreographic Quick Changes: First-Class Location (Set) Polymorphism
- 2507.14812-A Black-Box Approach for Exogenous Replenishment in Online Resource Allocation
- 2507.15448-Frames over finite fields with respect to the Galois inner product and their connections to self-orthogonal and LCD codes
- 2508.21572-NewsReX: An Open-Source Multi-Framework for Neural News Recommendation
- 2509.17028-Impact of Packetization on Network Calculus Analysis
- 2510.17342-AoA Services in 5G Networks: A Framework for Real-World Implementation and Systematic Testing
- 2511.04088-Efficient and rate-optimal list-decoding in the presence of minimal feedback
- 2511.17264-Notes on Stack Machines and Quantum Stack Machines
- 2601.15318-On the closest balanced game