Traditional information retrieval models often treat documents as monolithic blocks of text. While this works well for short abstracts, it introduces severe bottlenecks when applied to long-form documents. Long documents suffer from thematic drift—where a single text transitions through multiple unrelated topics—and noise accumulation, where highly relevant sentences are diluted by pages of irrelevant background text. Furthermore, modern Transformer-based ranking architectures (like BERT) are bound by strict input length limitations, typically capping inputs at 512 tokens.
To bridge this gap, modern retrieval systems decompose long documents into granular segments, score those segments independently, and synthesize a global document-level relevance score. This paradigm, which traces its roots back to early hierarchical "global-local" matching frameworks proposed by Salton et al., allows neural rankers to scale to long texts without losing precision.
Granular Decomposition Strategies
Before a system can aggregate relevance signals, it must first break a document down into manageable units. The choice of segmentation strategy directly impacts the resolution of the evidence collected:
- Fixed-Length Sliding Windows: Documents are sliced into uniform token or character windows (e.g., 100-word blocks). Wang et al. and Ai et al. utilize fixed-length windows to isolate specific topics and prevent thematic drift from penalizing localized matches.
- Logical Paragraphs: Text is segmented by natural paragraph boundaries. Wang et al. demonstrate that treating the paragraph as the primary unit of relevance allows systems to analyze the close proximity of query terms and entities within a coherent semantic block.
- Sentence-Level Indexing: The most granular approach treats individual sentences as independent units. White et al. show that ranking sentences independently can surface highly precise snippets that would otherwise be buried under a low document-level score.
The Aggregation Dilemma: Max vs. Average
Once a document's constituent passages have been scored against a query, the system faces a fundamental challenge: how should these local scores be combined into a single, representative document score? Two primary philosophies dominate this space.
The Max-Pooling Hypothesis (Single-Hit Relevance)
The simplest approach operates on the assumption that a single highly relevant passage is a sufficient indicator of a document's overall utility. In this setup, the document's final score is simply the maximum score achieved by any of its segments.
Chen et al. and Dai et al. demonstrated that maximum passage scoring consistently outperforms summation strategies in document re-ranking. The intuition behind why this works is detailed by Nguyen et al.: summing or averaging scores across a long document accumulates noise from the vast majority of passages that are completely irrelevant to the query. Selecting the maximum score acts as a robust filter, ensuring that a single concentrated answer can successfully surface the parent document.
The Top-K / Average Hypothesis (Distributed Relevance)
While max-pooling is highly effective for factoid queries, it falls short when a document contains distributed relevance signals—where a comprehensive answer is scattered across multiple sections, or where a document addresses multiple sub-topics of a complex query.
To capture these broader signals, Chen et al. and Chen et al. proposed aggregating scores by averaging the top-$k$ highest-ranking passages (such as the top two or three segments). This strategy provides a more stable proxy for document utility, ensuring that documents with sustained, distributed evidence of relevance are prioritized over those that contain only an isolated keyword match.
Moving to Representation-Level Fusion
While scalar pooling methods (like taking the maximum or average score) are computationally efficient, they treat passage scores as entirely independent numbers. This ignores the rich semantic interactions and contextual flow between different parts of the same document. As Akkalyoncu et al. point out, simple linear combinations fail to model how disparate pieces of evidence collectively satisfy a user's information need.
To resolve this, modern architectures shift from pooling scores to pooling deep representations.
Instead of condensing a passage into a single scalar score via a neural network, the system extracts dense embedding vectors for each passage. Li et al. demonstrated that feeding these passage-level representations into a shallow Transformer layer yields significantly better document-level rankings than max-pooling. By allowing the passages to attend to one another at the representation level, the model can dynamically weigh how different segments complement each other.
Taking this a step further, Gao et al. introduced modular re-rankers utilizing a joint query-to-all-chunk attention mechanism. This architecture allows the query to look across all document blocks simultaneously during the encoding phase. By synthesizing information across disjoint regions before computing final scores, the model successfully preserves local granularity while capturing the cumulative, global context of long-form text.
Key Foundational Work
Selection of Passages for Information Reduction (1996) — contributed 2 concept elements to this direction.
Approaches to Passage Retrieval in Full Text Information Systems (1993) — contributed 1 concept elements to this direction.