Document Understanding, Measurement, and Manipulation Using Category Theory

1 Introduction

• The framework is described as modality agnostic, covering text, images, audio, video, sensor data, and mixtures, with high-level multimodal semantics also considered.The introduction notes that much of the present work uses text examples.Source: §1 · pp. 1, 3
• Orthogonalization identifies atomic, non-overlapping QAs spanning the document's derived QAs, supporting measures of its information and entropy.The introduction also describes these methods as dividing information in one or more documents into non-overlapping pieces.Source: §1 · pp. 2–3
• A lattice built from orthogonalized QAs and rhetorical structure organizes possible summaries and can be extended through exegesis to create document extensions.The introduction describes the lattice's hierarchical subsets as summaries and identifies exegesis as a way to extend the document.Source: §1 · p. 3
• Documents are represented as categories of question-answer pairs, built by extracting rhetorical structure into an abstractive DAG and decomposing its core QAs.Each abstractive-DAG node summarizes a document chunk; the resulting category includes core and decomposed QAs.Source: §1 · pp. 2–3
• The listed contributions also include methods for relating documents and rate-distortion analysis of summarization.Source: §1 · p. 3
• The introduction proposes using constraints from the framework to generate feedback that can be fed back to a model.It also lists implementations using large pretrained models.Source: §1 · p. 3

2.1 Equivalence between assertions and question-answer pairs

• A core QA is equivalent to an assertion when it can answer the same questions, although individual question-to-assertion mappings are not unique.The equivalence classes of assertions and core QAs are unique.Source: §2.1 · pp. 3–4

2.2.1 Definition

• The category’s objects are equivalence classes of QAs, and its morphisms encode question-answering ability, with composition reflecting transitivity.Identity morphisms hold because each core QA can answer its own QA.Source: §2.2.1 · p. 4

2.2.2 Some examples

• Figure 1 gives equivalent representations of three related assertions, including a symmetric form and one built by adding assertions sequentially.The figure’s right-hand representation starts from A and B, then adds C.Source: §2.2.2 · p. 4, Figure 1 · p. 5

2.3.1 Operations on QA pairs

• QA decomposition separates each pair into QA1-only, shared, and QA2-only information while preserving the information of the original pair.The decomposition is expressed through complements and intersection; the three pieces combine to an equivalent union of the original QAs.Source: §2.3.1 · p. 5
• Consistent QAs are united with logical AND, while inconsistent QAs require isolating the disagreement and combining the differing answers with OR.The consistent pieces are combined with AND in either case.Source: §2.3.1 · p. 5

2.3.2 A metric between assertions and QA pairs

• The assertion metric is symmetric, assigns distance 0 to equivalent assertions, and gives orthogonal assertions the maximum distance of 1.For QAs, the distance between pairs is defined through their core assertions.Source: §2.3.2 · p. 6

2.3.3 A metric between documents

• Comparing two documents involves merging their QA categories and merging QAs that are isomorphic.The document metric extends the assertion metric to this comparison.Source: §2.3.3 · p. 6

2.3.4 Orthogonalization

• Orthogonalization compares QAs pairwise and decomposes overlaps into equivalent or entirely orthogonal pieces, producing atomic QAs that span the original QA space.The procedure can revisit and decompose QAs already in the processed pool; the authors postulate that non-overlap persists under decomposition.Source: §2.3.4 · p. 6

3.1 Extracting rhetorical structure and the abstractive DAG

• An abstractive DAG represents rhetorical structure with one-sentence summaries of document chunks, from a whole-document root to sentence-level or partial-sentence leaves.A pretrained model selects semantically meaningful chunks and summarizes each; each DAG level provides a complete summary at a different abstraction level.Source: §3.1 · p. 7
• Abstractive summaries can make inferences implied by, but not directly stated in, the original document.The method uses large pretrained models to extract rhetorical structure and is abstractive rather than extractive.Source: §3.1 · p. 7

