Documentation

# Radiology Transcript Linguistic and Semantic Refinement Specification ## Status This document is an example-driven draft. It is a lower-level refinement of the project-level `RADIOLOGY-TRANSCRIPT-INTERPRETATION-CONCEPTUAL-SPEC.md`. It assumes the concepts, boundaries, and invariants established by that specification. If the two documents conflict, the project-level conceptual specification is authoritative. This document begins to define the linguistic annotation profile, initial construction vocabulary, semantic frame vocabulary, intermediate representation, and source-to-semantics translation obligations needed for an implementation. It is not yet an implementation plan. Names and structures marked **provisional** are hypotheses to test against the corpus before they become normative interfaces. --- # Table of Contents 1. Vision 2. Technical Introduction 3. Survey of Decisions and Translation Rules 4. Reference Appendices --- # 1. Vision ## 1.1 Purpose of This Refinement The conceptual specification says what a valid transcript interpretation means. This refinement describes how representative radiology language is decomposed into those concepts closely enough to guide implementation. The refinement proceeds from examples rather than from an attempt to enumerate English grammar in advance: ```text representative transcript expression ↓ source-anchored linguistic observations ↓ recognized construction and bound elements ↓ candidate discourse and frame interpretation ↓ domain grounding candidates ↓ validated semantic result, ambiguity, or diagnostic ``` The governing principle is: > Define only as much linguistic and semantic machinery as is required to > explain representative corpus examples without losing source evidence or > inventing transcript-level facts. ## 1.2 Relationship to the Conceptual Specification This document does not redefine `Source`, `Mention`, `Construct`, `Construction Element`, `Referent`, `Frame`, `Composition`, `Grounding`, `Ambiguity`, `Provenance`, `Diagnostic`, or `Validation`. It refines them by specifying: - which corpus expressions initially exercise them; - which linguistic evidence may support them; - which construction and frame roles are required; - how objects are represented in an implementation-facing IR; - which external authority informs each decision; - which mappings are normative, provisional, or deliberately unresolved. The conceptual transformation order remains authoritative: ```text R0 Source Transcript ↓ R1 Linguistically Annotated Source ↓ R2 Construction-Annotated Source ↓ R3 Discourse and Frame Model ↓ R4 Domain-Grounded Semantic Model ↓ R5 Validated Structured Output ``` These are semantic representation levels. An implementation is not required to materialize five files, make five network calls, or execute five isolated processes. ## 1.3 Initial Empirical Boundary This refinement uses only the locally retained generated transcription inputs. The reports from which those inputs were historically derived are not retained and are not part of an interpretation execution. The initial corpus is sufficient for: - discovering recurring language forms; - proposing construction and frame vocabularies; - creating executable interpretation examples; - testing source preservation, ambiguity, diagnostics, and validation; - demonstrating architectural viability. It is not sufficient for claiming production performance on real human dictation or speech-recognition output. Real dictation may introduce repairs, hesitations, punctuation commands, speaker-specific shorthand, abandoned clauses, recognition substitutions, and other forms that are not adequately represented by the current corpus. Those phenomena will extend this profile later; they must not silently change the meaning of existing annotations. ## 1.4 Non-Goals This draft does not attempt to: - define a complete grammar of radiology dictation; - annotate all 3,573 generated inputs before implementation begins; - treat heuristic corpus labels as gold annotations; - decide every possible FrameNet or MoCCA alignment; - infer clinical truth beyond the transcript; - validate robustness to real human transcription; - prescribe service boundaries, programming languages, or deployment topology. --- # 2. Technical Introduction ## 2.1 Corpus Inventory and Evidential Status The checked-in corpus contains 3,573 generated transcription inputs in JSON Lines form. The records use two generated surface variants: | Variant | Records | Meaning | | --- | ---: | --- | | `sectionless_dictation` | 1,982 | A deterministic removal of selected section labels. | | `spoken_punctuation` | 1,591 | A deterministic replacement of selected punctuation with comma-like forms. | The current artifact was produced by the historical derivation procedure in `research/radiology-corpus/generated/generate_transcript_corpus.py`. It is a deterministically generated corpus, not an LLM-generated corpus. The upstream source material has been removed by project decision. Its historical hash is retained as provenance, but its content is unavailable to the interpreter. Future LLM-generated inputs may be added as a distinct provenance tier. Such records should