File size: 3,386 Bytes
53c0c2a
 
 
 
 
 
fa86202
 
e0600a7
53c0c2a
15390e9
53c0c2a
 
 
 
 
 
99fd96c
 
 
53c0c2a
 
 
62d918e
 
 
 
 
 
53c0c2a
 
2b053b6
53c0c2a
b86f54f
 
53c0c2a
 
 
 
 
 
 
 
 
b86f54f
 
53c0c2a
 
7869c42
b86f54f
53c0c2a
db490b1
 
53c0c2a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: mit
datasets:
- web_nlg
language:
- en
widget:
- text: "Bourg-la-Reine is located in France and I love this town. I'm from People's Republic of China. [SEP] A Chinese, Loves, Bourg-la-Reine"
- text: "Bucharest is a city in Romania. [SEP] Romania | is located in | Bucharest"
---

# Model card for Inria-CEDAR/FactSpotter-DeBERTaV3-Small

## Model description

This model is related to the paper **"FactSpotter: Evaluating the Factual Faithfulness of Graph-to-Text Generation"**. 

Given a triple of format "subject, predicate, object" and a text, the model determines if the triple is present in the text. 

The delimiter can be ", " or " | ".

Different from the paper using ELECTRA, this model is finetuned on DeBERTaV3. 

We also provide Base and Large versions of this model:

https://huggingface.co/Inria-CEDAR/FactSpotter-DeBERTaV3-Base

https://huggingface.co/Inria-CEDAR/FactSpotter-DeBERTaV3-Large

## How to use the model

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification


def sentence_cls_score(input_strings, predicate_cls_model, predicate_cls_tokenizer):
    tokenized_cls_input = predicate_cls_tokenizer(input_strings, truncation=True, padding=True,
                                                  return_token_type_ids=True)
    input_ids = torch.Tensor(tokenized_cls_input['input_ids']).long().to(torch.device("cuda"))
    token_type_ids = torch.Tensor(tokenized_cls_input['token_type_ids']).long().to(torch.device("cuda"))
    attention_mask = torch.Tensor(tokenized_cls_input['attention_mask']).long().to(torch.device("cuda"))
    prev_cls_output = predicate_cls_model(input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids)
    softmax_cls_output = torch.softmax(prev_cls_output.logits, dim=1, )
    return softmax_cls_output


tokenizer = AutoTokenizer.from_pretrained("Inria-CEDAR/FactSpotter-DeBERTaV3-Small")
model = AutoModelForSequenceClassification.from_pretrained("Inria-CEDAR/FactSpotter-DeBERTaV3-Small")
model.to(torch.device("cuda"))

# pairs of texts (as premises) and triples (as hypotheses)
cls_texts = [("the aarhus is the airport of aarhus, denmark", "aarhus airport, city served, aarhus, denmark"),
             ("aarhus airport is 25.0 metres above the sea level", "aarhus airport, elevation above the sea level, 1174")]
cls_scores = sentence_cls_score(cls_texts, model, tokenizer)
# Dimensions: 0-entailment, 1-neutral, 2-contradiction
label_names = ["entailment", "neutral", "contradiction"]
```
## Citation
If the model is useful to you, please cite the paper

```
@inproceedings{zhang:hal-04257838,
  TITLE = {{FactSpotter: Evaluating the Factual Faithfulness of Graph-to-Text Generation}},
  AUTHOR = {Zhang, Kun and Balalau, Oana and Manolescu, Ioana},
  URL = {https://hal.science/hal-04257838},
  BOOKTITLE = {{Findings of EMNLP 2023 - Conference on Empirical Methods in Natural Language Processing}},
  ADDRESS = {Singapore, Singapore},
  YEAR = {2023},
  MONTH = Dec,
  KEYWORDS = {Graph-to-Text Generation ; Factual Faithfulness ; Constrained Text Generation},
  PDF = {https://hal.science/hal-04257838/file/_EMNLP_2023__Evaluating_the_Factual_Faithfulness_of_Graph_to_Text_Generation_Camera.pdf},
  HAL_ID = {hal-04257838},
  HAL_VERSION = {v1},
}
```

## Questions
If you have some questions, please contact through my email zhangkun@ieee.org or kun.zhang@inria.fr