Scholarlink

The fake news running rampant in this modern day, it’s vital to catch it to dismiss the myth. Traditional analytics like Bag of Words (BoW) and TF-IDF could not capture any relationship that exists between words and hence is able to extract minimal context only. Deep learning methods such as BERT and Word2Vec can comprehend deeper context now, but in so doing, it requires word-level context along with a sequence dependency to be formed by models. The improved fake news detection by applying a BERT + LSTM framework on the LIAR dataset, giving detailed accuracy ratings of political statements. The benchmark of BERT + LSTM model against the previously used approaches such as BoW, Word2Vec, and GloVe and have shown that BERT + LSTM can capture nuances in word and sequence relationships. The results show that proposed method achieves higher accuracy of 62.9, precision of 63.94, recall of 91.29, and F1-scores of 75.69 than the traditional methods.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing Fake News Detection in Political Claims: A High-Recall BERT and LSTM Approach with the LIAR Dataset

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Saurabh Kumar Deepak,
  • Sanyam Shukla,
  • R. K. Pateriya,
  • Saurabh Shrivastava

摘要

The fake news running rampant in this modern day, it’s vital to catch it to dismiss the myth. Traditional analytics like Bag of Words (BoW) and TF-IDF could not capture any relationship that exists between words and hence is able to extract minimal context only. Deep learning methods such as BERT and Word2Vec can comprehend deeper context now, but in so doing, it requires word-level context along with a sequence dependency to be formed by models. The improved fake news detection by applying a BERT + LSTM framework on the LIAR dataset, giving detailed accuracy ratings of political statements. The benchmark of BERT + LSTM model against the previously used approaches such as BoW, Word2Vec, and GloVe and have shown that BERT + LSTM can capture nuances in word and sequence relationships. The results show that proposed method achieves higher accuracy of 62.9, precision of 63.94, recall of 91.29, and F1-scores of 75.69 than the traditional methods.