identify the source report, model, model version, prompt or prompt version, generation parameters, generation date, and any human review. The current `phenomena` and `expected_constructs` fields are regular-expression selection aids. They are neither linguistic analysis nor gold semantic annotation. In particular, the current heuristic label `ambiguous_attachment` means only that a measurement and selected spatial language co-occur; it does not establish actual ambiguity. ## 2.2 Corpus Roles The project distinguishes two corpus roles: ```text DEVELOPMENT INPUT CORPUS ───────────────────────────────────── Generated transcription records, including deterministic variants now and possible LLM-generated transcriptions later. Used as executable interpreter input. VALIDATION CORPUS ───────────────────────────────────── Future real human transcriptions with appropriate authority and handling. Used to evaluate generalization to actual dictation. ``` The input boundary is exact: ```text source(execution) = record.transcript_text ``` Only characters in `transcript_text` may supply linguistic evidence. Historical lineage fields, generation metadata, and unavailable upstream material may be used for provenance or test selection, but may not influence interpretation. Generated text may contain words that resemble report headings or other report structure. They are ordinary source tokens unless the transcription itself licenses a spoken discourse function. The system has no hidden Findings or Impression sections. Evaluation claims must identify whether their evidence comes from generated or future human transcription. ## 2.3 Example Selection The first annotated development set should be small and stratified. It should contain the simplest positive example and progressively add one linguistic complication at a time. The initial selection should cover at least: | Family | Representative form | | --- | --- | | Finding description | `Borderline cardiomegaly.` | | Normality | `The lungs are clear.` | | Negation | `No pneumothorax.` | | Negated coordination | `There is no pneumothorax or pleural effusion.` | | Location | `opacities in the left lung apex` | | Finding size | `an 8mm nodule in the left lower lobe` | | Device distance | `tube tip is 5 cm above the carina` | | Comparison | `stable from prior radiographs` | | Uncertainty | `could represent a cavitary lesion` | | Alternative interpretation | `probably scarring ... difficult to exclude a cavitary lesion` | | Recommendation | `CT chest with contrast is recommended` | | Cross-sentence identity | `There is a nodule. It measures 6 mm.` | | Repeated reference | a later utterance that redescribes an earlier finding | | Redaction damage | a construction containing `[REDACTED]` | Heuristic labels may retrieve candidates. A human must inspect each selected span before it becomes a normative example. ## 2.4 Annotation Units ### 2.4.1 Document and source span A document is one interpreter input. A source span is a half-open character interval in the exact input text: ```text span = [start, end) ``` The source is exactly the record's `transcript_text` value. Neither the record's historical `source_row` nor any pre-generation document is part of the source. The stored text for the span must equal the corresponding substring of the immutable source. Tokens and sentences may be associated with spans, but they do not replace character anchoring. ### 2.4.2 Linguistic observation A linguistic observation records imported or derived language analysis such as tokenization, lemma, part of speech, morphology, dependency relation, sentence boundary, or candidate normalization. Universal Dependencies is the initial authority for morphosyntactic labels. Parser output is evidence, not unquestionable fact. A later rule may disagree with a parser analysis only by recording an alternative observation or a diagnostic; it must not silently mutate the imported analysis. ### 2.4.3 Mention A mention is a source-anchored expression that participates in interpretation. Initial mention categories are **provisional annotation conveniences**, not new ontological kinds: ```text entity expression anatomical expression quantity expression unit expression state expression relation trigger discourse expression action expression ``` A mention may have more than one applicable category. Mention categorization does not establish discourse identity, ontology identity, polarity, or patient- world existence. ### 2.4.4 Construct and construction element A construct is a recognized form–meaning pattern over source-backed linguistic material. Every construct binds named construction elements licensed by its construction type. UCxn supplies the annotation model for construction instances and elements. MoCCA may classify or align a construction when an appropriate comparative concept exists. A radiology-specific construction remains locally defined when no external category expresses the needed distinction. ### 2.4.5 Referent A referent records discourse identity. Mentions can introduce, redescribe, or refer back to a referent. A referent under the scope of negation or uncertainty remains a discourse object, not a positive assertion