/fact_checking

Primary LanguageJupyter NotebookMIT LicenseMIT

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

False

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 predictions are evaluated on a subset of the FEVER dev dataset, restricted to the SUPPORTING and REFUTING options:

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.