Six-Stage Workflow

How one target lemma occurrence becomes a provisional, reviewable sense-and-function row.

Pilot material — not validated CEFR data.
The workflow creates structured candidate annotations for human review.

Workflow overview

Stage 1: Corpus preparation

Document source, licence, language, date and register.

Stage 2: Processing

Tokenise, POS-tag, lemmatise and calculate corpus evidence.

Stage 3: Sampling

Create stable target-occurrence rows.

Stage 4: Lexical-sense tagging

Propose and approve inventories, create an initial sense annotation and run informed review.

Stage 5: Function tagging

Create an initial sentence-function annotation and run informed taxonomy-based review.

Stage 6: QA and expert review

Adjudicate selected cases, join outputs and aggregate by sense.

The data unit

Each row represents one target lemma occurrence in one corpus sentence:

row_id
language
sentence_id
sentence
target_token
target_lemma
target_pos

If a sentence contains two sampled targets, it appears in two rows with different row IDs.

Annotation and review sequence

StepSenseFunction
Pass 1Initial target-lemma sense proposal from the approved inventoryInitial whole-sentence function proposal from the controlled taxonomy
Pass 2Reviews Pass 1 sense while seeing Pass 1 function and both rationalesReviews Pass 1 function while seeing Pass 1 sense and both rationales
AdjudicationResolves changed, uncertain, OTHER, UNCLEAR or low-confidence casesResolves changed, uncertain or low-confidence cases

Pass 2 is deliberately informed because the production objective is the strongest defensible candidate annotation. However, the reviewer must treat the other annotation as contextual evidence rather than proof.

Blind validation sample

A smaller sample is also processed with Pass 1 hidden. This allows the project to estimate reliability and possible anchoring without making blind annotation the default production workflow. Blind and informed results are stored separately.

Human decision points

  1. Approve or revise each lemma-specific sense inventory.
  2. Review changes, uncertainty, low confidence, OTHER and UNCLEAR.
  3. Review a random sample of accepted high-confidence rows.
  4. Keep final human sense and function decisions separate.
  5. Compare blind, informed and human outcomes.
  6. Align reviewed language-specific senses to concept IDs only afterwards.