of a patient-world entity. Assertion status is represented by a frame rather than encoded into referent identity. ### 2.4.6 Frame A frame represents a semantic situation with named roles. FrameNet is a source of candidate general frames and role vocabularies. The initial internal frame profile is a small hybrid: it may reuse FrameNet alignments, but local frames are normative when radiology examples require distinctions not cleanly supplied by FrameNet. ### 2.4.7 Grounding Grounding relates a mention or referent to one or more RadLex concept candidates. RadLex governs medical concept identity and ontology context. It does not govern source spans, discourse identity, assertion status, or the creation of transcript referents. ### 2.4.8 Interpretation alternative An interpretation alternative is a complete or partial candidate structure supported by evidence. Alternatives may differ in attachment, referent identity, frame type, role filler, grounding, scope, or semantic projection. An ambiguity groups supported alternatives when current rules do not justify a single choice. ## 2.5 Authority Map | Question | Primary authority | Local responsibility | | --- | --- | --- | | What morphosyntactic relation was observed? | Universal Dependencies | Preserve parser evidence and alternatives. | | What form–meaning construction occurred? | UCxn conventions | Define the radiology constructicon and recognition rules. | | How is a construction classified cross-linguistically? | MoCCA where suitable | Record alignment without forcing a match. | | What semantic situation and roles are expressed? | FrameNet where suitable | Define a minimal radiology frame profile and mappings. | | Which medical concept is denoted? | RadLex | Generate and constrain grounded candidates. | | Which mentions share discourse identity? | This specification | Define referent formation and resolution rules. | | What does polarity, certainty, or comparison scope over? | This specification, informed by linguistic evidence | Define frame composition and ambiguity rules. | | Which relations appear in final output? | This project | Define canonical projections and validation. | No authority is allowed to answer a question merely because it has a nearby concept. Authority follows semantic responsibility. ## 2.6 Initial Construction Families The initial constructicon is organized by linguistic function rather than by individual lexical triggers. | Construction family | Core elements | Initial examples | | --- | --- | --- | | `finding_predication` | `finding`, optional `state` | `The lungs are clear.` | | `telegraphic_finding` | `finding`, optional modifiers | `Borderline cardiomegaly.` | | `existential_presentation` | `presented` | `There is a nodule.` | | `negated_finding` | `negator`, `scope` | `No pneumothorax.` | | `coordination` | `conjuncts`, `coordinator` | `pneumothorax or effusion` | | `located_finding` | `figure`, `spatial_relation`, `ground` | `opacities in the left apex` | | `measured_entity` | `entity`, `quantity`, optional `dimension` | `8 mm nodule` | | `spatial_measurement` | `figure`, `distance`, `relation`, `ground` | `tip 5 cm above the carina` | | `comparison` | `entity`, `attribute`, `direction`, optional `baseline` | `stable from prior` | | `epistemic_qualification` | `content`, `cue`, `strength` | `could represent ...` | | `alternative_characterization` | `subject`, `alternatives`, cues | `probably X ... difficult to exclude Y` | | `recommendation` | `action`, optional `target`, `rationale` | `CT ... is recommended` | | `anaphoric_reference` | `anaphor`, candidate `antecedent` | `It measures 6 mm.` | The names are provisional. Each accepted construction type must eventually have a definition, licensed elements, recognition tests, counterexamples, and semantic contribution. ## 2.7 Initial Frame Profile The first implementation should require only frames demonstrated by accepted examples. | Frame | Roles | Purpose | | --- | --- | --- | | `FindingAssertion` | `content`, `polarity`, `certainty` | States whether and with what commitment finding content is presented. | | `PropertyState` | `entity`, `property`, `value` | Represents states such as normal, enlarged, or clear. | | `Location` | `figure`, `relation`, `ground` | Represents spatial organization. | | `Measurement` | `entity`, `value`, `unit`, optional `dimension` | Represents an entity attribute measurement. | | `SpatialMeasurement` | `figure`, `value`, `unit`, `relation`, `ground` | Represents a measured spatial relation. | | `Comparison` | `entity`, `attribute`, `direction`, optional `baseline` | Represents change or stability relative to a comparison context. | | `Characterization` | `subject`, `characterization`, `certainty` | Represents an interpretation of an observed finding. | | `Recommendation` | `action`, optional `target`, `rationale`, `certainty` | Represents a recommended future action. | `certainty` initially preserves a normalized category and the exact lexical cue. The first category vocabulary is provisional: ```text asserted probable possible cannot_exclude ``` These categories must not erase distinctions among source cues. A later corpus study may require a richer ordered or