A zero-shot relation extractor
Project description
Fact checking
This generative model - trained on FEVER - aims to predict whether a claim is consistent with the provided evidence.
Installation and simple usage
One quick way to install it is to type
pip install fact-checking
and then use the following code:
from transformers import (
GPT2LMHeadModel,
GPT2Tokenizer,
)
from fact_checking import FactChecker
_evidence = """
Justine Tanya Bateman (born February 19, 1966) is an American writer, producer, and actress . She is best known for her regular role as Mallory Keaton on the sitcom Family Ties (1982 -- 1989). Until recently, Bateman ran a production and consulting company, SECTION 5 . In the fall of 2012, she started studying computer science at UCLA.
"""
_claim = 'Justine Bateman is a poet.'
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
fact_checking_model = GPT2LMHeadModel.from_pretrained('fractalego/fact-checking')
fact_checker = FactChecker(fact_checking_model, tokenizer)
is_claim_true = fact_checker.validate(_evidence, _claim)
print(is_claim_true)
which gives the output
True
Probabilistic output with replicas
The output can include a probabilistic component, obtained by iterating a number of times the output generation. The system generates an ensemble of answers and groups them by Yes or No.
For example, one can ask
from transformers import (
GPT2LMHeadModel,
GPT2Tokenizer,
)
from fact_checking import FactChecker
_evidence = """
Jane writes code for Huggingface.
"""
_claim = 'Jane is an engineer.'
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
fact_checking_model = GPT2LMHeadModel.from_pretrained('fractalego/fact-checking')
fact_checker = FactChecker(fact_checking_model, tokenizer)
is_claim_true = fact_checker.validate_with_replicas(_evidence, _claim)
print(is_claim_true)
with output
{'Y': 0.95, 'N': 0.05}
Score on FEVER
The score on the FEVER dev dataset is as follows
precision | recall | F1 |
---|---|---|
0.94 | 0.98 | 0.96 |
These results should be taken with many grains of salt. This is still a work in progress, and there might be leakage coming from the underlining GPT2 model unnaturally raising the scores.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for fact_checking-0.0.1-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 90b6ca0be9fae36cb50c857bac4a012bb5fa1d2a689d80cc24f3b9d58e0d0cf5 |
|
MD5 | 2ef469f89a80d87cbfcbf9ff8380d294 |
|
BLAKE2b-256 | c052e4764db89fdea5380083d97a04e0278426c0a4a2b2a018c5439d1ffa06ee |