PaperlayerView exact revision on arXiv

Source-linked AI summary

5W1H+Which: Context-Valid Semantic Indexing with Progressive Ontology Binding

Yaxiao Liu, Pengbo Liu, Yiwen Liu, Yihua Guan, Jiaxing Song

arXiv:2609.35184v2cs.AIcs.CLcs.IR

TL;DR

Semantic indexing must preserve source information for reuse without prematurely committing to one business schema, while providing checkable premises for reasoning. The paper proposes 5W1H+Which to separate content extraction from progressive, context-constrained ontology binding and maintenance. No experiments have been conducted; the paper identifies matched comparisons and tests as the next step.

  • Problem

    An index must preserve facts for unknown future questions without hiding uses through task-specific representations or omitting context needed to judge validity.

    Source: §1 · p. 2

  • Method

    The proposed design organizes source-grounded content separately from contextualized ontology bindings and tracks dependencies for maintenance and revalidation.

    Source: §3.1 · p. 6, §11 · p. 27

  • Results

    No experiments have been conducted; the paper identifies fair comparisons and tests of cross-task evidence coverage, contextual misuse, and update cost as immediate priorities.

    Source: §11 · pp. 27–28

  • Takeaways & Limitations

    The proposed contract is intended to preserve unbound content while making bound information usable under explicit conditions and tracing consequences when conditions change.

    Source: §11 · p. 27

    Extraction, identity resolution, and ontology mapping can fail, and evidence links do not prove source evidence true, so consequential decisions require risk-proportionate verification and human review.

    Source: §10 · p. 27

Abstract

from arXiv · show

Semantic indexing must preserve reusable source information without committing too early to one business schema, while providing checkable premises for reasoning. We propose 5W1H+Which, a design that separates source-grounded content extraction from progressive ontology binding. The 5W1H questions organize independently addressable content units; Which records links to versioned ontology elements, mapping relations, applicability scopes, and validation states. Time, location, system environment, and participant roles constrain when facts, bindings, and rules may be used. Unbound content remains searchable, whereas formal reasoning requires explicit premise checks. The design distinguishes business valid time, system knowledge time, and operational traces, and tracks dependencies for revalidation and incremental maintenance. Worked examples illustrate cloud migration and temporal policy validity. The resulting linked corpus can support downstream training and evaluation organized around benchmarks, positive and negative examples, trajectories, and costs. Trajectories preserve steps, evidence-backed checkpoints, recovery, and local-to-global failure impact; reusable skills remain downstream abstractions requiring cross-episode validation. We specify falsifiable hypotheses and capability-matched baselines to test cross-task evidence coverage, contextual misuse, and update cost under equal information and budgets. The contribution is a proposed indexing and maintenance contract, not a new universal ontology or post-training algorithm. No experiments have yet been conducted, and no performance improvement is claimed.

1 Introduction

An index must retain information for future uses without tying it too closely to a current task or business ontology. The proposal separates source-grounded content from revisable ontological interpretation and constrains use by context.

1 Introduction

  • Task-specific indexing can hide other uses, while overly broad descriptions may omit the units, roles, negation, versions, and scope needed for valid answers.Ontology mappings also require evidence, judgment, and maintenance rather than being inherently correct.Source: §1 · p. 2
  • 5W1H organizes what a source expresses, while Which provides an interface for binding that content to an ontological interpretation.The questions are shared directions across domains, not a universal field set; precise measurements and domain structures still need typed payloads or extensions.Source: §1 · p. 2
  • A planned cloud migration is not a deployment fact, and changing a label after verified migration must not erase its historical applicability.Overwriting damages historical queries, while leaving a binding unchanged can apply old rules to a new environment.Source: §1 · p. 2
  • The proposal makes content units, bindings, and derived conclusions independently maintainable, preserves unbound content for queries, and specifies contextual constraints and change propagation.It also proposes controlled experiments to test whether this separation justifies its additional cost.Source: §1 · p. 2
  • The longer-term objective is to organize addressable content and evidenced relations by task and goal version for downstream agent adaptation, not to claim a validated fourth contribution.The paper identifies post-training, prompt and skill improvement, tool routing, and workflow repair as downstream uses.Source: §1 · p. 2

2 Related Work and Research Position

The paper positions its contribution as a testable combination of retained source content, contextual bindings, premise checks, and update maintenance—not as novelty in any one ingredient. It calls for capability-matched evaluation rather than claims based on terminology or simplified comparisons.

