Does each column represent a word?

In the classic entity-based construction, it represents a discourse entity and can combine several referring expressions.

Must every entity appear in every sentence?

No. Absence is explicitly represented and contributes to transition patterns.

Is the grid itself a coherence classifier?

No. It is a representation from which features can be derived; a scoring or learning method is an additional component.

AI Glossary / Natural language and speech / Intermediate

Entity grid

Definition

An entity grid is a table that represents how people, objects, and other discourse entities appear across the sentences of a document. In a classic version, rows represent sentences, columns represent entities, and cells record grammatical roles or absence. Patterns in these cells can help a model assess local coherence.

What it means

A readable document usually connects its sentences through continuing subjects and references. For example, a proposal may first be introduced, then described as “it,” and later mentioned by name again. An entity grid makes some of that continuity visible.
The columns refer to entities, rather than simply to matching words. Connecting “Maya” with “she” may therefore require coreference resolution: deciding which expressions refer to the same person.
The grid is a deliberately limited representation. It tracks participation and grammatical role, but does not by itself record every event, causal relationship, or factual claim in the document. A consistent pattern of references can occur in a text that is nevertheless incorrect or confusing.

How it works

Identify the discourse entities and their mentions in each sentence. Group mentions that refer to the same entity, then assign the appropriate grammatical role under the chosen annotation rules.
The original role-based model uses S for subject, O for object, X for another role, and a dash for absence. When an entity has multiple roles within one sentence, its construction gives priority to subject over object over other roles.
Read down the columns to obtain transitions between adjacent sentences. A two-sentence transition might be S - S, showing continued subject status, or O - S, showing an object becoming a subject. Frequencies of these patterns form features for a learned model. The treatment of entities, roles, transition lengths, and absent mentions must be specified.

Three practical examples

A writing researcher compares entity-transition patterns in an original article and a deliberately reordered version.
A summarization team examines whether a short summary introduces a participant and then refers to that participant coherently.
A document-review exercise uses a grid to reveal abrupt switches between entities that readers may struggle to follow.

Think of it this way

Think of a meeting attendance sheet that also records each participant’s role at successive stages. It shows who continues and who changes roles. The analogy stops at continuity: a grid does not hear the discussion or establish whether the participants’ statements make sense.

Worked example

Consider three invented sentences: “Maya reviewed the proposal.” “She revised it.” “The proposal impressed Maya.” Assume that “she” refers to Maya and “it” refers to the proposal.
The Maya column is S, S, O. The proposal column is O, O, S. There are two adjacent sentence pairs and two entity columns, producing four length-two transitions.
Maya contributes S - S and S - O. The proposal contributes O - O and O - S. Each transition occurs once, so each has frequency 1 / 4 = 0.25 under this simple normalization.
These values describe a pattern, not a 25% probability that the document is coherent. A trained model would use such features according to its learning objective.
This example also assumes correct reference links. If “she” were incorrectly treated as a new entity, the grid would change even though the original sentences had not changed. Entity detection and coreference errors can therefore affect downstream coherence assessments.

Illustrative case study

In a fictional editorial project, Harbor Briefs reviews generated updates about grant applications. Some updates repeatedly use “it” after introducing several possible referents.
The team creates small entity grids during diagnosis, checking the underlying mention links manually. The tables help distinguish an actual shift of subject from an incorrect reference assignment by the analysis software.
Editors keep the complete text beside each grid because grammatical continuity alone cannot establish which application was approved. The imagined project illustrates a diagnostic use; it is not evidence that grids automatically improve publication quality or outperform other coherence models.

Common misconceptions and mistakes

An entity grid is not merely a count of repeated nouns. Pronouns and differently worded mentions may belong to the same entity column.
Repeated subject status does not prove good writing. A document can repeat one entity while making contradictory statements.
The grid is also not a complete semantic representation. It omits much of what happened, why it happened, and how claims relate.
Finally, a transition frequency is a feature statistic. Treating it as an independently calibrated coherence score would misinterpret the representation.

Business use

Entity grids can support explainable experiments in document coherence, summarization, and writing analysis. Their compact structure makes some reference patterns easier to inspect than a single opaque score.
Before using them operationally, a team should evaluate mention detection, coreference, and grammatical analysis on its own document types. Reports full of tables, headings, fragments, or specialized names may require careful segmentation and annotation choices.
A useful assessment compares automated judgments with reader evaluations and concrete editing problems. The business value lies in finding actionable continuity problems, rather than assuming that maximizing a particular transition pattern will produce better prose.

Check your understanding

One entity has roles O, S, S across three sentences. Which length-two transitions does its column contribute?

Answer: O - S and S - S. There are two adjacent pairs, not three. Other entity columns contribute their own transitions to the document-level feature counts.

Related terms

Entity-based coherence
Coreference resolution
Discourse parsing
discourse entity
grammatical role
local coherence
sentence ordering
text summarization

Learn next

Learn coreference resolution first so that different mentions can be grouped reliably. Then compare entity-based coherence with discourse-relation approaches, which describe relationships such as explanation and contrast. For a practical exercise, manually construct a grid for a short paragraph and check every role against its sentence.

Sources

Primary research: Barzilay and Lapata, “Modeling Local Coherence: An Entity-Based Approach” (2008), https://aclanthology.org/J08-1001.pdf. The grid construction, grammatical-role priority, and transition-feature sections were consulted. Textbook context: Jurafsky and Martin, Speech and Language Processing, discourse coherence chapter, https://web.stanford.edu/~jurafsky/slp3/25.pdf. The entity-grid discussion was consulted. Independent review checked the original paper’s Sections 3.1–3.3, including role precedence, coreference grouping and the transition-frequency denominator.

Previous: Entity-based coherence Next: Entity integrity