1. Fundamental unit
The operational row is a target lemma occurrence in a corpus sentence. The lexical layer asks which sense the target occurrence realises; the pragmatic layer asks what the whole sentence is doing.
2. Distinct annotation objects
Sense and function remain separate analytical decisions. Sense belongs to the target lemma occurrence; function belongs to the whole sentence. A function label may provide contextual evidence about a sense, and a sense label may help interpret the sentence, but neither automatically determines the other.
3. Sense inventories
For each language + lemma + POS, sampled sentences are considered together. AI proposes a small coarse inventory; a language expert merges, splits, removes, renames or adds senses and marks the retained inventory approved before bulk tagging.
4. Sense-splitting rule
Create a separate sense only where the distinction could materially affect translation, grammatical construction, learner understanding, CEFR judgement or pedagogical treatment. Prefer defensible broad senses over dictionary micro-senses. Every inventory includes OTHER and UNCLEAR.
5. Pass 1 and informed Pass 2
Pass 1 produces initial sense and function proposals separately. Production Pass 2 then sees both Pass 1 labels and rationales and acts as an informed critical reviewer. It may accept, change or mark either proposal uncertain. The prompts explicitly require the reviewer to use the other annotation only as contextual evidence, not as proof.
6. Anchoring and blind validation
Informed review may improve correction but can also anchor the reviewer to Pass 1. The project therefore retains a blind mode for a smaller validation sample. Blind and informed outputs are stored separately and compared against expert decisions.
7. Targeted adjudication
Adjudication is prioritised where Pass 2 changes Pass 1, marks uncertainty, recommends review, selects OTHER or UNCLEAR, or identifies a possible inventory or taxonomy problem. Agreement is useful evidence but is not proof of correctness.
8. Multilingual analysis
Inventories are developed separately within each language. English senses are not translated and imposed on other languages. Reviewed senses may later be aligned to broader concept IDs, preserving one-to-many and constructional relationships.
9. Corpus limitations
Sense and function distributions describe the selected corpus. A news-heavy corpus will overrepresent reporting and description and underrepresent interactional functions; corpus composition must therefore accompany interpretation.
10. Validation
- Compare informed Pass 2, blind Pass 2 and human decisions.
- Measure model–human and human–human agreement.
- Inspect change rates, anchoring rates, confusion matrices and OTHER/UNCLEAR rates.
- Review disagreements and a random sample of accepted high-confidence rows.
- Compare lemma-level and sense-level functional profiles.
- Report results by language, register and POS.