Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

LLMDistill4Ads: Using Cross-Encoders to Distill from LLM Signals for Advertiser Keyphrase Recommendations at eBay

Created by
  • Haebom

Author

Soumik Dey, Benjamin Braun, Naveen Ravipati, Hansi Wu, Binbin Li

Outline

This paper studies improving a keyword recommendation system to improve the performance of eBay sellers' advertising campaigns. Because the existing embedding-based retrieval (EBR) model suffers from click data bias, we propose a two-step LLM distillation process that utilizes a large-scale language model (LLM) as a judge to remove this bias. First, we extract knowledge from the LLM judger using a cross-encoder as an intermediate step, and then distill this knowledge into a bi-encoder model through multi-task learning. Finally, we use the distilled bi-encoder model to recommend relevant keywords to sellers. Experimental results demonstrate that the proposed method improves the performance of the bi-encoder, which searches for relevant keywords for sellers on eBay.

Takeaways, Limitations

•
Takeaways:
◦
We demonstrate that the bias of the EBR model can be effectively removed through the knowledge distillation technique using LLM.
◦
We present an effective method to improve the performance of bi-encoder models through multi-task learning.
◦
It can help improve the performance of sellers' advertising campaigns on large e-commerce platforms such as eBay.
•
Limitations:
◦
Because of the high reliance on the LLM judge, the results may be affected by the performance of the LLM.
◦
Large datasets and computing resources may be required.
◦
The experiment was specific to the eBay environment and its generalizability to other platforms requires examination.
◦
There may be a lack of discussion about the bias of LLM judges themselves.
👍
Made with Slashpage