The paper "Real-time Ad Retrieval via LLM-generative Commercial Intention for Sponsored Search Advertising," authored by researchers from Tencent Inc., presents a novel framework designed to enhance ad retrieval processes in large-scale online systems. The framework, known as Real-time Ad REtrieval (RARE), employs customized LLMs to generate Commercial Intentions (CIs) as semantic intermediaries, facilitating efficient and scalable ad retrieval for user queries.
Problem Context
Traditional ad retrieval systems typically engage in a two-stage process, utilizing manually selected keywords to fetch ads. This approach, while functional, imposes significant inefficiencies and recall issues due to the gap between the user's queries and the keyword choices. Moreover, methods utilizing heavy DocIDs face challenges concerning real-time generation and scalability, particularly in handling vast ad libraries and dynamically updating candidate sets.
Proposed Framework
RARE addresses these inefficiencies by replacing keyword or DocID-based indexing with LLM-generated CIs. These CIs encapsulate the commercial intentions associated with ads and queries, allowing for direct and real-time retrieval of advertisements. The framework employs a two-pronged LLM strategy: the offline phase generates a corpus of CIs for ads, while the online phase facilitates real-time CI generation for user queries using constrained beam search.
Key Innovations
- Customized LLM Development: The authors enhance the base LLM through knowledge injection and format fine-tuning, incorporating domain-specific commercial knowledge. This enables the LLM to generate relevant CIs, improving semantic retrieval efficiency and extension.
- Constrained Beam Search: This decoding process ensures that meanings are constrained to the CI set, optimizing both the diversity and relevance of generated ads while maintaining decoding efficiency.
- Dynamic Indexing: RARE constructs a dynamic ad index, mapping CIs to ads in a one-to-many relationship, enabling a scalable and versatile retrieval model suitable for real-time environments.
Empirical Evaluation
RARE's implementation spans real-world applications with high query volumes, exceeding hundreds of millions per day. The paper reports quantitative improvements in operations using RARE, specifically a 5.04% increase in consumption, a 6.37% rise in Gross Merchandise Volume (GMV), a 1.28% enhancement in click-through rate (CTR), and a 5.29% increase in shallow conversions. The framework outperformed ten baseline methods across various metrics, including HR@500, MAP, and ACR.
Implications and Future Directions
The deployment of RARE in practical environments underscores its potential to transform ad retrieval mechanisms by enhancing both semantic alignment and system efficiency. This framework's embrace of LLMs in ad retrieval sets a precedent for exploring further integration of LLMs in commercial applications. Additionally, this method highlights the growing viability of generative models in real-time ad systems beyond traditional retrieval processes.
Future developments in this area might include refining generative capabilities to incorporate real-time semantic feedback and context-aware modifications in commercial intents. Additionally, exploring hybrid models that juxtapose generative and discriminative predictions could enhance alignment accuracy, thus expanding the application scope of sequence-generated systems in commercial technology platforms.
In summary, RARE represents a significant step in optimizing sponsored search systems through innovative use of LLMs, proving its merits both theoretically and practically in extensive online frameworks.