2.1 5W1H Extraction and Event Modeling

  • 5W1H-guided extraction and 5W1H event ontologies already exist, so the open question is whether retained addressable content, revisable bindings, and explicit contexts yield measurable benefits.The paper treats this distinction as a research position to test, not as proof that prior systems lack those capabilities.Source: §2.1 · p. 3

2.2 Dynamic Ontologies and Constrained Mapping

  • Dynamic ontology and mapping work provides close comparators; the proposed distinction is that source content remains available independently of successful mapping while bindings retain scope and lifecycle.If a dynamic ontology system already achieves equivalent results through those mechanisms, the contribution must be narrowed.Source: §2.2 · p. 3

2.3 Context, Time, and Provenance

  • Context, provenance, and temporal standards do not by themselves establish business-fact truth or implement application-specific bitemporal querying and access control.The proposed evaluation concerns whether content retention, contextualized binding, premise admission, and update maintenance improve reuse and reliability within a specified budget.Source: §2.3 · p. 3

2.4 Auditable Queries and Hypothetical Reasoning

  • Auditable-query and hypothetical-reasoning work motivates isolated scenario branches, while experiments should use strong controls with relevant capabilities rather than only keyword search versus the full system.Simulated what-if outputs should not automatically enter the factual store, and verbal explanations alone do not establish causal relationships.Source: §2.4 · p. 4

2.5 Agent Memory: Dimensions, Lifecycle, and Use-Time Enforcement

  • Related memory work already covers dimensional retrieval, reusable rules, implicit applicability, and verified updates; these precedents motivate separating stored state, evidence selection, and use.Prior results and settings require careful interpretation: a nonsignificant correctness difference does not establish equivalence, and controlled results do not establish enterprise deployment outcomes.Source: §2.5 · p. 4
  • The central memory question is whether independently retained content, contextual bindings, admission checks, and update dependencies improve cross-task reuse at matched cost.An equivalent result from a capability-matched baseline would narrow the contribution to an implementation or evaluation protocol.Source: §2.5 · p. 5

2.6 Stateful Testing and Compositional Security

The related-work discussion frames stateful testing and maintainable corpora as complementary prior areas, then treats downstream training and evaluation assets as applications rather than validated capabilities. The proposed advantages remain questions for matched-budget tests.

2.6 Stateful Testing and Compositional Security

  • Stateful API exploration, protocol fuzzing, and compositional policy testing are established related work, so the proposed test asks whether reusable evidence and context-valid transitions improve findings and evidence reuse under matched budgets.Strong prior implementations may eliminate the proposed advantage.Source: §2.6 · p. 5

2.7 Maintainable Corpora and Task-Adaptive Views

  • Task-adaptive evidence selection and maintaining valid state after evidence changes are complementary axes; the proposed separation of content, bindings, and view dependencies is a testable design choice at their intersection.Evaluation should compare against baselines with equivalent provenance, versioning, selection, and maintenance capabilities.Source: §2.7 · pp. 5–6

2.8 From Memory Material to Training and Evaluation Assets

  • The cited work motivates connected benchmark, example, trajectory, and cost views, but does not establish that these pillars are exhaustive or validate this indexing design.An indexed corpus may provide material for distinct forms of distillation, but improved external-memory access alone does not establish learned capability persistence without that access.Source: §2.8 · p. 6

3 Problem Definition and Design Boundaries

The section defines an index that keeps source content addressable while representing query context and reasoning premises explicitly. Its 5W1H organization does not require every field to be present, and formal conclusions remain conditional on declared premises and semantics.

