PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Natural Language Processing for Fake News Detection

Ray Oshikawa, Jing Qian, William Yang Wang

arXiv:1811.00770v2cs.CLcs.AI

TL;DR

Fake news detection is a critical NLP problem involving the assessment of news truthfulness, with manual verification limited by time and cost. The paper systematically reviews task formulations, datasets, NLP approaches, and empirical results. It finds strong performance in several settings while emphasizing dataset-quality and scope limitations and recommending more detailed, fair, and practical models.

  • Problem

    Automated fake news detection matters because news misinformation has substantial social and political impacts, while manual claim verification is time-consuming and expensive.

  • Method

    The paper provides a systematic survey comparing task definitions, datasets, model construction, performances, NLP approaches, and evidence-collection methods.

  • Results

    LSTM-based models generally outperform CNN-based models on LIAR, while metadata such as verdict reports and speaker credibility improves accuracy.

  • Takeaways & Limitations

    The survey recommends fine-grained veracity prediction, appropriate use of non-textual data, combining hand-crafted and neural features, and broader content verification.

  • Takeaways & Limitations

    Dataset construction can confound fake-news detection with website classification, and machine annotation of entire articles relies on strong assumptions.

Abstract

from arXiv · show

Fake news detection is a critical yet challenging problem in Natural Language Processing (NLP). The rapid rise of social networking platforms has not only yielded a vast increase in information accessibility but has also accelerated the spread of fake news. Thus, the effect of fake news has been growing, sometimes extending to the offline world and threatening public safety. Given the massive amount of Web content, automatic fake news detection is a practical NLP problem useful to all online content providers, in order to reduce the human time and effort to detect and prevent the spread of fake news. In this paper, we describe the challenges involved in fake news detection and also describe related tasks. We systematically review and compare the task formulations, datasets and NLP solutions that have been developed for this task, and also discuss the potentials and limitations of them. Based on our insights, we outline promising research directions, including more fine-grained, detailed, fair, and practical detection models. We also highlight the difference between fake news detection and other related tasks, and the importance of NLP solutions for fake news detection.

1. Introduction

The survey frames automated fake news detection as a critical NLP problem and systematically reviews its task definitions, datasets, models, and results. It also identifies future research directions and introduces NLP solutions to broader verification challenges.

  • Motivation: Automated fake news detection assesses the truthfulness of news claims, addressing socially consequential misinformation that can affect politics and public safety.Manual verification is time-consuming and expensive, motivating automated methods for internet content.
  • Scope: The survey systematically compares fake news detection task definitions, datasets, model construction, and empirical performances.It also discusses the pros, cons, pitfalls, and drawbacks of these tasks and methods.
  • Contributions: The paper presents the first comprehensive review of NLP solutions for automatic fake news detection.This contribution establishes the survey’s broad coverage of the research area.
  • Contributions: It analyzes how fake news detection aligns with existing NLP tasks and discusses assumptions and notable issues across problem formulations.The analysis helps clarify how different formulations shape the detection problem.
  • Contributions: It categorizes datasets, NLP approaches, and results while providing accessible introductions for researchers entering the field.The survey is intended to summarize available resources and findings for new researchers.

2. Related Problems

The paper distinguishes fake news detection from fact-checking, rumor detection, stance detection, and sentiment analysis, while situating these tasks alongside relevant datasets. These distinctions clarify whether systems assess veracity, verification status, debate position, or emotion.

