The Limitations of Stylometry forDetecting Machine-Generated Fake News

The Limitations of Stylometry for Detecting Machine-Generated Fake News

Abstract

Recent developments in neural language models (LMs) have raised concerns about their potential misuse for automatically spreading misinformation. In light of these concerns, several studies have proposed to detect machine-generated fake news by capturing their stylistic differences from human-written text. These approaches, broadly termed stylometry, have found success in source attribution and misinformation detection in human-written texts. However, in this work, we show that stylometry is limited against machine-generated misinformation. While humans speak differently when trying to deceive, LMs generate stylistically consistent text, regardless of underlying motive. Thus, though stylometry can successfully prevent impersonation by identifying text provenance, it fails to distinguish legitimate LM applications from those that introduce false information. We create two benchmarks demonstrating the stylistic similarity between malicious and legitimate uses of LMs, employed in auto-completion and editing-assistance settings.1 Our findings highlight the need for non-stylometry approaches in detecting machine-generated misinformation, and open up the discussion on the desired evaluation benchmarks.

\issue

xxx2020

1 Introduction

Many previous studies on stylometry—the extraction of stylistic features from written text—showed promising results on text classification. Two of stylometry’s common applications are: (1) Detecting the provenance of text (i.e. identifying the author) in order to prevent impersonations (Tweedie et al., 1996; Brennan et al., 2012; Afroz et al., 2014; Caliskan-Islam et al., 2015; Neal et al., 2017; Sari et al., 2017); (2) Detecting misinformation in text due to deception (Enos et al., 2007; Mihalcea and Strapparava, 2009; Ott et al., 2011; Feng et al., 2012; Afroz et al., 2012), fake news (Rashkin et al., 2017; Pérez-Rosas et al., 2018), or other false or illegal content (Choshen et al., 2019). In the former, the classifier identifies language features that correlate with a specific person or group. The latter, misinformation detection, relies on idiosyncrasies of lies, i.e. style and language characteristics that are unique to text that is false or misleading.

Stylometry has recently gained attention as a potential answer to concerns that LMs could be used to mass-produce malicious text (Vosoughi et al., 2018; Radford et al., 2019), that (1) impersonates a human author’s text and/or (2) is fallacious and misleading. Indeed, stylometry-based approaches have shown promising results for defending against human-impersonating LMs (Bakhtin et al., 2019; Zellers et al., 2019). However, as applications of text generation such as text auto-completion (Wolf et al., 2019) become widely used (P. House (2019); 12), labeling text as generated by a LM might not indicate anything at all about its trustworthiness. This motivates our core inquiry subject: can stylometry be used to distinguish malicious uses of LMs from legitimate ones?

We build the first benchmark for detection of LM-produced fake news that labels text as “real” or “fake” according to its veracity. Inspired by studies on deceitful behaviors in humans, showing that people try to diverge as little as possible from the truth when lying (Mazar et al., 2008), we focus on automatic false modifications or additions to otherwise truthful news stories.

Our datasets contain articles produced by both malicious and responsible uses of language models, and the detector’s task is to identify the malicious ones. In one dataset, we produce text by prompting a LM to extend news articles with relevant claims. We simulate malicious user, who only accepts the LM’s suggestion if the claim is factually false, and a responsible user, who only accepts correct claims. The produced sentences are short and concise statements, similarly to fake news and false claims as represented in human-generated datasets (Wang, 2017; Augenstein et al., 2019). In another dataset, we modify existing news articles to include false information by inverting article statements. In this case, the LM is used to automatically identify the most plausible edit locations. This is similar to (mis-)using an autocorrect tool that suggests local modifications.

We find that with the state-of-the-art stylometry-based classifier, even a single auto-generated sentence within a wall of human-written text is detectable with high accuracy, yet the truthfulness of a single sentence remains largely undecidable. Moreover, even a relatively weak LM can be used to produce statement inversions that the state-of-the-art stylometry-based model cannot detect. Thus, stylometry fails to distinguish malicious from responsible behaviors. This indicates that, unlike humans who expose stylistic cues when writing false content (Ott et al., 2011; Frank et al., 2008; Matsumoto et al., 2011), LMs maintain consistency for both true and false content. Worse yet, while a provenance classifier can effectively detect a potentially malicious author or publication venue (given enough examples), it might not distinguish malicious from legitimate authors if they are both using the same LM to generate their text. In this regard, malicious text generated by a LM might actually be harder to detect than hand-crafted malicious text.

