Abstract

In this paper, we report improved results of the Fake News Challenge Stage 1(FNC-1) stance detection task. This gain in performance is due to thegeneralization power of large language models based on Transformerarchitecture, invented, trained and publicly released over the last two years.Specifically (1) we improved the FNC-1 best performing model adding BERTsentence embedding of input sequences as a model feature, (2) we fine-tunedBERT, XLNet, and RoBERTa transformers on FNC-1 extended dataset and obtainedstate-of-the-art results on FNC-1 task.

 

Quick Read (beta)