multidimensional model. This is an empirical refinement task, not an owner-level design decision: retain the cue, use the smallest vocabulary that explains accepted examples, and extend it when a counterexample requires a semantic distinction. An internal frame directly uses a FrameNet identity only when its meaning and role constraints match the corpus interpretation. Otherwise the project defines a local frame and records any useful FrameNet relationship as an alignment. This rule avoids requiring an advance choice between wholly external and wholly local frame vocabularies. ## 2.8 Intermediate Representation Shape The IR is a typed, versioned graph. Its serialization format is secondary to its concepts and invariants. JSON is the initial external interchange form; Prolog terms may realize the same model inside the semantic interpreter. An interpretation execution has the following provisional shape: ```json { "ir_version": "draft-0", "execution_id": "exec-...", "source": { "document_id": "kaggle-rad-reports-000048-sectionless_dictation", "text": "...", "provenance": {"corpus": "generated-transcript-corpus", "variant": "sectionless_dictation"} }, "linguistic_observations": [], "mentions": [], "constructs": [], "referents": [], "frames": [], "groundings": [], "ambiguities": [], "diagnostics": [], "assertions": [], "validation": {} } ``` Every derived object contains: ```text id type status evidence[] derived_by ``` where `status` is one of: ```text accepted candidate rejected unresolved ``` A rejected candidate is retained when its rejection explains a consequential choice. Implementations need not retain every mechanically generated candidate. ## 2.9 Provenance Graph Provenance is represented as edges among identified objects rather than as an unstructured explanation string. Initial edge types are: ```text anchored_in(mention, source_span) observed_in(linguistic_observation, source_span) recognized_from(construct, linguistic_observation_or_span) binds(construct, role, object) evokes(frame, construct) fills(frame, role, object) refers_to(mention, referent) grounded_as(mention_or_referent, ontology_concept) derived_from(assertion, object) validated_by(object, validation_rule) ``` An implementation may serialize these relationships inline or as explicit edges, provided their identity and direction remain recoverable. --- # 3. Survey of Decisions and Translation Rules ## 3.1 Example Method Each example distinguishes five things: ```text SOURCE EXPRESSION What the transcript contains. LINGUISTIC EVIDENCE What observable form supports an interpretation. CONSTRUCTION Which conventional form–meaning pattern is recognized. SEMANTIC CONTRIBUTION Which candidate referents, frames, roles, or assertions are licensed. LIMIT What the example does not license. ``` The examples below specify semantic obligations. They do not prescribe exact Grew rules, Prolog predicates, parser calls, or JSON layout. ## 3.2 Telegraphic Finding Description ### Source example From `kaggle-rad-reports-000002-sectionless_dictation`: ```text Borderline cardiomegaly. ``` ### Linguistic interpretation The expression is a verbless radiology clause. Its head denotes a finding or state; `borderline` qualifies its degree or category boundary. The absence of a finite verb is licensed by the radiology dictation sublanguage and is not, by itself, an incomplete parse diagnostic. ### Design decision The local `telegraphic_finding` construction licenses a finding assertion from a nominal or adjectival fragment when transcript context and lexical evidence support that reading. Illustrative annotation: ```text Construct telegraphic_finding finding → "cardiomegaly" qualifier → "Borderline" Frame FindingAssertion content → cardiomegaly discourse referent polarity → positive certainty → asserted ``` `borderline` is preserved as source-backed qualification. Whether it becomes a `PropertyState`, a degree value, or part of RadLex grounding is open pending additional examples. ### Formal obligation ```text telegraphic_finding(c) ∧ binds(c,finding,m) ∧ radiology_transcript_context(c) → ∃ r,a : refers_to(m,r) ∧ FindingAssertion(a,r,positive) ``` ### Does not entail The construction does not establish that cardiomegaly is clinically true. It records that the transcript positively presents that content. ## 3.3 Negation and Coordination ### Source example From `kaggle-rad-reports-000005-sectionless_dictation`: ```text There is no pneumothorax or pleural effusion. ``` ### Linguistic interpretation An existential/presentational clause contains a negator whose scope includes a coordination. The coordination introduces two finding descriptions. The shared negation distributes to both conjuncts unless syntactic or constructional evidence supports a narrower scope. ### Design decision Coordination is represented before polarity is projected. Negation applies to the coordinated semantic contents, producing two negative finding assertions with shared scope provenance. ```text Construct coordination c1 conjunct → "pneumothorax" conjunct → "pleural effusion" coordinator → "or" Construct negated_finding c2 negator → "no" scope → c1 Frame FindingAssertion a1 content → pneumothorax content polarity → negative Frame FindingAssertion a2 content → pleural-effusion content polarity → negative ``` ### Formal obligation ```text neg_scope(n, coordination(c,{x₁,...,xₙ})) → ∀ xᵢ ∈ {x₁,...,xₙ} : negative_assertion(xᵢ,n) ``` This rule applies only to a coordination licensed as wholly inside negation scope. ### Does not entail The mentions do not create positive pneumothorax or effusion findings. RadLex concept matches do not reverse the polarity supplied by the construction. ## 3.4 Finding Measurement and Location ### Source example From `kaggle-rad-reports-000048-sectionless_dictation`: ```text There is an 8mm nodule in the left lower lobe. ``` ### Linguistic interpretation The noun phrase contains a measured finding and a prepositional location modifier. The measurement and location share the nodule as their semantic participant. ### Design decision The expression evokes separate `Measurement` and `Location` frames. Composition unifies their entity/figure roles through one finding referent. ```text Mentions m1 → "8mm" m2 → "nodule" m3 → "left lower lobe" Referents f1 → introduced by m2 a1 → introduced by m3 Frame Measurement entity → f1 value → 8 unit → mm dimension → size, unresolved subtype Frame Location figure → f1 relation → in ground → a1 ``` The raw quantity text, normalized numeric value, and normalized unit are all preserved. A normalization is a derived representation, not a replacement for the source mention. ### Formal obligations ```text measured_entity(c,m_entity,m_quantity) → ∃ r,f : refers_to(m_entity,r) ∧ Measurement(f,entity=r,quantity=m_quantity) located_finding(c,m_figure,m_relation,m_ground) → ∃ r₁,r₂,f : refers_to(m_figure,r₁) ∧ refers_to(m_ground,r₂) ∧ Location(f,figure=r₁,relation=m_relation,ground=r₂) ``` ### Does not entail The preposition `in` does not automatically become a RadLex object property. The measurement does not establish which anatomical dimension was measured unless the construction or domain evidence licenses that conclusion. The working rule is conservative: a dimension is accepted only when it is explicit in the transcription or licensed by a tested construction rule. Ontology knowledge alone does not supply a transcript-level dimension. ## 3.5 Spatial Measurement Is Not Entity Size ### Source example From `kaggle-rad-reports-000057-sectionless_dictation`: ```text The tracheostomy tube tip is 5 cm above the carina. ``` ### Linguistic interpretation The quantity measures the distance expressed by the spatial relation `above`. It does not measure the tube tip itself. ### Design decision This example requires a `spatial_measurement` construction and a `SpatialMeasurement` frame distinct from `measured_entity` and `Measurement`. ```text Frame SpatialMeasurement figure → tube-tip referent value → 5 unit → cm relation → above ground → carina referent ``` ### Formal obligation ```text spatial_measurement(c,figure,q,relation,ground) → SpatialMeasurement(figure,q,relation,ground) ∧ ¬ entity_size(q,figure) ``` The final negative term expresses a translation prohibition, not necessarily a stored negative assertion. ### Does not entail The tube tip is not five centimetres in size. The carina is not a measured entity. Linear proximity in the sentence is insufficient to determine the measurement target. ## 3.6 Comparison and Stability ### Source example From `kaggle-rad-reports-000057-sectionless_dictation`: ```text There are prominent diffuse bilateral interstitial opacities, stable from prior radiographs. ``` ### Linguistic interpretation The participial/adjectival comparison expression predicates stability of the opacities relative to a prior-study baseline. `prominent`, `diffuse`, and `bilateral` describe the current finding; `stable` relates an attribute or overall state across observations. ### Design decision The comparison is represented independently from the positive finding assertion: ```text Frame FindingAssertion content → interstitial-opacities referent polarity → positive Frame Comparison entity → same referent attribute → unresolved overall finding state direction → unchanged baseline → prior radiographs ``` The exact baseline may remain a discourse description rather than a fully grounded study referent when redaction or missing context prevents resolution. ### Formal obligation ```text comparison(c,entity,cue="stable",baseline) → Comparison(entity,attribute=?,direction=unchanged,baseline) ``` The unresolved attribute is explicit and valid. ### Does not entail `stable` does not mean normal, benign, absent, or clinically insignificant. It does not identify the baseline date when the source does not supply one. ## 3.7 Epistemic Qualification and Alternative Characterization ### Source example From `kaggle-rad-reports-000004-sectionless_dictation`: ```text Probably scarring in the left apex, although difficult to exclude a cavitary lesion. ``` ### Linguistic interpretation The expression offers at least two characterizations of an observed finding with different epistemic cues. `probably` supports scarring more strongly; `difficult to exclude` keeps a cavitary lesion as a live alternative. The second characterization is not negated merely because `exclude` occurs in the