Open Exploratory Ontology Construction: Adapting Models to Evolving Research Questions
This article presents a five-step methodology for open exploratory ontology construction where research questions evolve iteratively: scope each round, separate observations from hypotheses, check existing model coverage, propose minimal candidate definitions, verify with independent materials, and hand off with documented evidence, versions, and unresolved issues.
The article introduces an open exploratory approach to ontology construction for situations where research questions change continuously, such as shifting from brand reputation analysis to celebrity topic analysis in public opinion monitoring. Instead of fixing requirements upfront, the method lets questions and models co-evolve through iterative rounds.
Step 1: Set Scope for This Round
Define a starting point without committing to a final domain goal. Agree on available materials, sources, time range, roles (who reads, who reviews new definitions), and a checkpoint to review progress. New leads may redirect the effort, but the reason for redirection and any unresolved items must be recorded.
Step 2: Examine Materials, Separate Observations from Hypotheses
Preserve provenance (source, timestamp, original text, context) before categorizing. The article illustrates with three fictional social media posts:
Account A: "I waited half an hour at Brand A's store today, terrible experience." → Direct negative expression about service; waiting experience is still Account A's claim.
Account B: "Someone said 'Brand A service is terrible', I haven't been there, no comment." → A negative statement is quoted, but the publisher explicitly disclaims evaluation.
Account C: "I like Artist B's stage arrangement this time." → Positive expression about a specific aspect (arrangement) involving a person.
Observations (what the text directly shows) are kept separate from hypotheses (e.g., whether quoting a negative statement counts as the publisher's own negative view). A new term alone does not justify adding an ontology concept.
Step 3: Pick One Question, Check Existing Expressions
Select a tractable question that advances judgment. The example chooses "opinion attribution": who said what, targeting what, based on which text span; unresolved attributions are left open. The article references Ontology Development 101 (Noy & McGuinness) for using purpose and questions to bound scope while allowing them to change during design.
Two model starting points are compared:
Model already represents brands, organizations, and persons → add the new person as an instance; check if existing relations capture specific works or performances.
Model only supports brands, with data mappings and queries hard-coded for brands → cannot simply stuff a celebrity name into "brand name"; must examine type system, relations, and dependent programs for required changes.
If no model exists, extract necessary terminology from the current materials and check reusable vocabularies. Clarify content, publisher, expression, and attribution first — no need to pre-build a full classification for all future topics.
Step 4: Organize Candidate Definitions and Relations, Make Minimal Expressions
Returning to Account B's post: the old processing only recorded "account publishes content", "content involves brand", "content contains negative sentiment", and queries attributed the negative sentiment to the publishing account. First check if the existing model can distinguish quotation (perhaps the relation exists but was not extracted or queried). Only if no expression slot exists, propose a candidate.
Candidates are kept as drafts for discussion, not finalized. The article proposes three roles: "publisher" (account that emitted the content), "expression subject" (who, per the original text, is speaking), "discussion target" (who or what the utterance is about). An account emitting content does not mean every sentence reflects its own view.
One candidate expression: record each specific expression and its attribution separately — Account B published the content; it quotes a negative claim; the quoted speaker is unknown; Account B's own wording is "I haven't been there, no comment." Each item retains its source text span.
New or modified definitions must document meaning, applicability, source material, candidate version, and unconfirmed items, then be handed to the responsible domain expert. Gaps remain on a candidate list.
Account C's post raises a further question: even if the old system never answered incorrectly, analyzing person-centric discussions may require distinguishing which performance and which aspect (e.g., "arrangement") are evaluated. Simply asking "positive or negative about this person" may no longer suffice. Check whether old definitions can separate person, performance, and evaluation aspect; if not, propose candidates to keep them distinct, leaving the performance identity as "to be verified" rather than inventing an event name.
Step 5: Verify Candidates with Materials, Decide Accept/Modify/Continue
Verification uses the original three posts: Account A's direct criticism should be preserved; Account B must not be labeled a negative reviewer merely for quoting; for Account C, check whether the candidate distinguishes the person, the referenced performance, and the evaluation aspect, retaining the unverified performance identity and not expanding "likes arrangement" into "endorses everything about the person."
Beyond the design materials, hold out unused materials to test conflicting statements, missing attributions, and confusing cases. New scope uses new materials; existing capabilities are regression-tested with shared old materials.
AI may suggest candidate relations and examples, but cannot treat its own output as ground truth. Divergent interpretations are retained with evidence; if the material cannot decide, do not force a definite subject. The article emphasizes: "This round can verify what the original text said and who the utterance belongs to. What the speaker actually thinks is not in the material — record 'don't know'."
