The fake news story refers to information falsified or made up to manipulate or deceive the public. It is a serious issue that undermines the credibility of information sources and influences public opinion, discourse, and important decisions and outcomes. The content and source of news articles and stories must be considered when identifying and classifying fake news. Detecting fake news is an important and active research area with many potential applications and implications for society. Many challenges are involved in determining whether a news article is authentic. Several deep learning models that employ natural language processing (NLP) have shown excellent results in detecting fake news. To assess the truthfulness of news articles, our methodology is based on state-of-the-art language models based on transformers. Bidirectional Encoder Representation from Transformers (BERT) and Robustly Optimized BERT Technique (RoBERTa) is one of the most advanced models. Our findings reveal that the BERT model achieved an accuracy of 64%, while the RoBERTa model slightly outperformed it with an accuracy of 66%. These results are particularly significant when compared to similar research in this domain, which reported a maximum accuracy of 62% for both models on Liar dataset.

Social media like Twitter have become very popular in recent decades. Hashtags are new kind of metadata which make non-structured tweets into searchable semistructured content. There are varied previous methods which recommend hashtags for new tweets. However, to the best of our knowledge, there is no previous word that uses BERT embedding for this purpose. In this paper, we propose a new method called EmHash that uses neural network based on BERT embedding to recommend new hashtags for each tweet. Unlike other word embeddings, BERT embedding constructs different vectors for the same word in different contexts. Emhash succeeded in outperforming three methods LDA, SVM, and TTM with respect to recall measure.

Community Question-Answering websites, such as StackOverflow and Quora, expect users to follow specific guidelines in order to maintain content quality. These systems mainly rely on community reports for assessing contents, which has serious problems, such as the slow handling of violations, the loss of normal and experienced users' time, the low quality of some reports, and discouraging feedback to new users. Therefore, with the overall goal of providing solutions for automating moderation actions in Q&A websites, we aim to provide a model to predict 20 quality or subjective aspects of questions in QA websites. To this end, we used data gathered by the CrowdSource team at Google Research in 2019 and fine-tuned pre-trained BERT model on our problem. Based on our evaluation, model achieved value of 0. 046 for Mean-Squared-Error (MSE) after 2 epochs of training, which did not improve substantially in the next ones. Results confirm that by simple fine-tuning, we can achieve accurate models in little time and on less amount of data.

Intent classification is one of the important tasks in natural language understanding which aims to classify queries based on their intent, goal, or purpose which is implied in the content. However, the problem in this task is the lack of data labeled by a human agent. This problem leads to weakness in the generalization of models, especially when the models face rare words. Using pre-trained models can be useful in the generalization of language representation. BERT pre-trained language model which has been currently published has made a major impact in the natural language processing field. This model which is trained on unlabeled large-scale corpora, with fine-tuning could achieve state-of-art results in various natural language processing tasks e. g. question answering and sentiment analysis. In this paper, we compared the BERT model with conventional machine learning models and shown that the BERT model has a better performance than conventional machine learning models in few-shot learning.