mssense-eval-benchmark / docs /inter_annotator_agreement.md
Novylab's picture
Upload 13 files
b6a6d0d verified
|
Raw
History Blame Contribute Delete
4.6 kB

Inter-Annotator Agreement — mssense-eval-benchmark v1.1

V5.15.h.3 artefact — JOT submission readiness.

Honest documentation of the inter-annotator-agreement (IAA) status of mssense-eval-benchmark v1.1. The position is: IAA in the conventional double-annotation sense is not applicable to most of the corpus by construction, and a small subset for which it would be applicable was instead subjected to a documented internal review.

Why IAA is a relevant question

IAA — typically reported as Cohen's κ, Krippendorff's α, or Fleiss's κ — measures consistency between two or more human annotators. It serves three purposes:

  1. Reliability evidence: if independent annotators agree, the labels reflect a property of the data, not idiosyncrasies of a single labeller.
  2. Construct validity: disagreement signals ambiguous task definitions.
  3. Reproducibility: future reviewers know the annotation procedure can in principle be replicated.

A reviewer of mssense-eval-benchmark v1.1 will reasonably ask: "what is the IAA?" The answer depends on which sub-corpus is asked about.

Composition and IAA applicability per block

Block Source N Annotation procedure IAA applicable?
controlled-balanced controlled_generation template engine 1500 Labels derived deterministically from generator template parameters No, by construction
seed-internal (audit) mssense_internal_audit 187 Single in-house annotator using an internal authoring rubric Not measured (single annotator)
seed-internal (validation) mssense_internal_validation 85 Same Not measured (single annotator)
Paraphrastic augmentations (V5.15.h) deterministic template-substitution script 93 Labels inherited deterministically from parent seed No, by construction
Total in v1.1-eval JSONL 1865

The 1500 controlled-balanced records and the 93 augmentations cannot have an IAA because their labels are not the output of human annotation. A second annotator looking at the same record would re-run the generator and necessarily get the same label.

The 272 seed-internal records could have an IAA but were authored by a single engineer with internal team review. This is documented as a limitation rather than masked.

What was done in lieu of formal IAA

Three procedures were applied to the 272 seed-internal records:

  1. Rubric-based authoring. Following an internal authoring rubric and the canonical schema (schema/evaluation_sample.v1_1.schema.json). Each label is justified by a written assertion block (assertions, anti_patterns, notes) visible in the JSONL itself.
  2. Schema validation. Records pass the strict v1.1 schema (schema/evaluation_sample.v1_1.schema.json).
  3. Team review on hard cases. Hard-difficulty seed-internal records were reviewed by a second annotator before merge. Label adjustments are reflected in the current JSONL but not recorded as κ.

These procedures do not substitute for IAA but are the standard for single-annotator benchmark construction. JOT itself has published dataset papers (XCorpus, Dietrich et al. 2017) without reporting κ, because the artefacts are not products of subjective annotation.

Planned v1.2 improvement

A future v1.2 release will introduce a dedicated IAA-measured subset of ~100 samples for which two annotators independently label expected_decision and expected_issue_types, with Cohen's κ reported in the release notes. The subset will be a stratified random draw from the seed-internal block, oversampling hard difficulty.

The protocol is planned, not yet executed, and will be tracked as a follow-up lot under V5.15 in plan/ARCHITECTURE_LIVING_PLAN_V5.md.

Suggested manuscript disclosure

Inter-annotator agreement is not applicable to 1593 of the 1865 evaluation samples (1500 controlled-balanced records and 93 paraphrastic augmentations) because their labels are derived deterministically from a generator or inherited from a parent seed through documented template substitution (a deterministic script, seed 20260610). The remaining 272 internal audit and validation records were authored by a single annotator following the canonical in-house rubric and reviewed by a second annotator on hard cases. A formal κ measurement on a stratified subset is planned for the v1.2 release.

Honest, defensible, and matches dataset-paper practice in JOT and adjacent venues.