  • Fact-Checking: Fact-checking assesses the truthfulness of claims made by public figures, whereas fake news detection usually focuses more narrowly on news events.The paper notes that researchers often treat the two tasks similarly because both assess claim truthfulness.
  • Rumor Detection: Rumor detection lacks a consistent definition, but one survey defines it as separating personal statements into rumor or nonrumor.Rumors are described as unverified information at posting time that can be verified rather than subjective opinions or feelings.
  • Datasets: Table 1 summarizes various fake news detection-related datasets and identifies Facebook with the abbreviation FB.The table is presented as a dataset overview rather than a single-task benchmark.
  • Datasets: Table 2 gives an example LIAR entry and orders credibility-history counts as barely true, false, half true, mostly true, and pants on fire.These history counts include only inaccurate statements.
  • Stance Detection: Stance detection assesses which side of a debate an author takes, measuring consistency rather than claim veracity.It can serve as a subtask for searching documents for evidence in fake news detection.
  • Sentiment Analysis: Sentiment analysis extracts personal emotions such as favorable or unfavorable impressions, unlike fake news detection’s objective claim verification.The distinction is based on whether the task analyzes emotions or evaluates truthfulness.

3. Task Formulations

The survey formulates fake news detection primarily as classification, while also considering regression and the practical difficulties of representing partial truth and discrete labels. These formulations determine the input, output, and evaluation choices used by detection systems.

  • Task Formulations: Fake news detection identifies false stories that appear to be news, with text inputs ranging from short statements to entire articles.Additional information, such as speaker identity, may also be appended depending on the dataset.
  • Task Formulations: Most studies formulate fake news detection as classification or regression, with classification used more frequently.The general goal is to identify fake news through these predictive task formulations.
  • Classification: Binary classification labels news as fake or real, but partial truth motivates adding intermediate or multiple truthfulness classes.Multi-class outputs treat truthfulness degrees as independent labels learned from the data.
  • Regression: Regression formulates detection as predicting a numeric truthfulness score.This formulation differs from classification by producing a continuous-valued output.
  • Regression: Regression evaluation can use prediction differences from ground-truth scores or Pearson/Spearman correlations.The approach is used by Nakashole and Mitchell (2014).
  • Regression: Discrete ground-truth labels create a challenge for regression because they must be converted into numeric scores.The conversion from discrete labels to numeric values is a central issue in this formulation.

4. Datasets

The survey organizes fake-news datasets into claims, entire articles, and social-networking-service data, while emphasizing challenges in dataset size, quality, labeling, and topical scope.

  • Dataset categories: Fake-news datasets are categorized as claims, entire articles, or social-networking-service data with structured account, post, and non-text information.Claims contain one or a few verifiable sentences, whereas articles contain related sentences forming a whole.
  • Claims: Early manually labeled claim datasets were small, with 221 statements in the first public dataset and 300 claims in EMERGENT, limiting machine-learning assessment.LIAR later expanded claim data to 12,836 six-grade-labeled statements with speaker, subject, party, and context information.
  • Claims: FEVER contains 185,445 Wikipedia-derived claims labeled Supported, Refuted, or Not Enough Info, with cited sentences as evidence.Its evidence structure enables systems to predict claim truthfulness together with supporting evidence.
  • Articles and SNS data: Entire-article and social-networking datasets combine article content, engagement information, or publisher and account data, but their labels and coverage vary.FAKENEWSNET includes headlines, bodies, and Twitter engagements; some SNS datasets rely on browser-generated or publisher-identity labels.
  • SNS data: SNS datasets tend to cover limited topics and may be less related to news, constraining their suitability for fake-news detection.
  • SNS data: Twitter datasets such as PHEME and CREDBANK are limited by narrow topics or non-news tweets and are therefore used more often for rumor detection.CREDBANK covers 60 million tweets across 1,049 events, while PHEME contains 330 threads from nine events.

5. Methods

The surveyed methods preprocess text into lexical, linguistic, embedding, rhetorical, or metadata features and apply supervised classifiers, neural models, or evidence-retrieval approaches suited to dataset structure.