3.2 Creating a category of QAs from a abstractive DAG

• Each DAG node’s assertion becomes one or more core QAs; the method then decomposes them into sub-QAs and orthogonal pieces while retaining mappings to the DAG.Question-answering morphisms link decomposed QAs to their parents, and isomorphic overlapping QAs are identified.Source: §3.2 · p. 7

3.3 Lattices, summaries, and extensions

• The lattice represents possible summaries from the empty summary to the whole document, and identifies hierarchical summaries that preserve rhetorical order.In the example, a hierarchical summary preserves higher assertion A before assertions B and C.Source: §3.3 · p. 9
• All summaries decompose into quotient and subdocument summaries, and all extensions into superdocument and elaboration extensions.The paper treats exegesis as adding information to a document to create an extension.Source: §3.3 · pp. 8–9
• A subdocument summary removes entire chains, while a quotient document summary removes parts of information along chains and retains main points.Unlike subdocument summaries, quotient summaries have a map from the original document to the summary; typical summaries can combine both forms.Source: §3.3 · pp. 9–10
• An elaboration extension adds detail to existing information, whereas a superdocument extension adds new heads of chains.A typical extension can include both types.Source: §3.3 · p. 10
• Every extension can be written as a superdocument extension followed by an elaboration extension.The construction first adds new main points, then elaborates on the chains, including the new ones.Source: §3.3 · p. 10

Subdocument

• The sample summary retains the Address’s account of founding a new nation, the civil war, and dedicating a battlefield, while omitting the closing sentence.The supplied passage contains the summarized text.Source: Subdocument · p. 11

Quotient document

• The passage labeled “Summary synopsis information” presents the same excerpt as the passage labeled “Document.”Source: Quotient document · p. 11
• The passage labeled “Document” contains the original Gettysburg Address excerpt beginning “Four score and seven years ago.”Source: Quotient document · p. 11
• The extension adds details including three months, the Atlantic and Pacific oceans, prosperity, and a description of the day.It also changes the reference to the nation's founders from “fathers” to “fathers and mothers.”Source: Quotient document · p. 11

3.4 Measures of document information

• The section presents these measures as different ways to quantify document information, with usefulness depending on the purpose.Source: §3.4 · p. 12
• Information content counts a document’s atomic orthogonal QA pairs, and information density normalizes that count by document length.The length may be measured in words, pages, or another notion of document length.Source: §3.4 · p. 12
• Mutual information between two documents counts their atomic orthogonal QA pairs in common after orthogonalization, without normalization by the merged category’s total QAs.Source: §3.4 · p. 12
• Information gain from an additional document is defined using its information content minus its mutual information with the existing document.The passage identifies the two quantities used in the information-gain expression but does not provide the expression itself.Source: §3.4 · p. 12
• Content entropy is the logarithm of the number of QA chains, equivalently their heads, with each head containing the information of its chain.The measure is additive across documents: CE(D + D′) = CE(D) + CE(D′).Source: §3.4 · p. 12
• Content entropy density divides a document’s content entropy by its length, such as its number of words or pages.CEd(D) = CE(D)/|D|.Source: §3.4 · p. 13
• The combined diversity-and-depth entropy measure codes atomic orthogonal QAs and counts bits needed to transmit unique QAs.Its primary hybrid method codes all QAs independently while transmitting each unique QA only once; E0 and E1 provide lower and upper bounds.Source: §3.4 · p. 13

3.5 Constraints and feedback

• Compositionality and related category constraints can guide fine-tuning when model outputs fail to meet the required mathematical properties.The section also mentions DRESS’s reinforcement learning with AI feedback and a DPO-like approach as alternative fine-tuning avenues.Source: §3.5 · p. 13
• RLVR can use composability and closure constraints, including summary closure under intersections and unions, as automatically computable self-supervision tasks.The framework’s graph structure can generate many such constraints automatically.Source: §3.5 · p. 13
• Experimental study of whether these automatically generated constraints improve model accuracy is deferred to forthcoming work.Source: §3.5 · p. 13

