Multi-Document Fusion

The foundational step of identifying when different textual expressions across multiple documents refer to the same real-world entity. In legal fusion, the plaintiff 'Acme Corp' in Document A must be algorithmically linked to 'Acme Corporation' and 'the Claimant' in Documents B and C. Without this resolution, the fusion engine will treat them as separate entities, generating a fragmented and redundant summary. This process often relies on entity linking against a knowledge base or anaphora resolution across document boundaries.

Fusing documents requires reconstructing a single, non-contradictory timeline from multiple narratives. A contract's effective date in one document must be sequenced against a breach notice date in another. The fusion engine must detect and resolve temporal inconsistencies—such as a filing date that precedes a cited event—by applying temporal reasoning to normalize all events onto a single, coherent axis. This is critical for legal narrative construction where the sequence of events determines liability.

A core algorithmic defense against repetitive output. Using Maximum Marginal Relevance (MMR) , the system iteratively selects passages that maximize relevance to the central topic while minimizing similarity to already-selected content. In a multi-document corpus where five sources describe the same precedent, MMR ensures the final summary includes the precedent only once, citing the most authoritative or comprehensive version. This is computed using cosine similarity between candidate passage embeddings.

Unlike single-document summarization, fusion must actively identify conflicting factual claims. A Natural Language Inference (NLI) model classifies whether a statement from Document A contradicts one from Document B. When a contradiction is detected, the fusion engine cannot simply average the claims. Resolution strategies include: deferring to the most recent document, prioritizing the source with higher authority weight (e.g., a court ruling over a party brief), or explicitly flagging the unresolved conflict in the output for human review.

Every fused statement must retain a pointer back to its originating document. This source attribution is non-negotiable in legal applications where citation integrity is paramount. The fusion output is not a monolithic block of text but a structured object where each claim is annotated with its provenance. This enables downstream citation verification systems to validate the output and allows an attorney to click on any sentence in the fused summary and instantly retrieve the exact source paragraph.

When the total token count of the corpus exceeds a model's context window, a hierarchical fusion strategy is required. The corpus is partitioned into manageable chunks, each chunk is independently fused into an intermediate summary, and then those intermediate summaries are recursively fused in a second pass. This 'map-reduce' approach for summarization allows the system to synthesize information across hundreds of documents, though it requires careful prompt engineering to prevent information loss during the intermediate condensation steps.