After verification, decide: accept the draft, return for revision, reuse the original model, or gather more materials. Document scope of applicability and evidence for accepted items; for undecided items, document what is missing and who continues. Separately verify that extraction programs follow the model and that claims in materials are independently corroborated — model correctness and query accuracy do not substitute for these two checks.
A minimal runnable example was manually checked with structured inputs for the three posts' partial expressions and query behaviors. It did not automatically read original texts, did not validate human annotations, and did not fully model specific performances or fine-grained evaluation aspects — therefore it cannot be claimed that the public opinion system "understands" these utterances.
If topic, materials, and sentiment analysis method all change simultaneously, record each change separately; improvement in results cannot be automatically credited to the ontology modification.
Step 6: Close This Round, Decide Next Round
Materials can keep accumulating, but a time-boxed checkpoint forces the team to lay out evidence, current judgments, and remaining ambiguous interpretations. The example records: whether the candidate preserved the publish/quotation distinction; whether the existing model sufficed; which checks passed and which did not. If only data is lacking, note "currently undecidable" with reasons. The round can pause with the original question still open.
Next round might shift to counting content volumes across themes — deciding whether to count raw posts, reposts, deduplicated discussions, and how to handle multi-theme content. This may only affect query and analysis configuration, not shared concepts.
Another example: two posts titled "Artist B Responds to Controversy" lack date, location, or context to distinguish events. Record both as involving the same person; event identity remains "to be verified". Next step is to gather evidence, not to merge by adding a relation.
New questions are worth pursuing on their own merit, not only to fix old system errors. Selection criteria: what judgment does clarifying it enable? Are materials sufficient? Who will judge? Rough time estimate? If key data or ownership is missing, pause — no need to modify the ontology every round just to show progress.
When switching from brand to celebrity topics, record why the shift occurred and how far the original problem was resolved. Existing conclusions stand; new questions get their own scope; next round's findings must not masquerade as completion of the previous round.
Unified Exploration Record Connecting the Five Steps
The article adapts the "construction card" from the previous installment, adding fields needed for exploration. A summary table (converted to list below) specifies the minimum artifacts for each phase:
Round Starting Point: Research direction, material scope, access rights, time and resource allocation; who judges, who takes over. Example: enter from brand and person discussions, agree sources and round scope; real projects fill in specific people and review dates.
Examine Materials: Original text, provenance, context, observations, hypotheses, pending questions. Example: retain three fictional posts; distinguish direct expression, quotation, and aspect-specific evaluation; experiences not independently verified.
Select Question, Check Existing Expressions: Current question, model and implementation version, reusable parts and gaps. Example: first verify opinion attribution; if quotation relation exists check implementation, else propose candidate.
Build Candidates: Meaning, applicability, relations, mapping method, material evidence, unconfirmed items. Example: separate publisher from expression subject; new question checks how person, performance, and evaluation aspect relate; candidates handed to relevant personnel.
Verify and Decide: Check materials, expected basis, actual results, confirmation records, failed and unexecuted items. Example: verify direct evaluation vs quotation and distinctions needed for new question; program only checked part of manually structured inputs; performance and evaluation aspects not fully modeled; automatic extraction not validated.
Round Closure and Next Round: Evidence-backed understandings, unknowns, adoption/modification/retention/deferral decision, pivot reasons and continuation conditions. Example: record accept, modify, retain, or defer; event identity and theme counting become subsequent questions, not yet implemented.
The five phases can iterate back and forth; candidates may be accepted, modified, retained, or deferred. The next round builds on current understandings and remaining questions — not every round must modify the ontology.
No need to explain the entire public opinion domain before closing a round. Document how far the selected question was pursued, why it remains undecided, along with materials, versions, limitations, and handover owner. After research handover, application integration still requires verifying that programs actually use the model.
When the next round changes topic, revisit this record: which definitions carry over, which judgments have evidence, which are still unclear. Thus questions continue to expand while the team knows whether each step advances on solid ground or remains speculative.
The next article returns to the equipment case to elaborate the modeling phase in both construction flows: how to cross-check business goals, reality constraints, and execution arrangements together to reduce omissions.
Reference
Natalya F. Noy, Deborah L. McGuinness, Ontology Development 101: A Guide to Creating Your First Ontology , Section 3 and Step 1. Cited for iterative development, revisable purpose and scope, and evaluation of drafts via applications or expert discussion. The five phases, public opinion materials, candidate expressions, phase exit criteria, and exploration record are the author's design, not a prescribed public opinion analysis method from the guide. URL: https://protege.stanford.edu/publications/ontology_development/ontology101-noy-mcguinness.html
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
