Skip to content

Latest commit

 

History

History
52 lines (34 loc) · 3.15 KB

README.md

File metadata and controls

52 lines (34 loc) · 3.15 KB
Raw

Revisiting English Winogender Schemas for Consistency, Coverage, and Grammatical Case

This repository contains code and data for our paper, which presents WinoPron, an expanded dataset that also fixes issues with Winogender Schemas (Rudinger et al., 2018).

Data

We provide our newly constructed templates in a zipped data file. Please unzip it with the password vogelbeobachtung131719, using a command like unzip -P PASSWORD FILE.zip. It contains two files:

  • new_templates.tsv: WinoPron templates, where each row contains a sentence for a given occupation, participant and pronoun type, as well as a single-entity sentence that contains a version of the sentence with similar content but only one of the entities (either the occupation or participant) such that the coreference is maintained. The answer column is either 0 for a coreference with the occupation or 1 for coreference with the participant.
  • old_templates.tsv: Winogender schemas (Rudinger et al., 2018) templates for comparison, where each row contains a sentence for a given occupation, participant and pronoun type; the answer column is either 0 for a coreference with the occupation or 1 for coreference with the participant.

After unzipping, run:

python3 scripts/verification.py new_templates.tsv to verify the templates with some manual checks, which creates 4 files: automatically_validated_singles.tsv and automatically_validated.tsv for human validation with Google sheets; and double.tsv and single.tsv with single- and double-entity sentences for evaluation.

Run python3 scripts/prep_old_templates.py old_templates.tsv to instantiate and convert the Winogender schemas templates into double_old.tsv, which is in the same format as our WinoPron templates.

Running FLAN-T5

python3 scripts/score_models.py double.tsv single.tsv double_old.tsv (uses scripts/prompt.py behind the scenes) Evaluates FLAN-T5 models of five sizes on the data files with 10 x 3 prompts and stores results in subfolders named after the data files.

Running SpanBERT

python3 scripts/spanbert_inputs.py path/to/stimuli path/to/vocab/file Converts stimuli into jsonlines format for SpanBERT and creates two files: single.jsonlines and double.jsonlines

Running CAW-coref

python3 scripts/convert_for_caw.py Converts double.tsv, single.tsv and double_old.tsv to jsonlines format for the CAW-coref system.

Follow the setup instructions in the CAW-coref repository and produce output files named single_output.jsonlines, double_output.jsonlines, and double_old_output.jsonlines.

python3 scripts/convert_caw_output.py Converts single_output.jsonlines, double_output.jsonlines, and double_old_output.jsonlines into TSV output files in subfolders named after the data files.

Running LingMess

python3 scripts/evaluate_lingmess.py Evaluates LingMess on the data files and stores TSV results in subfolders named after the data files.

Compute performance

python3 scripts/compute_performance.py Computes accuracy for FLAN-T5, CAW-coref and LingMess on WinoPron (double-entity and single-entity versions), and Winogender schemas (Rudinger et al., 2018).