Our human evaluation tests show that humans are also fooled by machine-generated misinformation, but that access to external information sources can help. Therefore, we recommend future research on machine-generated misinformation to focus on non-stylometry strategies. Finally, we discuss what benchmarks are required for evaluating the performance of such detectors.

2 Background and Related Work

Stylometry for Human-written Text

The use of statistical methods for analyzing human-written documents has been studied extensively since the early days of the field. One common application is provenance detection. For example, Mosteller and Wallace (1963) used word counts to predict the authorship of historical documents. Tweedie et al. (1996) extracted other stylistic features and applied a neural network to the same task. While these classifiers could be fooled by an aware writer that intentionally imitates other’s style (Brennan et al., 2012), this approach was found useful for de-anonymizing cybercriminals in forums (Afroz et al., 2014), identifying programmers (Caliskan-Islam et al., 2015) and more (Neal et al., 2017). In a related line of study, stylometry was applied to, rather than detecting the specific person, identifying characteristics of the author, such as age and gender (Goswami et al., 2009), political views (Potthast et al., 2018), or IQ (Abramov and Yampolskiy, 2019). Recently, as detailed later below, stylometry was used to distinguish machine- from human-writers.

Another common application of stylometry is detecting human-written misinformation. Mihalcea and Strapparava (2009) found specific words that are highly correlated with true and false statements. Ott et al. (2011); Feng et al. (2012) used a richer set of features such as POS tags frequencies and constituency structure to deceptive writing. Following these observations and the increasing interest in fake news, recent studies applied stylometry on entire news articles (Pisarevskaya, 2017), short news reports (Pérez-Rosas et al., 2018), fact and political statements (Nakashole and Mitchell, 2014; Rashkin et al., 2017) and posts in social media (Volkova et al., 2017). The success of these studies is mostly attributed to stylistic changes in human language when lying or deceiving (Bond and Lee, 2005; Frank et al., 2008). In this work, we evaluate the viability of this approach on machine-generated text, where stylistic differences between truth and lie might be more subtle.

Machine-generated Text Detection

Detecting text’s provenance is similar to authorship attribution and, therefore, stylometry can be effective. Indeed, Gehrmann et al. (2019) show the existence of distributional differences between human-written texts and machine-generated ones by visualizing the per-token probability according to a LM. Bakhtin et al. (2019) learn a dedicated provenance neural classifier. While their classifier achieves high in-domain accuracy, they find that it overfits the generated text distribution rather than detecting outliers from human texts, resulting in increased “human-ness” scores for random perturbations. Nevertheless, an advantage of such neural approaches over more traditional stylometry methods is that, given enough data, the model learns hidden stylistic representations without the need to manually define any features.

Building on the above observation, Zellers et al. (2019) focus on fake news and create a Transformer-based LM (Vaswani et al., 2017) dubbed Grover and train it on a large news corpus. Grover also includes a “neural fake news detector”, a linear classifier on top of the hidden state of the last token of the examined article, fine-tuned to classify if the news text was machine-generated or not. The experiments in this paper are based on the Grover-Mega classifier, fine-tuned for the target task (see section 3).

Fake News Detection Approaches Beyond Stylometry

The other most extensively studied NLP-based approach for fake news detection is based on fact-checking. This approach have recently gained increasing attention thanks to several synthetic (Thorne et al., 2018) and real-world datasets (Hanselowski et al., 2018; Wang, 2017; Popat et al., 2018; Augenstein et al., 2019). The performance of current models is still far from humans (Thorne et al., 2019; Schuster et al., 2019), but with their advancements they can still play a positive role in detection.

Another line of work for fake news detection utilizes non-textual information such as how content is propagated, by which users, its originating URL, and other metadata (Castillo et al., 2011; Gupta et al., 2014; Zhao et al., 2015; Kochkina et al., 2018; Liu and Wu, 2018), as well as incorporating users’ explicit feedback, such as abuse reporting Tschiatschek et al. (2018). Social network platforms, ISPs, and even individual users can employ such methods to moderate content exposure. These approaches are beset by the challenges Shu et al. (2017) of noisy and incomplete data, especially given the need to detect fake news early Liu and Wu (2018) (before propagation and user engagement patterns are fully formed).

