On Synthetic Data Strategies for Domain-Specific Generative Retrieval
This work addresses scalability issues in domain-specific generative retrieval by reducing reliance on manual annotation, though it appears incremental in its methodological contributions.
This paper tackles the challenge of scaling generative retrieval models for domain-specific corpora by investigating synthetic data generation strategies, including LLM-generated queries with domain constraints and hard negative mining for preference learning. Experiments on public datasets across diverse domains demonstrate the effectiveness of these approaches.
This paper investigates synthetic data generation strategies in developing generative retrieval models for domain-specific corpora, thereby addressing the scalability challenges inherent in manually annotating in-domain queries. We study the data strategies for a two-stage training framework: in the first stage, which focuses on learning to decode document identifiers from queries, we investigate LLM-generated queries across multiple granularity (e.g. chunks, sentences) and domain-relevant search constraints that can better capture nuanced relevancy signals. In the second stage, which aims to refine document ranking through preference learning, we explore the strategies for mining hard negatives based on the initial model's predictions. Experiments on public datasets over diverse domains demonstrate the effectiveness of our synthetic data generation and hard negative sampling approach.