Abstract
Generative Document Retrieval (GDR) has emerged as a promising alternative to traditional "index-then-retrieve" pipelines by training models to generate unique document identifiers (IDs) directly from queries. However, existing GDR methods often treat documents as flat sequences of text, ignoring internal logical structures and using semantically shallow identifiers. This paper introduces Event GDR, a framework that redefines documents as structured collections of events. By extracting event relations through a multi-agent Large Language Model (LLM) approach and mapping them to a hierarchical taxonomy, Event GDR creates semantically rich document representations and identifiers. Experimental results on English and Chinese benchmarks demonstrate that this event-centric approach significantly outperforms state-of-the-art GDR baselines.
Introduction
Traditional document retrieval systems typically rely on dense or sparse vector representations to match queries against a document corpus. Generative Document Retrieval (GDR) simplifies this process by tasking a single model with generating a document's unique ID. Despite its elegance, GDR faces two primary challenges: it often ignores the internal context and causal correlations within a document, and it frequently utilizes flat, arbitrary identifiers that lack semantic depth.
To address these issues, the authors propose that events—defined by actors, actions, and their causal links—serve as the natural units of information. By indexing the "structural skeleton" of a document rather than just its raw text, retrieval models can better capture the underlying logic of the content.
Methodology
Event GDR introduces three core innovations to transform the retrieval process:
1. Exchange-then-Reflection (ExR) Extraction
To ensure high-fidelity event data, the framework employs a multi-agent LLM strategy called Exchange-then-Reflection (ExR). Instead of relying on a single model, multiple LLM agents (e.g., ChatGPT and GLM-130B) first exchange insights regarding the events and relations within a document. They then reflect on this collective feedback to refine the final list of events and their causal links.
2. Event-Centric Representation
Documents are represented through EReps (individual event mentions) and ERReps (pairs of events linked by causal relations). This structured approach ensures the model understands the "connective tissue" of the document, such as how one event leads to another.
3. Taxonomy-based Identifiers (ETIds)
Rather than using random integers or simple titles, Event GDR maps extracted events to a structured hierarchy (an event taxonomy). A document’s identifier becomes a semantic path through this tree (e.g., Change -> Change_state). This "semantic addressing" allows the model to navigate a meaningful map of document contents during the retrieval phase.
Experimental Results
The model was evaluated on two major benchmarks: NQ_10K (English Wikipedia questions) and DuReader_50K (Chinese passage retrieval).
- Performance Gains: Event GDR consistently outperformed established baselines. On NQ_10K, it achieved a Hits@1 of 73.25, surpassing the previous best GDR variant (DSI-QG-SEM) by nearly 3 points. On DuReader_50K, it reached a Hits@1 of 63.96, outperforming the best baseline by over 2 points.
- Ablation Insights: Studies revealed that incorporating event relations specifically provided a 7.7% improvement in representation quality, confirming that the links between events are as critical as the events themselves.
- Scalability: Performance scaled positively with model size (tested up to 13B parameters), indicating that the method benefits from more powerful underlying architectures.
Conclusion and Future Work
Event GDR demonstrates that treating documents as structured event graphs significantly enhances generative retrieval. While the multi-agent ExR process introduces some computational overhead during indexing, the resulting semantic identifiers bridge the gap between query intent and document content. Future research will explore expanding the framework to include temporal and sub-event relations to create even richer document representations.
Summary updated