3.1 Inputs, Index, and Queries

  • The indexer maps versioned sources with addressable evidence spans into content, binding, rule, and derived-record sets; multimodal span addressing and extraction quality remain outside scope.Sources may include text, structured records, or operational traces.Source: §3.1 · p. 6
  • Queries include business time, knowledge cutoff, typed context, goal version, access principal, and response mode; a source assertion alone does not establish verification.Modes distinguish source lookup, business-state assessment, and hypothetical reasoning.Source: §3.1 · p. 6
  • Historical queries distinguish what was warranted by information then known from what current knowledge says happened in the past, while both enforce current access authorization.Late evidence may revise the latter answer but must not be silently added to the former.Source: §3.1 · p. 6

3.2 The Content Layer Is Not Six Mandatory Strings

  • 5W1H organizes extraction and browsing without requiring six filled fields; unstated, inapplicable, unextracted, and permission-hidden values represent different states.An unstated property indicates missing knowledge, not that the business property is absent.Source: §3.2 · pp. 6–7
  • Compound narratives should be split into claims so values, units, negation, conditions, quantifiers, and modality remain attached to the right evidence.Document-level summaries can aid entry but cannot replace claim-level addressing.Source: §3.2 · p. 7

3.3 Probabilistic Extraction and Conditional Formal Reasoning

  • LLMs may propose uncertain content and mappings, but processing and evidential status—not a fixed probabilistic-versus-deterministic division—distinguish 5W1H from Which.A 5W1H unit can record a verified measurement, while a Which candidate can be incorrectly mapped.Source: §3.3 · p. 7
  • Formal reasoning derives conclusions relative to declared ontology versions, rule languages, and premises, without guaranteeing input truth or completeness.An LLM self-reported score is a ranking signal, not a calibrated probability of truth without validation.Source: §3.3 · p. 7

4 Representation: Content, Bindings, and Derived Results

The representation separates source-grounded content, ontology bindings, contextual applicability, and derived conclusions into maintainable records. This preserves unbound content while making binding lifecycle, changes, and operational evidence explicit.

4 Representation: Content, Bindings, and Derived Results

  • Figure 1 distinguishes source exploration, formal reasoning, and hypothetical reasoning; its arrows show permitted data flow, not an implemented safety guarantee.Reads require current authorization, and outputs or tool actions require a fresh policy check.Source: Figure 1 · p. 7

4.1 Source-Grounded Content Units

  • Content units retain 5W1H organization, typed payloads, evidence, epistemic and validation status, and separate business-valid and system-visible intervals.Evidence can identify source version, span, and access policy; unknown fields remain unknown, while typed payloads extend the content directions.Source: §4.1 · p. 8, Table 2 · p. 8

4.2 Which: Explicit Ontology Binding

  • Which bindings connect content to a specific ontology version and element with a mapping type, applicability context, evidence, and validation record; relation types are not interchangeable.The manuscript keeps source references, contextual associations, process links, and conclusion dependencies outside Which.Source: §4.2 · pp. 8–9
  • When no suitable binding exists, content retains an empty binding; candidate, validated, revalidation, rejected, and retired states distinguish its lifecycle.Only a possible concept gap—not insufficient evidence, ambiguity, irrelevance, or mapper error—is a candidate for ontology extension.Source: §4.2 · p. 9

4.3 Context Space and Three Levels of Change

  • Context uses typed coordinates, and maintenance distinguishes world-state changes, interpretation changes, ontology-schema changes, and independent rule-applicability changes.A deployment change does not by itself imply deleting an ontology class.Source: §4.3 · p. 9

4.4 Derived Records and Support Sets

  • Derived records retain support paths across content, binding, and rule versions; loss of one path does not necessarily falsify a conclusion.Withdrawal of evidence is not treated as proof of the opposite, and conflicts enter an explicit conflict set.Source: §4.4 · p. 9

4.5 Optional Transition and Trajectory Records

  • Transition records link hypotheses, actions, observed effects, and evidence into trajectories, but observed responses are not hidden state or proof of a complete minimal machine.Wall-clock proximity alone does not establish causality or total order in distributed systems.Source: §4.5 · p. 10
  • Behavioral models describe what a system did, whereas separately approved policies specify what it should permit; policy provenance and enforcement details require separate records.Learning behavior and its normative oracle from the same unreviewed trace risks normalizing an existing defect.Source: §4.5 · p. 10

5 Method: Progressive Binding and Context-Constrained Queries

The method preserves source-grounded content independently of ontology bindings, admits premises only after explicit checks, and proposes support-aware updates with conservative fallbacks. It also separates exploratory retrieval and hypothetical reasoning from established factual conclusions.