3 Adversarial Setting

“Fake News”: Our Working Definition

Our attackers focus on automatically introducing false information into otherwise trustworthy content. We call the resultant manipulated articles “fake news”. This definition matches the one of Zhou and Zafarani (2018) and is in line with the disinformation focused view of Wardle and Derakhshan (2017). Also, this is in line with how false claims are represented in many human-generated fake news datasets (Wang, 2017; Augenstein et al., 2019). Conversely, Zellers et al. (2019) focus on entirely fabricated articles, a different type of fake news where the goal is mostly to create “viral and persuasive” content.

Our choice of creating articles with only a limited number of false statements is aligned with the way humans tend to deceive or lie. Psychological studies (Mazar et al., 2008) support the age-old notion that, when lying, “the best policy for the criminal is to tell the truth as nearly as possible” [Raskolnikov, “Crime and Punishment” Dostoyevsky (1866)]. This helps preserve an honest self-image and, perhaps more importantly, reduces the chances of the lie being detected. For example, a study on the longest-surviving known fake Wikipedia article (O. Benjakob (2019); 22) revealed that many of the presented facts were only slightly altered from other, true facts.

Figure 1: Examples of the fake class in our experiments. (a) In the news question answering (Section 4), a CNN article is presented with two examples of questions (bold) from newsQA (Trischler et al., 2017) and Grover’s generated answer (red). The first answer is verified by a human annotator to be false and the second as true. (b) In article modification () (Section 4), the negations are marked with a cross-line for deletions and underline for addition. (c) In the article extension case (Section 5), the bold red text is the generation of GPT-2 Medium to extend the prefix.

Attack and Defense Capabilities

We adopt an adversarial setting similar to that of Zellers et al. (2019). Our attacker wishes to generate fake text, that contains unverified or false claims, en masse, using a language model to automate the process. The attacker’s goal is to produce fake text that the verifier classifies as real. Our verifier is adaptive: it receives a limited set of examples generated by the attacker, and trains a discriminator to detect the attacker’s texts from legitimately-produced, real text, containing exclusively human-verified claims (news articles from relatively reputable sources, like the The New York Times are assumed to be real). We also experiment with a non-adaptive, zero-shot setting, where the verifier does not receive the attacker’s examples.

Training and evaluating the detector

In each experiment, we collected a dataset with a “real” text class and a “fake” text class and used separate samples for testing and for fine-tuning. We used a Grover-Mega discriminator for all of the experiments. The model’s weights were initialized from a checkpoint provided by Zellers et al. (2019) and fine-tuned for 10 epochs with our training samples. For evaluating the zero-shot defense, we applied a pretrained Grover-Mega discriminator by querying its Web interface. We report human performance on some of the attacks.

4 Stylometry Fails to Detect Machine-Generated Misinformation

We create two datasets, simulating two different uses of LMs to automatically produce fake news. In the first, the extension scenario, an auto-completion text generator extends a news article. A responsible user of this generator verifies the correctness of the output (producing real text), whereas an attacker verifies incorrectness (producing fake text). In the second, the modification scenario, the attacker uses a human-written news article and performs subtle modifications to semantically modify statements. Specifically, we add and remove negations. This follows the intuition that, if we take care to add negations in a syntactically-correct fashion, the new sentence is a negative inversion of the original (Rudanko, 1982). Yet, such changes are subtle enough to retain the original article’s style and distributional features. See below for the full details on the creation of the datasets.

Additionally, we used about 100 examples from each dataset to test human performance in detecting this form of misinformation. For the extension scenario, we assigned two subject volunteers with two different tasks. The first had the same task as Grover’s detector, and the second was allowed to use external sources to verify facts (which Grover cannot do). For the modification scenario, the participants had the same task as Grover but we highlighted negations to help them focus on the relevant parts.

(1) Creating the Extension Dataset

We use the newsQA dataset (Trischler et al., 2017) that contains Cable News Network (CNN) articles with corresponding questions and answers (segments copied verbatim from the text). Following Radford et al. (2019) and von Davier (2019), we embed the question in a template that is appended to the end of the input text:
We attempt to answer: <question>
Answer:

Then, the first sentence produced by Grover’s generator is used as a potential answer. We choose this template by examining common formats of questions in news articles and finding the one that empirically produces mostly reasonable answers by Grover’s generator. For example, given text about a suspect caught in Washington D.C., and the question “Where was the suspect stopped?”, Grover answered “2 blocks from the U.S. Capitol.”

