jennhu/metalinguistic-prompting: Materials for "Prompting is not a substitute for probability measurements in large language models" (EMNLP 2023)

[irthomasthomas]

• jennhu/metalinguistic-prompting: Materials for "Prompting is not a substitute for probability measurements in large language models" (EMNLP 2023)

Title

jennhu/metalinguistic-prompting: Materials for "Prompting is not a substitute for probability measurements in large language models" (EMNLP 2023)

Description

"Prompting is not a substitute for probability measurements in large language models

This repository contains materials for the EMNLP 2023 paper "Prompting is not a substitute for probability measurements in large language models" (Hu & Levy, 2023). The preprint is available on arXiv.

If you find the code or data useful in your research, please use the following citation:

@inproceedings{hu_prompting_2023,
title = {Prompting is not a substitute for probability measurements in large language models},
author = {Hu, Jennifer and Levy, Roger},
year = {2023},
booktitle = {Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing},
url = https://arxiv.org/abs/2305.13264
}

Evaluation materials

Evaluation datasets can be found in the datasets folder. Please refer to the README in that folder for more details on how the stimuli were assembled and formatted.

Evaluation scripts

The scripts folder contains scripts for running the experiments. There are separate scripts for models accessed through Huggingface (*hf.sh) and the OpenAI API (*openai.sh).

For example, to evaluate flan-t5-small on the SyntaxGym dataset of Experiment 3b, run the following command from the root of this directory:

bash scripts/run_exp3b_hf.sh syntaxgym google/flan-t5-small flan-t5-small

Please note that to run the OpenAI models, you will need to save your OpenAI API key to a file named key.txt in the root of this directory. For security reasons, do not commit this file (it is ignored in .gitignore).

Results and analyses

The results from the paper can be accessed by extracting the results.zip file. This will create a folder called results, which is organized by experiment:

• exp1_word-prediction
• exp2_word-comparison
• exp3a_sentence-judge
• exp3b_sentence-comparison

A few notes about the results:

• Each result file is named in the following format: <eval_type>.json. For the experiments where option order matters, there is an addition _<option_order> suffix in the name.
• Each result file is formatted as a JSON file, with two dictionaries:
  • meta contains meta information about the run (e.g., name of model, timestamp of run, path to data file)
  • results contains the results from the run, formatted as a list of dictionaries (one per stimulus item)

The results from the direct evaluation method are identical across Experiments 3a and 3b (see paper for details).

The figures from our paper can be reproduced using the analysis.ipynb notebook.

URL

https://github.com/jennhu/metalinguistic-prompting

Suggested labels

{'label-name': 'EMNLP-2023', 'label-description': "Materials and information related to the EMNLP 2023 paper 'Prompting is not a substitute for probability measurements in large language models'", 'confidence': 56.95}

[irthomasthomas]

Related content

#684 - Similarity score: 1.0

#238 - Similarity score: 0.91

#309 - Similarity score: 0.89

#136 - Similarity score: 0.88

#706 - Similarity score: 0.88

#715 - Similarity score: 0.87