acred models and data

Models and data needed to run and reproduce results for the acred credibility review system, as described in our paper accepted at the International Semantic Web Conference 2020.

Models:

• semantic_encoder.zip is a RoBERTa base model fine-tuned on STS-B for encoding sentences for semantic similarity
• saved_fnc1_classifier is a RoBERTa base model with sentence-pair classifier fine-tuned on FNC-1 for stance detection
• check_worthiness is a RoBERTa base model with sentence classifier fine-tuned on a variety of datasets to predict the checkworthiness of a sentence

Data:

• claims-from-ClaimReviews-45K.csv provides 45K sentences for which we have found ClaimReviews, which serve as ground credibility signals
• sentences-extractedFrom-Articles-40K.csv provides 40K sentences extracted from a variety of news websites, which serve as ground credibility signals
• claim_dev_embs_85K_20200426.tar.gz provides embeddings for the 85K sentences in the csv files
• claimReviews-pruned-45K.jsonl provides a pruned version of the original claimReview. In particular, we do not include the full text of hte claimReview, but only the url, author information and main rating.