Upload 3 files
Browse files- app.py +6 -0
- negbleurt.py +72 -0
- requirements.txt +1 -0
app.py
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import evaluate
|
2 |
+
from evaluate.utils import launch_gradio_widget
|
3 |
+
|
4 |
+
|
5 |
+
module = evaluate.load("negbleurt")
|
6 |
+
launch_gradio_widget(module)
|
negbleurt.py
ADDED
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
2 |
+
import datasets
|
3 |
+
import evaluate
|
4 |
+
|
5 |
+
|
6 |
+
_CITATION = """\
|
7 |
+
tba
|
8 |
+
"""
|
9 |
+
|
10 |
+
_DESCRIPTION = """\
|
11 |
+
Negation-aware version of BLEURT metric.
|
12 |
+
BLEURT a learnt evaluation metric for Natural Language Generation. It is built using multiple phases of transfer learning starting from a pretrained BERT model (Devlin et al. 2018) and then employing another pre-training phrase using synthetic data. Finally it is trained on WMT human annotations and the CANNOT negation awareness dataset.
|
13 |
+
"""
|
14 |
+
|
15 |
+
_KWARGS_DESCRIPTION = """
|
16 |
+
Calculates the NegBLEURT scores between references and predictions
|
17 |
+
Args:
|
18 |
+
predictions: list of predictions to score. Each prediction should be a string.
|
19 |
+
references: single reference or list of references for each prediction. If only one reference is given, all predictions will be scored against the same reference
|
20 |
+
batch_size: batch_size for model inference. Default is 16
|
21 |
+
Returns:
|
22 |
+
negBLEURT: List of NegBLEURT scores for all predictions
|
23 |
+
Examples:
|
24 |
+
>>> negBLEURT = evaluate.load('negbleurt')
|
25 |
+
>>> predictions = ["Ray Charles is a legend.", "Ray Charles isn’t legendary."]
|
26 |
+
>>> reference = "Ray Charles is legendary."
|
27 |
+
>>> results = rouge.compute(predictions=predictions, references=reference)
|
28 |
+
>>> print(results)
|
29 |
+
{'negBLERUT': [0.8409, 0.2601]}
|
30 |
+
"""
|
31 |
+
|
32 |
+
@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
|
33 |
+
class NegBLEURT(evaluate.Metric):
|
34 |
+
def _info(self):
|
35 |
+
return evaluate.MetricInfo(
|
36 |
+
description=_DESCRIPTION,
|
37 |
+
citation=_CITATION,
|
38 |
+
inputs_description=_KWARGS_DESCRIPTION,
|
39 |
+
features=[
|
40 |
+
datasets.Features(
|
41 |
+
{
|
42 |
+
"predictions": datasets.Value("string", id="sequence"),
|
43 |
+
"references": datasets.Sequence(datasets.Value("string", id="sequence")),
|
44 |
+
}
|
45 |
+
),
|
46 |
+
datasets.Features(
|
47 |
+
{
|
48 |
+
"predictions": datasets.Value("string", id="sequence"),
|
49 |
+
"references": datasets.Value("string", id="sequence"),
|
50 |
+
}
|
51 |
+
),
|
52 |
+
],
|
53 |
+
codebase_urls=["https://github.com/MiriUll/negation_aware_evaluation"]
|
54 |
+
)
|
55 |
+
|
56 |
+
def _download_and_prepare(self, dl_manager):
|
57 |
+
model_name = "tum-nlp/NegBLEURT"
|
58 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
59 |
+
self.model = AutoModelForSequenceClassification.from_pretrained(model_name)
|
60 |
+
|
61 |
+
def _compute(
|
62 |
+
self, predictions, references, batch_size=16
|
63 |
+
):
|
64 |
+
single_ref = isinstance(references, str)
|
65 |
+
if single_ref:
|
66 |
+
references = [references] * len(predictions)
|
67 |
+
|
68 |
+
scores_negbleurt = []
|
69 |
+
for i in tqdm(range(0, len(references), batch_size)):
|
70 |
+
tokenized = self.tokenizer(references[i:i+batch_size], candidates[i:i+batch_size], return_tensors='pt', padding=True, max_length=512, truncation=True)
|
71 |
+
scores_negbleurt += self.model(**tokenized).logits.flatten().tolist()
|
72 |
+
return {'negBLEURT': scores_negbleurt}
|
requirements.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
transformers~=4.25.1
|