phrase. ### Design decision Negation detection must operate over constructions and scope, not keyword presence. The interpretation contains two `Characterization` candidates linked to the same observed-content referent, preserving the different cues. ```text Characterization c1 subject → observed apical abnormality characterization → scarring certainty → probable cue → "Probably" Characterization c2 subject → same observed abnormality characterization → cavitary lesion certainty → cannot_exclude cue → "difficult to exclude" ``` The relationship between these candidates is represented as an alternative- characterization set. A downstream application may order the alternatives but must retain both and their source wording. ### Formal obligation ```text alternative_characterization(subject,{(x,cue₁),(y,cue₂)}) → Characterization(subject,x,normalize(cue₁)) ∧ Characterization(subject,y,normalize(cue₂)) ∧ alternatives(x,y) ``` ### Does not entail The transcript does not positively establish either diagnosis as clinical truth. `difficult to exclude` does not establish absence. An ontology hierarchy between the candidates does not authorize collapsing the alternatives. ## 3.8 Recommendation Is Not a Finding ### Source example From `kaggle-rad-reports-000009-sectionless_dictation`: ```text CT chest with contrast is recommended. ``` ### Linguistic interpretation The passive predicate presents a recommended future imaging action. It does not state that the CT has occurred. ### Design decision The `recommendation` construction evokes a `Recommendation` frame. Its action may be medically grounded, but it is not projected as a current examination or finding. ```text Frame Recommendation action → CT chest manner_or_protocol → with contrast rationale → unresolved or linked from discourse context certainty → asserted recommendation ``` ### Formal obligation ```text recommendation(c,action,target,rationale?) → Recommendation(action,target,rationale?) ∧ ¬ performed(action) ``` Again, the final term is a prohibited inference rather than a required stored negative assertion. ### Does not entail The recommended examination has not necessarily been ordered, scheduled, or performed. The recommendation does not itself validate its clinical rationale. ## 3.9 Cross-Sentence Reference ### Canonical example ```text There is a nodule in the right upper lobe. It measures 6 mm. ``` ### Linguistic interpretation `a nodule` introduces a discourse referent. `It` is a distinct mention whose candidate antecedent is that referent. The measurement frame uses the resolved referent as its entity. ### Design decision Anaphora resolution is candidate-based. Agreement, discourse salience, constructional role, semantic type, and locality may constrain candidates. No single token-distance heuristic is authoritative. ```text Mention m1 → "a nodule" Mention m2 → "It" Referent f1 introduced_by → m1 referred_to_by → m2 Frame Measurement entity → f1 value → 6 unit → mm ``` ### Formal obligation ```text anaphor(m) ∧ candidates(m)={r₁,...,rₙ} ∧ uniquely_preferred(rᵢ) → refers_to(m,rᵢ) anaphor(m) ∧ multiple_undominated_candidates(m) → ambiguity(reference,m,candidates(m)) ``` ### Does not entail The pronoun does not introduce a second finding merely because it is a second mention. A nearby noun is not necessarily its antecedent. ## 3.10 Repetition Does Not Supply Hidden Report Identity The current input may repeat similar expressions, but the interpreter sees only a transcription character stream. It may not consult a source report or infer a Findings→Impression relationship from the historical origin of the record. Two repeated expressions are distinct mentions. They resolve to one referent only when transcript-internal discourse evidence uniquely supports that identity. Otherwise identity remains unresolved. ```text same wording + compatible grounding ⇏ same referent ``` ╔════════════════════════════════════════════════════════╗ ║ INPUT BOUNDARY ║ ║ ║ ║ Historical report structure supplies no evidence ║ ║ for transcript mention identity. ║ ╚════════════════════════════════════════════════════════╝ An actual generated or human transcription exhibiting repeated reference must be annotated before stronger identity rules are specified. ## 3.11 Redaction and Damaged Constructions ### Source examples The corpus preserves `[REDACTED]` markers, sometimes inside expressions needed for interpretation. ### Design decision `[REDACTED]` is an opaque source token. It is never normalized into guessed content. A construction may bind an explicitly unresolved element when the remaining source supplies enough evidence to recognize the construction. ```text Construction element role → baseline filler → unresolved evidence → [REDACTED] source span Diagnostic type → redacted_required_element severity → partial_interpretation ``` If the redaction prevents recognition itself, the system emits a diagnostic rather than fabricating a construction. ### Does not entail The redaction marker does not denote a person, date, finding, anatomy, or other domain entity merely because one of those would make the sentence grammatical. ## 3.12 Mention and Referent Grounding ### Upstream meaning RadLex provides abstract