MIPVU Annotation Guide

This guide defines the first-pass metaphor-identification layer for the workbench. MIPVU decisions come before CMT mapping and Koenigsbergian interpretation.

MIPVU answers a narrow question:

Is this lexical unit metaphor-related in this context?

It does not answer whether the passage supports Koenigsberg’s Law of Sacrifice, what conceptual metaphor is active, or whether a historical practice has been enacted. Those are downstream questions.

Workflow

  1. Generate a source-language worklist:

    python3 scripts/generate-mipvu-worklist.py --case <case-id>

  2. Review every lexical unit in cases/<case>/corpus/mipvu/<document_id>_mipvu.json.

  3. Assign one decision_type to every lexical unit.

  4. Add detailed rationale for metaphor-related or uncertain decisions.

  5. Create CMT/Koenigsbergian annotations only from MIPVU-positive or uncertain lexical units.

For a pilot, training, or reliability sample, review complete sentences rather than isolated exciting words. Every lexical unit in the sampled sentence should receive a decision. Non-metaphor decisions may remain brief; metaphor-related and uncertain decisions require the full rationale fields below.

Decision Types

  • non_metaphor: contextual meaning does not contrast with a more basic meaning in a metaphorically relevant way.
  • mipvu_indirect: contextual meaning contrasts with a more basic meaning and can be understood by comparison with it.
  • mipvu_direct: direct metaphor, simile, analogy, or explicit comparison.
  • mipvu_implicit: omitted or implied lexical material is metaphor-related.
  • mipvu_personification: nonhuman entity is construed through human agency, emotion, body, or action.
  • uncertain: evidence is insufficient or contested; preserve the uncertainty.
  • excluded_nonlexical: generated item should not have been treated as a lexical unit.

Do not use mipvu_indirect, mipvu_direct, mipvu_implicit, or mipvu_personification merely because a term is theoretically interesting. The contextual meaning must contrast with a more basic meaning and be understandable by comparison with it. If that contrast is plausible but not settled, use uncertain.

Required Rationale

For mipvu_indirect, mipvu_direct, mipvu_implicit, mipvu_personification, and uncertain, fill in:

  • contextual_meaning
  • basic_meaning
  • basic_meaning_source
  • contrast_explanation
  • comparison_basis
  • confidence
  • review_notes

Manual dictionary/source citation is the v1 policy. The citation can name a dictionary, critical edition note, scholarly lexicon, or other basic-meaning source used by the annotator.

When a term is listed in cases/lincoln/references/historical-semantics-notes.md, also fill:

  • semantic_shift_risk
  • a review_notes explanation of how the period-control note affected the decision

Candidate source/target domain hints may be recorded for later CMT review, but they do not replace the MIPVU decision. CMT mappings should still be made in their own layer.

Review Status

  • pending: lexical unit has not been reviewed.
  • needs_review: annotator has flagged the unit for later review but has not accepted a decision.
  • reviewed: annotator has made a first-pass decision.
  • accepted: decision has passed adjudication or final human review.
  • rejected: prior decision was reviewed and rejected.

Validation treats any non-pending unit as intentionally reviewed. Reviewed, accepted, rejected, or needs-review units must have a valid decision_type. Metaphor-related and uncertain decisions must also have all required rationale fields.

Lincoln Pilot Sample

The current Lincoln pilot sample is documented in cases/lincoln/metadata/mipvu-pilot-sample.json.

It reviews full sentences from:

  • the Lyceum Address live/die/suicide national survival formulation;
  • the Gettysburg Address burial/sacrifice/national life sentence;
  • the Gettysburg Address final dedication/new birth sentence;
  • the Second Inaugural blood repayment/divine judgment sentence.

This sample demonstrates the workflow and creates a future reliability target. It does not complete MIPVU review for the Lincoln corpus.

Reliability And Adjudication

The Lincoln reliability workflow is documented in cases/lincoln/quality/reliability-sample.json and cases/lincoln/quality/reliability-report.md.

Use the pilot sample for joint training and codebook calibration. Use the separate reliability sample for independent double-coding. The v1 agreement policy is:

  • primary measure: Cohen’s kappa for metaphor-related vs. not metaphor-related;
  • secondary measure: percent agreement on the full decision_type;
  • adjudication log: cases/lincoln/quality/adjudication-log.csv.

Disagreements must be categorized as lexical segmentation, contextual meaning, basic meaning, metaphor decision, confidence, or source-domain ambiguity. Adjudicated decisions should update the MIPVU worklist with review_status set to accepted, unless the unit remains explicitly deferred.

Language Policy

German and French documents are annotated in the source language. Add gloss_en only as an analytical aid. Do not make the MIPVU decision on an English gloss when the source text is German or French.

Use gloss_notes when translation choices affect the source-domain mapping, especially for politically loaded words such as Volk, Opfer, Vernichtung, gloire, patrie, honneur, or sacrifice.

Uncertainty

Use uncertain when:

  • basic and contextual meanings are difficult to separate;
  • OCR or translation risk affects the lexical unit;
  • the unit may be conventional but still politically active;
  • multiple source-domain readings remain plausible.

Uncertain MIPVU decisions may support exploratory CMT annotations, but aggregate claims should report rates with and without uncertain units.