File size: 2,314 Bytes
01cb32d 856850b 01cb32d 856850b 01cb32d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 |
## 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
```bash
pip install fact_checking
```
and then use the following code:
```python
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
```bash
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
```python
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
```bash
{'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.
|