5.1 Content Extraction and Retention

  • Extraction retains addressable claims, unknown fields, and source-supported but unbound content, while declaring span-selection and storage budgets rather than promising zero information loss.High-risk claims require authorized external records or human review; failure analysis distinguishes extraction, binding, and retrieval errors.Source: §5.1 · p. 10

5.2 Which Candidate Generation and Validation

  • Which candidates are checked for identity, relation direction, units, negation, modality, target version, and scope; name similarity alone does not establish equivalence.Bindings retain evidence, checks, failure reasons, and revalidation triggers, while ranking remains distinct from task-specific eligibility.Source: §5.2 · pp. 10–11

5.3 Premise Admission and Formal Reasoning

  • Formal premise admission requires evidence, time, scope, context, binding, rule, and authorization checks; unknown required checks block reasoning without making data nonexistent.Source assertions and business facts remain distinct, and joins must check shared context rather than assume persistence, identity merging, or unit conversion.Source: §5.3 · p. 11

5.4 What-else: Finding Omitted Content and Interpretations

  • What-else separates omitted evidence, new relation candidates, and possible concepts, while leaving unbound content available for later questions and verification.Finding an unbound term alone does not establish a new concept; accepted changes create a new version and justification.Source: §5.4 · p. 12

5.5 What-if: Isolated Hypothetical Branches

  • What-if reasoning uses a fixed base snapshot and scoped assumptions, labeling conclusions that depend on those assumptions without changing the factual record.A hypothetical migration does not prove migration or responsibility transfer; causal claims require a separately specified causal model and assumptions.Source: §5.5 · p. 12

5.6 Change Propagation and Conservative Fallback

  • Change propagation versions changes, revalidates bindings, reassesses support, derives new conclusions, handles recursion, falls back to recomputation, and publishes snapshots atomically.If dependencies are incomplete or scope cannot be bounded, the procedure recomputes a relevant context slice or the finite knowledge base.Source: §5.6 · p. 12
  • Incremental optimization is conditional on agreement with full recomputation under supported assumptions, and access revocation instead requires immediate blocking rather than delayed maintenance.Historical retention remains subject to authorization and deletion policy; the initial implementation should establish full recomputation and differential tests before optimizing.Source: §5.6 · p. 13, Figure 2 · p. 13

5.7 Maintenance Budget and Temporary Non-Use

  • The maintenance ledger charges extraction, mapping, source rechecks, dependency repair, storage, and human review; tool receipts support only scoped evidence and also incur validation costs.Learned recheck scheduling is a later extension, not a claimed contribution of the draft.Source: §5.7 · pp. 13–14

6 Constructed Examples: Changing Deployments and Temporal Policies

Constructed examples show how business-valid time, knowledge cutoffs, and evidence type constrain deployment queries. A delayed-execution security trace likewise remains a candidate until contextual checks and validation controls support a conclusion.

6.1 From On-Premises to Cloud Deployment

  • In the constructed migration case, a plan does not establish completion, and later-acquired evidence can support a past business-state query only at a later knowledge cutoff.The example concerns only service S’s production environment; dates and spacing are illustrative, not a measured timeline.Source: §6.1 · p. 14, Table 3 · p. 14, Figure 3 · p. 15

6.2 Revocation Before Delayed Execution

  • In the constructed revocation case, enqueue, revocation, and later execution yield only a candidate violation unless identity, resource, policy, timing, and delegation conditions are established.Reset-and-replay and negative controls adjudicate candidates; unknown ordering, identity, or policy oracle leaves the result inconclusive.Source: §6.2 · pp. 15–16, Table 4 · p. 16, Figure 4 · p. 17

7 Falsifiable Research Hypotheses

The paper proposes three falsifiable hypotheses about deferred binding, contextual validity, and dependency-based maintenance. Each is paired with conditions or outcomes that would narrow the claim rather than establish universal superiority.

