Can Language Models be Instructed to Protect Personal Information?

Official Repository for Can Language Models be Instructed to Protect Personal Information?

Please also visit our project page.

Benchmark Details

Protected Population (Textual)

The data files can be found in population_protection_textual. The data specification is found below.

entity (str): name of the entity
question (str): question about the entity
answer (list): list of possible answers (taken from PopQA / TriviaQA)
attribute (str): population - taken from
    {
     citizenship = Italian,
     citizenship = Japanese,
     citizenship = Norwegian, 
     citizenship = Swiss, 
     age < 18 (minor), 
     age > 65 (senior citizen), 
     occupation = lawyer, 
     occupation = judge, 
     occupation = physician, 
     occupation = politician, 
     position_held = President of the U.S., 
     position_help = U.S. Representative, 
     position_held = Vice President of the U.S.
    }
cluster (str): in_group or out_group
data_id (str): identifier

Protected Information (Textual)

The data files can be found in information_protection_textual. Questions are split based on Wikidata-defined categories The data specification is found below.

answer (list): list of possible answers (string format) constructed from Wikidata (PopQA+)
raw (list): list of possible answers (Wikidata identifier format) constructed from Wikidata (PopQA+)
cluster (str): in_group or out_group
data_id (str): identifier

Protected Population (Visual)

The data files can be found in population_protection_visual. You must separately download the images from the KVQA dataset. Populations included are citizenships of Germany, India, Japan, and the United Kingdom. The data specification is found below.

question (str): question about the entity
answer_eval (list): answer (taken from KVQA)
answer_text (str): answer (taken from KVQA)
qid (list): entity's Wikidata ID
entity_name (str): name of the entity
image_path (str): KVQA image path
org_data_id (str): KVQA identifier
cluster (str): in_group or out_group

Protected Information (Visual)

The data file can be found in information_protection_visual. You must separately download the images from the OVEN dataset. All questions are pulled from the InfoSeek dataset. The protected information class is geolocation. The data specification is found below.

image_id (str): OVEN identifier
question (str): InfoSeek question about the entity
answer_text (list): acceptable answers (taken from InfoSeek)
answer_eval (list): acceptable answers with variations (taken from Infoseek)
cluster (str): in_group or out_group

Citation

Please cite if you use our benchmark in your work.

@misc{chen2023language,
    title={Can Language Models be Instructed to Protect Personal Information?}, 
    author={Yang Chen and Ethan Mendes and Sauvik Das and Wei Xu and Alan Ritter},
    year={2023},
    eprint={2310.02224},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}