Global Coherence for Entity Resolution

164 concepts123 papers

When a natural language processing system encounters an ambiguous name like "Ford" or "Gates," the words immediately surrounding the mention—the local context—are often insufficient. A sentence like "Ford announced a new strategy today" provides almost no clues to determine whether it refers to the motor company, the historical president, a theater, or a completely different entity.

Traditional entity resolution and information extraction systems treat mentions as independent units, processing them in isolation. However, this independent resolution strategy ignores the semantic relationships and mutual constraints between different mentions within the same document, frequently leading to logical inconsistencies across a text, as noted by Yiping Zhou et al. and Thomas Demeester et al..

To resolve these ambiguities, researchers have shifted toward global document coherence: an algorithmic paradigm that resolves all mentions across a document simultaneously by enforcing thematic and semantic consistency.

The Local-Global Context Gap

Why does local context fail? In long-form narratives, the critical evidence required to establish an entity's category or identity is often scattered across non-contiguous paragraphs rather than being concentrated in a single sentence (Hongyin Tang et al.). Standard Transformer architectures exacerbate this issue due to their limited context windows (typically 512 tokens), which cut off vital information (Arthur Amalvy et al.).

Without a mechanism to capture broader document-level semantics, systems struggle to maintain topical consistency (Angel X Chang et al.). Jianfeng Gao et al. and Li Dong et al. demonstrate that global document-level context is frequently required to distinguish names that appear consistently throughout a text. If a model only has access to a narrow window of text, it suffers from a form of "tunnel vision."

The Intuition of Collective Optimization

The core intuition behind global coherence is that entities appearing within the same document are usually topically related. If a document mentions "Ford," "Model T," and "Detroit," the collective presence of these terms makes it highly probable that "Ford" refers to the automotive company.

Instead of resolving each mention independently, global coherence frameworks optimize for the entire document at once. In their seminal work, Sayali Kulkarni et al. showed that integrating topical relatedness across all entities serves as a vital precision-enhancing mechanism. By simultaneously optimizing for local compatibility and global document coherence, the system allows the clear identity of one entity to resolve the ambiguity of another nearby mention (Veselin Stoyanov et al.).

[Isolated Processing]
"Ford"  ---> ? (Could be Henry Ford, Ford Motor Co., or Gerald Ford)
"Model T" ---> Ford Model T

[Global Coherence Processing]
"Ford" <--- (Topically Related) ---> "Model T"
Result: "Ford" is resolved to Ford Motor Co. due to document-wide consistency.

Algorithmic Paradigms for Global Resolution

To bridge the local-global gap and enforce document-wide consistency, researchers have explored several distinct mechanisms:

1. Joint Inference and Feedback Loops

Rather than treating information extraction as a rigid, one-way pipeline, some approaches use downstream tasks to fix upstream errors. Heng Ji et al. introduced feedback loops where coreference resolution and relation detection inform name tagging, allowing global discourse context to override local ambiguities.

Similarly, Hannaneh Hajishirzi et al. demonstrated that solving coreference resolution and named-entity linking jointly allows world knowledge to fix coreference errors while document context fixes linking errors. In event extraction, Shasha Liao et al. leveraged cross-event inference to ensure that an entity's role remains consistent across an entire narrative.

2. Sequential and Reinforcement Learning

Global optimization can be computationally expensive if every candidate entity is compared against every other candidate. To mitigate this, Yanan Cao et al. framed joint entity linking as a sequence decision problem using Deep Reinforcement Learning. By sorting mentions from "easiest to hardest," the model uses previously linked, unambiguous entities to dynamically enrich the context for harder, more ambiguous mentions later in the sequence.

3. Generative Topic and Graph Models

Another approach is to explicitly model the document's underlying theme. Xianpei Han et al. proposed a generative entity-topic model that jointly captures dependencies between words, mentions, entities, and topics. This allows global topic coherence to act as a secondary evidence source, boosting the resolution of obscure or non-salient mentions that lack strong local cues. Other frameworks leverage external structural knowledge, such as Wikipedia's hyperlink graph, to enforce semantic consistency across the text (Ander Barrena et al.).

The Balancing Act

While global coherence provides a powerful macro-level signal, it cannot entirely replace local context. As Phong Le et al. point out, local context remains essential for distinguishing between entities that share similar thematic domains. For example, in an article about physics, global coherence might correctly identify the domain but fail to distinguish between "Einstein" and "Bohr" without immediate textual descriptors. The most robust systems must therefore find a principled way to balance immediate local cues with document-wide semantic constraints.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 130 papers
  2. Version 1 · 2026-10-04 · 221 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_global_document_coherence_e89a022d,
  title = {global document coherence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/global-document-coherence/rdr_e89a022dab4d4e57991db61bee20bea2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_global_document_coherence_e89a022d_v2,
  title = {global document coherence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/global-document-coherence/rdr_e89a022dab4d4e57991db61bee20bea2/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake