metadata

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-Large

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 Small models

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

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

How to use the model

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-Large")
model = AutoModelForSequenceClassification.from_pretrained("Inria-CEDAR/FactSpotter-DeBERTaV3-Large")
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 [email protected] or [email protected]