We manually labeled the generated answers. We filtered out nonsensical ones (29%) and labeled the rest as real or fake texts by correctness. Answers containing at least one false statement (judging by the article and/or external knowledge) were labeled fake. We measured inter-annotator agreement on a subset of 100 examples labeled by two different annotators, and found it to be substantial (Cohen’s kappa score of ).

Finally, we remove the sentence containing the answer from each article by removing the sentence with the highest TF-IDF-weighted word-count similarity with the question and answer (concatenated).

(2) Creating the Modifications Dataset

We use articles from The New York Times (Thompson, 2017). We perform the modifications by removing and adding negations from statements: first, we randomly delete “not” or “no” occurrences from the text. Then, we automatically find statements to add negations to, using the probabilities of a GPT-2 Medium language model. We randomly sample 100 locations in the article and choose the ones with the maximal score, defined as the probability for either “not” or “no”, multiplied by the probability for the word following the negation. By both adding and removing, we keep the total number of negations unchanged, avoiding anomalous negations counts. Original NYT articles are labeled as real and modified ones as fake.

4.1 Results

Machine-generated misinformation precision recall F1 accuracy
adaptive QA extension (false vs. true) 0.72 0.71 0.71 71%
modification () 0.53 0.52 0.53 53%
modification () 0.66 0.65 0.65 65%
modification () 0.73 0.47 0.63 65%
Table 1: Results: Section 4. We report (macro) F1 score and overall accuracy, as well as precision and recall of the “fake” class. Zero-shot performance (not included) was very low in all cases.

The detector’s performance figures are given in Table 1. Against QA extension, the detector reached 71% accuracy in distinguishing true vs. false extensions generated by Grover. While this is above a trivial majority baseline (51%), some of the success could be a side effect of the way that the dataset was constructed. Labeling instances with at least one false statement as fake makes these sentences generally longer, allowing a simple linear classifier that only uses the extension’s length to reach 56%. When evaluated only on the short false answers (up to 10 words), the accuracy drops from 71% to only 62%. Ultimately, however, despite performing better than simple baselines, the detection in this setting is much worse than in distinguishing human from machine text (Section 5).

Comparing this with the results of the human evaluation is particularly instructive: the first subject, given the same task as the detector, attained an F1 score of 0.68, only marginally below the detector. The second participant, who was allowed to use external sources, attained an F1 of 0.84. We conclude that the stylometry-based classifier doesn’t perform much better than humans in detecting potential misinformation, and that verifying against other resources can drastically improve results.

For the automatic article modification attack, the classifier fails completely to detect two modifications (), and is not much better even when statements are inverted. Even though the change includes negations that might introduce surprising factual claims, the language style remains too similar to detect it. Humans, evaluated with , had an F1 of 0.74, with 0.91 and 0.59 recall for real and modified articles, respectively. Thus, while humans perform better than the model, many of the modified statements were interpreted as real. Manually examining these, we found that the added negation actually changed the meaning in 60% of the instances that were misclassified.

Overall, the results show that both attacks can produce articles containing misinformation that evade the stylometry detector and mislead humans.

5 Stylometry Detects Machine-Human Impersonations

Provenance detection precision recall F1 accuracy
zero-shot full article 0.84 0.98 0.90 90%
vanilla extension () 0.52 0.20 0.45 51%
vanilla extension () 0.07 0.01 0.28 37%
adaptive full article 0.93 0.94 0.94 94%
vanilla extension () 0.90 0.97 0.95 95%
vanilla extension () 0.91 0.95 0.94 95%
QA extension (machine vs. human) 0.82 0.86 0.83 83%
Table 2: Detection results on the attacks of Section 5 in a zero-shot setting and the adaptive setting (where the discriminator is fine-tuned to a specific attacker). We report (macro) F1 score and overall accuracy. Precision, and recall of the “fake” class are reported as well.

One might suspect that the low performance in Section 4 is due to limited capacity of our detector or the small fraction of generated text. We now show that the detector performs well on the provenance task and is sensitive to differences in small text portions.

