One-Step Retrieval Framework for Real-Time Sponsored Search Ads Using Hierarchical Text Representations

简体中文

Core mechanism on public data. MovieLens 100K exercises hierarchical identifiers, joint scoring, and constrained retrieval, without Tencent's private ads logs or production index.

Paper information

Field	Details
Paper	arXiv v1
Organization	Tencent
First public release	2026-09-16 (arXiv v1)
Authors' code	No public author repository was found as of 2026-09-19.
Adapter	angle
Local reproduction code	src/auto_research/reproductions/angle/

Original paper

Motivation and method

The semantic hierarchy combines intent and summary IDs. Training joins generation, discrimination, and ranking objectives. Request-time beam constraints allow only valid identifiers.

flowchart LR
  H[Public history] --> I[Intent and summary hierarchy]
  I --> J[Joint generative, discriminative, and ranking scores]
  J --> B[Dynamically constrained beam]
  B --> E[Offline evaluation on the same split]

Key figure from the paper

Figure 1: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.

Core computation

The local score s = s_gen + 0.15 s_dis retains generative relevance and hierarchical-ID discrimination. Request-specific valid IDs constrain the beam; the fusion coefficient is selected only on validation.

Paper results

On Weixin Top Stories, online A/B measured consumption +1.81%, GMV +2.16%, clicks +1.50%, conversion +1.44%, and impressions +2.49%. These production results are not local MovieLens measurements.

Local reproduction

See the local results at seeds 42/43/44.

Local comparison: Baseline and treatment use the same MovieLens 100K candidates, temporal split, and evaluation. Local relative changes are in the metrics file, not substituted with online percentages.

Reproduction boundaries

MovieLens 100K exercises hierarchical identifiers, joint scores, and constrained retrieval. No private Tencent ads logs or production index are used.