Stage 4: Lexical-Sense Tagging

4A: Propose the inventory

python scripts/04a_create_sense_inventory.py \
 --samples data/en_sampled_occurrences.csv \
 --output data/en_sense_inventory.csv

The proposal records short glosses, distinguishing features, typical patterns and supporting row IDs.

4B: Human approval

A language expert may merge, split, rename, add or remove senses. Retain OTHER and UNCLEAR, then set retained rows to inventory_status=approved.

4C: Sense Pass 1

python scripts/04b_run_sense_pass1.py \
 --samples data/en_sampled_occurrences.csv \
 --inventory data/en_sense_inventory.csv \
 --output data/en_sense_pass1.csv

Pass 1 selects the most defensible sense from the approved inventory and records confidence, rationale and any plausible alternative.

4D: Informed Sense Pass 2

python scripts/04c_run_sense_pass2.py \
 --samples data/en_sampled_occurrences.csv \
 --inventory data/en_sense_inventory.csv \
 --pass1 data/en_sense_pass1.csv \
 --function_pass1 data/en_function_pass1.csv \
 --output data/en_sense_pass2.csv

Pass 2 sees the initial sense, the initial sentence function and both rationales. It accepts, changes or marks the sense uncertain. The function may support interpretation, but it must not determine the lexical sense.

4E: Optional blind validation sample

python scripts/04c_run_sense_pass2.py \
 --samples data/en_blind_sample.csv \
 --inventory data/en_sense_inventory.csv \
 --blind \
 --output data/en_sense_pass2_blind.csv

Blind output is stored separately and used to study reliability and possible anchoring.

4F: Adjudication

python scripts/04d_run_sense_adjudication.py \
 --pass1 data/en_sense_pass1.csv \
 --pass2 data/en_sense_pass2.csv \
 --inventory data/en_sense_inventory.csv \
 --only_problem_cases \
 --output data/en_sense_pass3.csv

Adjudication focuses on changed, uncertain, low-confidence, OTHER, UNCLEAR and likely inventory-problem cases.

Continue to Stage 5 →