(1) Fully-generated Articles

We perform a first evaluation of Grover against texts generated by a different model of similar size, namely GPT-2 XL (Radford et al., 2019). We include examples from the WebText test set (Radford et al., 2019), labeled real text, and the released unconditioned generations of the GPT-2 XL model, labeled fake text.

(2) Automatic Article Extension (Vanilla)

In order to experiment with mixed human-machine articles, we build an attacker that automatically extends unfinished human-written articles. We simulate this attacker by querying a GPT-2 Medium model, conditioned on the the first 500 words from an article. We include New York Times articles, truncated to match the length of the fake class, labeled as real text, and articles that were automatically extended, labeled fake text. We use different percentages () of machine-generated text, by stopping the generation after an appropriate number of sentences.

(3) Automatic QA-based Extension

In this experiment, we test whether the QA defense from Section 4 can perform better if we can assume that true texts are always human-written. To this end, we repeat the same setting but use the gold answers from the newsQA dataset for the “real text” class.

5.1 Results

Zero-shot Setting

As Table 2 shows, the zero-shot classifier is effective in detecting the fully generated articles of a different model, with a 0.9 F-1 score. The extended articles, containing a substantial amount of human-written text, are mostly classified as human-written in this setting. This is unsurprising considering the dataset of full articles the detector was originally trained on.

Adaptive Setting

After fine-tuning, Grover improves on the full articles and performs well against article extension generations, reaching 0.94 F1 score for articles with a single generated sentence (). This stylometry detector is thus highly effective in distinguishing human from machine.

In the QA setting, also containing a single machine-generated sentence, the detector performs worse by 0.11 F1 points. This might be due to using the (stronger) Grover-Mega generator for the QA setting, and because the appended template for QA was selected to empirically maximize human “reasonableness” scores. Still, this setting allows the stylometry detector to reach much higher performance than in the veracity-based setting (Section 4). This indicates that a restricted benchmark, that assumes no legitimate use of LMs, might not reflect the misinformation detection performance if the model is deployed in a world where LMs are used for both legitimate and malicious purposes.

6 Discussion

Advancements in LM technology and their various applications have introduced a new challenge: distinguishing truthful text from misinformation, when the text is generated or edited by a LM. Our experiments indicate that LM-generated falsified texts are very similar in style to LM-generated texts containing true content. As a result, stylometry-based classifiers cannot identify auto-generated intentionally misleading content.

We conclude with the following recommendations:

(1) Extending Veracity-based Benchmarks. In order to better evaluate detectors against LM-generated misinformation, we recommend extending our benchmarks by creating other veracity-oriented datasets, that represent a wide range of LM applications, from whole-article generation to forms of hybrid writing and editing.

(2) Improving Non-stylometry Methods. Other detection approaches, as surveyed at the end of Section 2, are less affected by the use of LMs. Therefore, advancements in such methods can improve the detection of both human- and machine-generated misinformation. Notably, the fact-checking setting makes fewer assumptions on the available auxiliary information and can be applied even if the text was sent to the verifier through a private channel such as E-mail. However, since fact-checking requires advanced inference capabilities, incorporating non-textual information, when available, can yield better results.

Conclusion

The potential use of LMs in creating fake news calls for a re-evaluation of current defense strategies. We examine the state-of-the-art stylometry model, and find it effective in preventing impersonation, but limited in detecting LM-generated misinformation. This new kind of misinformation could be created by the same model that is used by legitimate writers as a writing-assistance tool, hiding stylistic differences between falsified and truthful content. This motivates (1) constructing more instructive benchmarks for NLP-based approaches and improving non-stylistic methods, and (2) addressing a set of challenges that spans many disciplines beyond NLP, including social networks, information security, human-computer interaction, and others.

Acknowledgements

We thank the anonymous reviewers and the members of the MIT NLP group for their helpful comments. R.S is a member of the Check Point Institute of Information Technology. This work is supported in part by Google’s TensorFlow Research Cloud program; DSO grant DSOCL18002; by the Blavatnik Interdisciplinary Cyber Research Center (ICRC); by the NSF award 1650589; and by the generosity of Eric and Wendy Schmidt by recommendation of the Schmidt Futures program.