  • Preprocessing: Preprocessing commonly includes tokenization, stemming, word weighting, TF-IDF, LIWC, and pretrained word embeddings such as word2vec and GloVe.Entire articles may additionally require identifying central claims from raw text.
  • Classification models: Most existing research uses supervised learning, while semi-supervised and unsupervised methods are less common.
  • Classification models: SVM and Naive Bayes frequently serve as baseline classifiers, while logistic regression and random forests are also used.
  • Neural models: LSTM models capture longer-term dependencies and often perform strongly on LIAR and FEVER, while CNNs are used for text and metadata feature extraction.Some LSTM models combine word embeddings or LIWC vectors with other representations.
  • Rhetorical methods: RST analyzes rhetorical coherence and semantic roles, sometimes with VSM, to identify essential ideas and structural characteristics associated with fake news.
  • Evidence collection: RTE-based methods gather sentences supporting or opposing a claim and predict correctness, but require datasets containing textual evidence such as FEVER.

6. Results & Observations

Across LIAR, FEVER, and FAKENEWSNET, results favor attention and metadata or social-engagement information, while performance depends strongly on task formulation and available evidence.

  • Evaluation scope: The survey compares empirical results on LIAR, FEVER, and FAKENEWSNET because other datasets are smaller, less studied, or more focused on rumor detection.
  • LIAR: On LIAR, LSTM-based models generally outperform CNN-based models, and adding annotator verdict reports or larger speaker-credibility data raises accuracy.The reported increases are 4% with verdict reports and 21% when replacing credibility history with speaker2credit.
  • FEVER: On FEVER, attention-LSTM achieves the best score for both verification and evidence-collection tasks among the discussed models.The workshop-leading method selects evidence through semantic matching between retrieved-page sentences and the claim.
  • FAKENEWSNET: On FAKENEWSNET, RST- and LIWC-based models exceed 60% accuracy but perform below methods using social-engagement information.
  • FAKENEWSNET: Social-engagement-only modeling defeats text-only RST and LIWC modeling on FAKENEWSNET, while hybrid and graph-based methods use engagement or relationship data.The survey attributes very high accuracy to successful utilization of additional data.

7. Discussions, & Recommendations

The survey recommends more nuanced datasets and veracity labels, while cautioning that source composition and annotation choices can confound evaluation. It also identifies opportunities and risks in textual, neural, and non-textual detection methods.

  • Datasets and Inputs: Dataset construction should account for fake-news types, language and culture, delivery manner, timeframe, writing, length, and ground-truth verifiability.The survey expands nine corpus requirements and notes that satire can be distinguished from real and fake news through style analysis.
  • Recommendations: Fine-grained truthfulness labels are more realistic than binary labels because news can mix true and false statements.Existing fact-checking systems use labels such as MOSTLYFALSE, HALFTRUE, and Mixture, while binary classification can exceed 90% accuracy but 6-way text-only classification remains below 30%.
  • Recommendations: Ordinal learning could distinguish severe errors from neighboring truthfulness levels, improving the practicality of multi-class detection.Treating labels as independent classes makes judging True as False equivalent to judging True as Mostly True.
  • Datasets and Inputs: Sampling all true or fake items from one situation or publisher introduces confounding variables and can turn detection into website classification.The survey therefore recommends collecting claims or articles from varied speakers and publishers within the dataset scope.
  • Methods: Hand-crafted features, attention-based neural models, and metadata each offer potential benefits, but metadata must be used carefully.Attention and LSTM methods capture long-term contextual information, while speaker or publisher signals may silence minority voices; the source also notes a celebrity-reporting limitation for GossipCop.

8. Conclusion

The conclusion summarizes the survey’s analysis of fake-news detection definitions, datasets, and experimental methods, then proposes directions for future models. These directions include examining feature combinations, using non-textual information appropriately, and extending verification through content.

  • Conclusion: The survey compares benchmark datasets and experimental results across automatic fake-news detection methods.
  • Conclusion: Future research should investigate combining hand-crafted features with neural-network models and using non-textual data appropriately.
  • Conclusion: The survey recommends extending fake-news verification through content rather than relying only on classification.
Loading 1811.00770v2…