Boosting Ranking Performance with Minimal Supervision
06-19, 12:00–12:40 (Europe/Berlin), Kesselhaus

Using generative Large Language Models (LLMs) to generate synthetic labeled data to train in-domain ranking models. Distilling the knowledge and power of generative LLMs into effective ranking models.


Transformer language models are highly effective text rankers; however, training Transformer-based neural ranking models requires vast amounts of labeled supervised data, which is costly and time-consuming. What if you could teach a ranking model without behavioral click data or human annotations? Enter generative large language models (LLMs) such as GPT-3.

This talk showcases a novel approach to generating labeled data with minimal human supervision. First, with just three human-labeled queries and document examples, an open-source LLM generates synthetic questions for all documents in the index. Then, the synthetic data trains a much smaller, cost-efficient Transformer ranking model, which outperforms a strong BM25 baseline by 10 nDCG@10 points on a popular relevance dataset.

The innovative method saves on costly annotation efforts and enables faster adaptation to search ranking in new domains, and allows organizations to revolutionize their search capabilities without breaking the bank.

Jo Kristian is a Distinguished Engineer @Yahoo, where he spends his time working on the open-source Vespa.ai serving engine. Jo Kristian has 20 years of experience with deploying search systems at scale.