H1: Deferred Binding Improves Cross-Task Evidence Coverage

  • Deferred binding is expected to improve coverage of source-supported evidence needed by new tasks, but that difference may disappear with complete ontologies or strong content-retaining baselines.The proposed measure is task-required evidence, not node count; any benefit may arise from avoiding premature information discard rather than the 5W1H labels.Source: H1: Deferred Binding Improves… · p. 17

H2: Contextualized Binding Reduces Misuse

  • Contextualized valid time, knowledge time, and scope are expected to reduce stale-label errors, cross-environment misuse, and treating plans as facts at matched coverage and evidence.A typed fact graph with equivalent information and constraints achieving the same result would locate the benefit in general context management.Source: H2: Contextualized Binding Reduces… · p. 17

H3: Dependency Maintenance Reduces the Cost of Correct Updates

  • Dependency maintenance is expected to reduce revalidation and recomputation for local changes only when updated results agree with full recomputation and both withdrawals and additions are tested.Benefits may disappear for ontology-wide changes, dense dependencies, or broad authorization changes; performance claims must specify change types and scales.Source: H3: Dependency Maintenance Reduces the… · p. 18

8 Experimental Design

The proposed evaluation pairs varied data and tasks with capability-matched baselines, controlled comparisons, and separate coverage, reliability, update, and cost measures. No datasets or comparisons have yet been completed; an optional security study has its own controls and costs.

8.1 Data and Tasks

  • The planned evaluation uses synthetic scenarios and licensed public records in IT operations and manufacturing, with task splits that freeze the index before test questions are revealed.Synthetic outcomes are reported separately and cannot substitute for real-world business validity.Source: §8.1 · p. 18

8.2 Strong Baselines and Ablations

  • Comparators include rich summaries, static and dynamic ontology indexes, and typed fact graphs; controlled ablations isolate extraction, organization, binding, and contextual admission.Methods share model versions, evidence permissions, rule expressivity, and questions, while accounting separately for engineering and human-check costs.Source: §8.2 · pp. 18–19, Table 5 · p. 19

8.3 Metrics, Statistics, and Failure Interpretation

  • Metrics cover required-evidence recall, precision, qualifier retention, binding correctness, contextual misuse, answer coverage, update outcomes, and total costs; the proposed datasets and comparisons remain unrun.The design separates stored state, evidence views, agent use, and outcomes, and calls for grouped confidence intervals and frozen test criteria.Source: §8.3 · pp. 19–20

8.4 Optional Security Substudy

  • The optional authorized security study tests context-valid joins and counterexample-driven state revision against matched trajectory-graph baselines, charging exploration, replay, and review costs.It distinguishes policy violations from false positives and inconclusive outcomes; more alerts, generated tests, or refusals alone do not count as success.Source: §8.4 · p. 20

9 Extension: An Indexed Corpus for Training and Evaluation

Section 9 treats the indexed corpus as a reusable data foundation for task-specific training and evaluation views, then describes versioned process evidence and proposed downstream uses. It emphasizes that these are design proposals, not demonstrated performance gains.

9.1 Separating the Data Foundation from Downstream Systems

  • The corpus retains evidence, context, binding states, and cross-record relations; task-level views assemble benchmark, example, trajectory, and cost material for a specific task and goal version.These views are reproducible selections rather than new sources of truth, and unbound material remains discoverable.Source: §9.1 · pp. 20–21, Figure 5 · p. 21

9.2 Four Pillars, Not One-to-One Field Assignments

  • Pillar views require independently justified benchmarks, oracle-conditional labels, and deduplicated cost ledgers; calls, tokens, monetary cost, latency, and resource use remain separate.Benchmarks distinguish terminal outcomes from proxies, while missing evidence yields an unknown label rather than an automatic negative.Source: §9.2 · pp. 22–23

9.3 Process Evidence, Goal Revision, and View Maintenance

  • Episodes distinguish valid time, knowledge time, event order, and goal versions; source, goal, binding, or criteria changes can add or withdraw view material.Linked evidence carries scope and version, but ontology validation does not establish event identity, truth, authorization, success, or measured cost.Source: §9.3 · p. 23