medical concepts, labels, synonyms, hierarchy, and ontology relations. A source expression may lexically evoke one or more RadLex concepts. A discourse referent may be characterized by one or more mentions. ### Provisional decision The IR permits both mention grounding and referent grounding with distinct meanings: ```text mention grounding This expression is a lexical/contextual realization candidate for this RadLex concept. referent grounding The composed discourse interpretation characterizes this referent using this RadLex concept. ``` Mention grounding supplies evidence for referent grounding; it is not automatically copied. Composition, polarity, qualification, and competing mentions may affect the referent-level result. ### Formal obligations ```text mention_grounding(m,c) → Mention(m) ∧ RadLexConcept(c) referent_grounding(r,c) → Referent(r) ∧ RadLexConcept(c) ∧ supported_by_composed_evidence(r,c) ``` ### Does not entail A lexical match does not prove a unique grounding. Neither kind of grounding creates a referent. Grounding does not change assertion polarity or certainty. ╔════════════════════════════════════════════════════════╗ ║ WORKING RULE ║ ║ ║ ║ Accept grounding only at uniquely supported semantic ║ ║ specificity; otherwise preserve the candidates. ║ ╚════════════════════════════════════════════════════════╝ The working test combines source-backed lexical or semantic evidence, construction-role compatibility, composed-context compatibility, and the absence of an undominated incompatible candidate at the claimed specificity. Concrete examples may refine this rule without requiring a project-scope decision. ## 3.13 Candidate Preservation and Resolution ### Design decision Interpretation is candidate-producing. A pass may: - introduce a candidate supported by identified evidence; - accept a candidate because a stated rule is satisfied; - reject a candidate with a stated reason; - group undominated candidates into an ambiguity; - leave a role unresolved with a diagnostic. A pass may not silently discard a materially supported candidate. Candidate preference is represented as an evidence-bearing relation: ```text preferred(candidate_a, candidate_b, rule, evidence) ``` Acceptance requires either a unique supported candidate or an explicit rule that permits several compatible candidates to coexist. ### Formal obligation ```text supported(c₁) ∧ supported(c₂) ∧ ¬ dominates(c₁,c₂) ∧ ¬ dominates(c₂,c₁) ∧ incompatible(c₁,c₂) → preserve_ambiguity({c₁,c₂}) ``` ## 3.14 Validation Profile The initial validator checks at least the following obligations. ### Source integrity - Every source span is within the document boundary. - Every stored span text equals the immutable source substring. - Every mention has at least one source span. ### Construction integrity - Every construct has exactly one defined construction type. - Every construction element uses a role licensed by that type. - Required roles are filled or explicitly unresolved. - Recognition evidence is retained. ### Discourse and frame integrity - Every `refers_to` target is a defined referent. - Every referent is licensed by at least one mention or permitted derivation. - Every frame role is licensed by its frame type. - Every frame is evoked by a construct or permitted semantic rule. - Polarity and certainty attach to semantic content, not ontology concepts. ### Grounding integrity - Every grounding target exists in the pinned RadLex bundle. - Candidate and accepted groundings are distinguishable. - No grounding operation creates a referent. - Ontology implications do not populate transcript individuals. ### Ambiguity and diagnostic integrity - Every ambiguity contains at least two supported alternatives. - Every alternative identifies its distinguishing choice. - Rejected consequential alternatives retain a reason. - Every diagnostic identifies a source span or interpretation object. ### Provenance integrity - Every accepted frame and output assertion has a derivation path to source. - Normalized values retain the source expression from which they were derived. - External resource versions used by the execution are recorded. Validation success means that the representation obeys this contract. It does not mean every expression was interpreted or every ambiguity resolved. ## 3.15 Annotation Workflow Examples progress through the following states: ```text selected ↓ span-annotated ↓ construction-annotated ↓ semantically annotated ↓ reviewed ↓ accepted as executable example ``` Each transition records the annotation-profile version and author or process. Automated parser and grounding suggestions remain distinguishable from human- accepted annotations. Disagreement is represented as alternatives or an adjudication record. It is not overwritten without history. ## 3.16 Refinement Acceptance Criteria This draft is ready to become the first implementation contract when: - an initial stratified set of examples has been selected from the corpus; - every selected example has exact source-span annotations; - each accepted construction type has licensed roles and counterexamples; - each required frame has a defined role inventory; - at least one complete example traverses