3.6 The category of documents

• Document categories are objects and functors between them are morphisms, with suppression and exegesis as special examples.Source: §3.6 · p. 13
• A document-category morphism is a relation that can associate each source QA with zero, one, or multiple target QAs.Relational composition may be sub-optimal when it leaves assertions in either document unmapped despite a possible morphism containing more assertions.Source: §3.6 · p. 13

3.7 Rate distortion analysis

• Figure 3 uses word count for rate and incorrectly answered questions for distortion; dashed lines show raw points and solid lines the operational curve.The solid curve selects the lowest-distortion method at the given rate or lower.Source: Figure 3 · p. 14
• In rate-distortion analysis, rate is summary length and distortion is the fraction of document questions the summary cannot answer.Varying summary length yields an operational curve for comparing summarization schemes across rates.Source: §3.7 · p. 14

4.1 Multimodal categories

• Multimodal category nodes retain high-level semantic assertions while also pointing to low-level evidence such as text, images, video, audio, or sensor data.The section presents this as an extension of the abstractive DAG.Source: §4.1 · pp. 14–15

4.2 Probabilistic categories

• Probabilistic categories assign probabilities to scenarios in the DAG structure.Source: §4.2 · p. 15

4.3 Coherent extensions

• Document extensions add information to different parts, and coherence requires that additions across those parts do not conflict.The section relates this local-agreement idea to sheaves: information assigned to open sets forms a sheaf when it matches on overlaps.Source: §4.3 · p. 15, Figure 4 · p. 15
• The section leaves open how extensions behave for inconsistent documents and how sheaf-based methods might assess inconsistency across multiple documents.It also raises questions about whether local document information can indicate document dimension or missing information.Source: §4.3 · pp. 15–16

5.1 Departure from the state of the art

• The framework combines LLM-based assertion discretization, category representations, and a distance metric to support semantic retrieval and coherent semantic merging of documents.Preliminary summarization results are described as empirical grounding for the framework.Source: §5.1 · p. 16

5.2 Potential applications

• Structure extraction and information-theoretic measures could help writing learners inspect rhetorical structure, extraneous or missing detail, and document coherence.Source: §5.2 · p. 16
• Category merger followed by generation is proposed for combining multiple scientific papers into one coherent document.Source: §5.2 · p. 16
• Separating a document's content from its style is proposed as a way to generate versions customized to different audiences.Source: §5.2 · p. 17
• Comparing orthogonalized question-answer pairs can assess a document's alignment with principles, including whether system practices conform to harmlessness.The framework generalizes alignment assessment beyond ethical principles to any set of principles.Source: §5.2 · p. 17
• Exegesis is framed as transferring task knowledge by adjusting, removing, or adding components of a task description to form a related task description.The component-level process supports systematic transfer, unlike a monolithic representation.Source: §5.2 · p. 17

5.3 Prompt Science

• Representing prompts as categories could support prompt summarization with an accuracy-effort trade-off, as well as information measures, lattice-based summarization, and exegesis.The authors call this proposed research direction prompt science.Source: §5.3 · p. 17
• The prompt framework proposes consistency-based extension, content-style analysis, semantic prompt merging, and self-correction using compositionality.It also proposes using exegesis to refine or redirect prompts and to extend common-sense representations.Source: §5.3 · pp. 17–18

• Prior work includes empirical machine-learning approaches to question answering, lattice structures for questions, and dialogue-, inference-, or logic-based QA formulations.Source: §6 · p. 18
• The authors position their framework as formalizing document structure through large-model decomposition of natural-language assertions into QA pairs and incorporation of rhetorical structure.They also describe the QA-based distance metric as enabling orthogonalization and rate-distortion analysis of summarization.Source: §6 · p. 18
• The discussion identifies possible extensions through efficient questions, dynamic QA modeling, improved assertion-to-QA conversion, and category-based context representation.It also proposes semantic-space content entropy measures to assess information gain from one document to another.Source: §6 · p. 19