Motivations, Methods and Metrics of Misinformation Detection: An NLP Perspective

Abstract

ive summarization is also a relevant task that can be useful for facilitating misinformation detection. Specifically, the summarization model can be applied to identify the central claims of the input texts and serves as a feature extractor prior to misinformation detection. For example, Esmaeilzadeh et al. [24] use a text summarization model to first summarize an article and then input the summarized sequences into a RNN-based neural network to do misinformation detection. The experimental results are compared against the task using only the original texts, and finally demonstrate higher performance. Fact checking is the task of assessing the truthfulness of claims especially made by public figures such as politicians [25]. Usually, there is no clear distinction between misinformation detection and fact checking since both of them aim to assess the truthfulness of claims, thoughmisinformation detection usually focuses on certain pieces of information while fact checking is broader [26]. However, fact checking can also be a relevant task of misinformation detection when a piece of information contains claims that need to be verified as true or false. Rumor detection is often confused with fake news detection, since rumor refers to a statement consisting of unverified information at the posting time. Rumor detection task is then defined as separating personal statements into rumor or nonrumor [27]. Thus, rumor detection can also serve as another relevant task of misinformation detection to first detect worth-checking statements prior to classifying the statement as true or false. This can help mitigate the impact that subjective opinions or feelings have on the selection of statements that need to be further verified. Sentiment analysis is the task of extracting emotions from texts or user stances. The sentiment in the true and misrepresented information can be different, since publishers of misinformation focus more on the degree to impress the audience and the spreading speed of the information. Thus, misinformation typically either contains intense emotion which could easily resonate with the public, or Q. Su et al. / Natural Language Processing Research 1(1-2) 1–13 3 controversial statements aiming to evoke intense emotion among receivers. Thus, misinformation detection can also utilize emotion analysis through both the content and user comments. Guo et al. [28] propose a Emotion-based misinformation Detection framework to learn contentand comment-emotion representations for publishers and users respectively so as to exploit content and social emotions simultaneously for misinformation detection. 1.3. An Overview of the Survey This survey aims to present a comprehensive review on studying misinformation in terms of its characteristics and detection methods. It first introduces the related concepts and highlights the significance of misinformation detection. It then uses a two-dimensional model to decompose this task: the internal dimension of descriptive analysis (i.e., the characterization of low-credibility information) and the external dimension of predictive modeling (i.e., the automatic detection of misinformation). In particular, the publicly available datasets and the state-of-the-art technologies are reviewed in terms of the detection approaches, feature representations and model construction. Finally, challenges of misinformation detection are summarized andnewprospects are provided for futuremisinformation detection works.