Robust Pronoun Use Fidelity with English LLMs: Are they Reasoning, Repeating, or Just Biased?

This repository contains code, data and human annotations for our paper.

Getting started

Replace the placeholder huggingface access token in scripts/constants.py with your own. This is necessary for running Llama-2 models. Refer to the Llama-2 documentation for the specifics.

Code

• constants.py: secrets, API keys and such; not a runnable script
• pronouns.py: parametrized list of pronouns we use in the paper (simply extend this dictionary to evaluate on more pronouns); not a runnable script
• add_context.py: given task templates and context templates, create pronoun use fidelity data with an explicit introduction and various numbers of distractors; run with python3 scripts/add_context.py data/task.tsv data/context.tsv
• sample_templates.py: sample templates for the evaluation in our paper; run with python3 sample_templates.py
• score_models.py: scoring all the models in the paper; run with, e.g., python3 score_models.py 13_eo_task.tsv or python3 score_models.py 19*.tsv, which will create directories for each TSV file and populate them with a results file for each model
• prompt.py: prompting code for all the chat models in the paper, used by score_models.py; not a runnable script on its own
• sample_for_humans.py: sample templates for human evaluation of pronoun use fidelity; run with python3 sample_for_humans.py, which will create the file sampled_for_humans.tsv

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:

• task.tsv: task templates where each row contains a sentence for a given occupation, participant and pronoun type; the word column is equal to occupation because it is the answer for the coreference
• context.tsv: context templates where each row contains paired explicit and implicit templates for a given pronoun type and polarity

After unzipping, run

python3 scripts/add_context.py data/task.tsv data/context.tsv

to generate the 5 million+ instances of our complete dataset. This will generate the following files:

• eo_task.tsv (7200 lines): explicit occupation introduction + task
• eo_ep_task.tsv (86400 lines): explicit occupation introduction + explicit participant distractor + task
• eo_ep_ip_task.tsv (345600 lines): explicit occupation introduction + explicit participant distractor + 1 implicit participant distractor + task
• eo_ep_ip_ip_task.tsv (1036800 lines): as above, but with 2 implicit distractors
• eo_ep_ip_ip_ip_task.tsv (2073600 lines): as above, but with 3 implicit distractors
• eo_ep_ip_ip_ip_ip_task.tsv (2073600 lines): as above, but with 4 implicit distractors

To sample the 3 x 2,160 instances we use per setting to replicate our model evaluation, run

python3 scripts/sample_templates.py

which will generate 18 more files named like the ones above, prefixed with the random seeds 13, 17 and 19, each with 2,160 lines.

These files can be used to reproduce our reported numbers.