9.3.1 Trajectory as a Versioned Process-Evidence Graph

  • A trajectory is an episode-specific process-evidence graph linking attempted steps, checkpoint assessments, recovery, and verdicts under task, goal, and observation versions.Checkpoint criteria, applicability, criticality, assessment, and evidence are recorded separately; success at one decision layer does not imply success at another.Source: §9.3.1 · p. 23, Table 7 · p. 24

9.3.2 Local Failure, Recovery, and Overall Impact

  • Failure consequences depend on the acceptance policy: retries may recover a failed attempt, optional failures may leave outcomes intact, and non-compensable violations may block completion.Recovery adds a new assessment without erasing the original failure or cost; dependency reachability alone does not prove causation.Source: §9.3.2 · p. 24, Figure 6 · p. 25

9.4 Constructed Example: A Migration Task

  • The constructed migration task separates plans, tool receipts, acceptance specifications, independent verification, and usage records as distinct evidence for its four views.A verified retry can satisfy a restore checkpoint while preserving the initial failure and both costs; missing integrity evidence leaves the verdict unknown unless another verified failure determines it.Source: §9.4 · pp. 25–26

9.5 Supported Optimization and Evaluation Boundaries

  • The corpus may support downstream demonstrations, analysis, optimization, and evaluation, but training algorithms and policies remain separate choices and candidate skills require validation across held-out episodes.Failures can inform diagnosis or negative examples without becoming successful demonstrations; a single successful trace does not establish skill generalization.Source: §9.5 · p. 26
  • Evaluation should first test evidence coverage, pillar reuse, and view maintenance against field-equivalent baselines, with lineage-based splits and a separate downstream study of performance and total cost.The proposed tests report no training gains.Source: §9.5 · pp. 26–27

10 Limitations, Risks, and Ethics

The paper identifies representational, reliability, implementation, and ethics risks that constrain how the proposed index can be used and evaluated.

Limitations, Risks, and Ethics

  • 5W1H dimensions are not inherently orthogonal, and typed payloads and domain extensions remain necessary; no finite index guarantees support for arbitrary future questions.Source: §10 · p. 27
  • Extraction, identity resolution, and ontology mapping can fail; evidence links aid auditability but do not prove truth, so consequential decisions require risk-proportional verification and review.Source: §10 · p. 27
  • Separating content, bindings, and dependencies adds storage and implementation complexity, while richer indexes may increase irrelevant retrieval and user burden.The proposal may amount to an implementation convention if strong baselines match its outcomes at the same budget.Source: §10 · p. 27
  • User behavior logs can contain sensitive information, so collection and processing require authorization, purpose limitation, minimization, access isolation, and deletion propagation.De-identification does not eliminate re-identification risk.Source: §10 · p. 27
  • Deleting an index entry does not unlearn information already absorbed into model weights, requiring separate training-lineage and model-remediation processes.Evaluating a system against labels extracted from that same system risks circular validation.Source: §10 · p. 27

11 Conclusion

The conclusion proposes 5W1H for evidenced content, Which for revisable ontological interpretations, and contextual validity for governing queries and updates. It presents a reusable corpus and an evaluation agenda, while noting that experiments have not yet established the proposal’s contribution.

Conclusion

  • The design aims to preserve unbound information, condition bound information on explicit contexts, and trace consequences when those conditions change.Source: §11 · p. 27
  • The longer-term product is a reusable indexed corpus connecting benchmarks, examples, trajectories, and costs rather than four collections of isolated facts.Task-level evidence views can serve post-training and non-parametric optimization while retaining separate validation and authorization responsibilities.Source: §11 · p. 27
  • The manuscript proposes tests of cross-task evidence coverage, contextual misuse, and update cost against a fact graph containing the same information; no experiments have yet been conducted.View reuse and maintenance are also slated for testing before assessing whether the mechanism merits an independent contribution.Source: §11 · pp. 27–28

Related papers

Loading 2609.35184v2…