R0 through R5; - mention and referent grounding are tested against concrete examples; - negation, uncertainty, comparison, recommendation, and redaction remain semantically distinct; - ambiguity and partial interpretation have machine-representable examples; - the IR can represent every accepted example without ad hoc fields; - validators enforce the conceptual specification's cross-cutting invariants; - external knowledge and rule versions are recorded in every execution; - open decisions that block the first vertical slice are resolved, while later decisions are explicitly deferred. --- # 4. Reference Appendices ## Appendix A. Provisional Object Schemas The schemas below define semantic fields, not a required physical serialization. ### A.1 Source span ```text SourceSpan id document_id start_character end_character text ``` ### A.2 Mention ```text Mention id spans[1..n] categories[0..n] normalization_candidates[0..n] evidence[1..n] ``` ### A.3 Construct ```text Construct id construction_type elements[1..n] recognition_evidence[1..n] authority_alignment[0..n] status ConstructionElement role filler evidence[1..n] ``` ### A.4 Referent ```text Referent id introduced_by[1..n] referred_to_by[0..n] grounding_candidates[0..n] status ``` ### A.5 Frame ```text Frame id frame_type roles[1..n] evoked_by[1..n] authority_alignment[0..n] status ``` ### A.6 Grounding ```text Grounding id subject ontology concept_id status lexical_evidence[0..n] contextual_evidence[0..n] constraint_evidence[0..n] resource_version ``` ### A.7 Ambiguity ```text Ambiguity id kind subject alternatives[2..n] unresolved_because[1..n] ``` ### A.8 Diagnostic ```text Diagnostic id type severity subject evidence[1..n] detail ``` ## Appendix B. Required Counterexample Pairs Each construction family should be tested with contrasts that prevent shallow keyword translation. | Superficially similar forms | Required distinction | | --- | --- | | `8 mm nodule` / `tip 5 cm above carina` | Entity measurement / spatial-relation measurement | | `No pneumothorax` / `difficult to exclude pneumothorax` | Negative assertion / live uncertain alternative | | `stable opacity` / `normal lung` | Unchanged state / normal state | | `CT is recommended` / `CT demonstrates` | Future recommended action / evidential examination statement | | `nodule in the lobe` / `nodule near the fissure measuring 6 mm` | Clear shared participant / potentially ambiguous attachment | | repeated `nodule` / two explicitly enumerated nodules | Coreferent mentions / distinct same-type referents | | `[REDACTED] lobe` / `left lobe` | Unresolved anatomy / grounded anatomy | ## Appendix C. Version Metadata Every interpretation execution should record the materially relevant versions: ```text IR profile version source corpus and record version UD parser and model version UD specification version UCxn schema and rule version local constructicon version MoCCA database version FrameNet version local frame-profile version RadLex bundle version and checksum semantic-rule version validation-rule version ``` ## Appendix D. Open Design Decision ╔═══════════════════════════════════════════════════════╗ ║ ⚠ OPEN DESIGN DECISION ║ ║ ║ ║ Define which frames project into canonical R5 output ║ ║ predicates and which remain frame structures. ║ ╚════════════════════════════════════════════════════════╝ This is genuinely product-defining because it determines the public semantic contract of validated output. It does not block the first examples: until a projection is accepted, R5 may preserve the validated frame structure itself. ## Appendix E. Design Provenance This appendix is non-normative. ### E.1 Refinement Before Implementation **Joint refinement.** The design conversation identified the need for a layer between the conceptual specification and an implementation plan. That layer is example-driven and defines a machine-representable semantic contract by decomposing corpus expressions through the appropriate linguistic and domain authorities. ### E.2 Lower-Level Dependency **Human-led clarification.** This document is explicitly a lower-level specification that assumes the higher-level conceptual specification. It does not reopen the established ontology or duplicate its full argument. ### E.3 Initial Corpus Sufficiency **Human-led scope decision.** The retained generated transcriptions are accepted as sufficient development input to get the interpreter running, without productionizing it or making claims about future real human transcriptions. ### E.4 Authorities Have Bounded Roles **Joint refinement.** The corpus is decomposed using UD, UCxn, MoCCA, FrameNet, and RadLex where each is authoritative. External vocabulary is reused without allowing one authority to absorb responsibilities belonging to another layer. ### E.5 Transcription Is the Complete Input **Human-led correction.** An example incorrectly treated Findings and Impression sections from a radiology report as transcript structure. The source reports were removed, normalized report-copy records were removed from the generated corpus, and the interpretation boundary was restated: only the generated `transcript_text` character stream supplies linguistic evidence.