Stage 5: Communicative-Function Tagging

Distinct-decision rule: the lexical sense may provide useful context, but it must not determine the whole-sentence function.

Function Pass 1

python scripts/05a_run_pass1.py \
 --sentences data/en_sampled_occurrences.csv \
 --taxonomy taxonomy/cefr_function_taxonomy_v0_2.csv \
 --output data/en_function_pass1.csv

Pass 1 classifies what the whole sentence is doing communicatively. It does not classify the sampled lemma, its lexical sense or merely the sentence topic.

Informed Function Pass 2

python scripts/05b_run_pass2.py \
 --sentences data/en_sampled_occurrences.csv \
 --taxonomy taxonomy/cefr_function_taxonomy_v0_2.csv \
 --pass1 data/en_function_pass1.csv \
 --sense_pass1 data/en_sense_pass1.csv \
 --output data/en_function_pass2.csv

Pass 2 sees the initial function, the initial lexical sense and both rationales. It accepts, changes or marks the function uncertain and records whether the sense helped, conflicted with or was irrelevant to the decision.

Optional blind validation sample

python scripts/05b_run_pass2.py \
 --sentences data/en_blind_sample.csv \
 --taxonomy taxonomy/cefr_function_taxonomy_v0_2.csv \
 --blind \
 --output data/en_function_pass2_blind.csv

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

Adjudication

python scripts/05c_run_pass3.py \
 --pass1 data/en_function_pass1.csv \
 --pass2 data/en_function_pass2.csv \
 --taxonomy taxonomy/cefr_function_taxonomy_v0_2.csv \
 --only_problem_cases \
 --output data/en_function_pass3.csv

Adjudication focuses on changed, uncertain and low-confidence cases or possible taxonomy problems.

Example

Sentence	Target sense	Sentence function
Apply the cream twice daily.	put a substance onto a surface	giving an instruction
The nurse applied the cream.	put a substance onto a surface	reporting an event

The same sense can occur in different functions, and the same function can contain many senses.

Continue to Stage 6 →