\starttwocolumn

Footnotes

  1. Data: https://people.csail.mit.edu/tals/publication/are_we_safe/

References

  1. Automatic iq estimation using stylometric methods. In Handbook of Research on Learning in the Age of Transhumanism, pp. 32–45. Cited by: §2.
  2. Detecting hoaxes, frauds, and deception in writing style online. 2012 IEEE Symposium on Security and Privacy, pp. 461–475. Cited by: §1.
  3. Doppelgänger finder: taking stylometry to the underground. 2014 IEEE Symposium on Security and Privacy, pp. 212–226. Cited by: §1, §2.
  4. MultiFC: a real-world multi-domain dataset for evidence-based fact checking of claims. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, pp. 4685–4697. External Links: Link, Document Cited by: §1, §2, §3.
  5. Real or fake? learning to discriminate machine from human generated text. arXiv preprint arXiv:1906.03351. Cited by: §1, §2.
  6. The fake nazi death camp: wikipedia’s longest hoax, exposed. Note: Haaretz\urlhttps://tinyurl.com/y3m2wd9k Cited by: §3.
  7. Language of lies in prison: linguistic classification of prisoners’ truthful and deceptive natural language. Applied Cognitive Psychology 19 (3), pp. 313–329. External Links: Document, Link, https://onlinelibrary.wiley.com/doi/pdf/10.1002/acp.1087 Cited by: §2.
  8. Adversarial stylometry: circumventing authorship recognition to preserve privacy and anonymity. ACM Transactions on Information and System Security (TISSEC) 15 (3), pp. 12. Cited by: §1, §2.
  9. De-anonymizing programmers via code stylometry. In 24th USENIX Security Symposium (USENIX Security 15), Washington, D.C., pp. 255–270. External Links: ISBN 978-1-931971-232, Link Cited by: §1, §2.
  10. Information credibility on twitter. In Proceedings of the 20th International Conference on World Wide Web, New York, NY, USA, pp. 67 – 684. External Links: ISBN 9781450306324, Link, Document Cited by: §2.
  11. The language of legal and illegal activity on the Darknet. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 4271–4279. External Links: Link, Document Cited by: §1.
  12. Doc. doc product medical q&a with deep language models. Note: \urlhttps://github.com/re-search/DocProduct Cited by: §1.
  13. Crime and punishment. The Russian Messenger. Cited by: §3.
  14. Detecting deception using critical segments. In International Speech Communication Association, pp. 2281–2284. Cited by: §1.
  15. Syntactic stylometry for deception detection. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Short Papers - Volume 2, ACL ’12, Stroudsburg, PA, USA, pp. 171–175. External Links: Link Cited by: §1, §2.
  16. Human behavior and deception detection. Wiley handbook of science and technology for homeland security, pp. 1–13. Cited by: §1, §2.
  17. GLTR: statistical detection and visualization of generated text. In Proceedings of the 57th Conference of the Association for Computational Linguistics: System Demonstrations, Florence, Italy, pp. 111–116. External Links: Link Cited by: §2.
  18. Stylometric analysis of bloggers’ age and genderStylometric analysis of bloggers’ age and gender. In International AAAI Conference on Web and Social Media, External Links: Link Cited by: §2.
  19. Tweetcred: real-time credibility assessment of content on twitter. In International Conference on Social Informatics, pp. 228–243. Cited by: §2.
  20. A retrospective analysis of the fake news challenge stance-detection task. In Proceedings of the 27th International Conference on Computational Linguistics, pp. 1859–1874. Cited by: §2.
  21. I, language robot. Note: Los Angeles Review of Books\urlhttps://lareviewofbooks.org/article/i-language-robot/ Cited by: §1.
  22. Hox. list of hoaxes on wikipedia. Note: \urlhttps://en.wikipedia.org/wiki/Wikipedia:List_of_hoaxes_on_Wikipedia Cited by: §3.
  23. All-in-one: multi-task learning for rumour verification. In Proceedings of the 27th International Conference on Computational Linguistics, Santa Fe, New Mexico, USA, pp. 3402–3413. External Links: Link Cited by: §2.
  24. Early detection of fake news on social media through propagation path classification with recurrent and convolutional networks. In AAAI Conference on Artificial Intelligence, External Links: Link Cited by: §2.
  25. Evaluating truthfulness and detecting deception. FBI L. Enforcement Bull. 80, pp. 1. Cited by: §1.
  26. The dishonesty of honest people: a theory of self-concept maintenance. Journal of marketing research 45 (6), pp. 633–644. Cited by: §1, §3.
  27. The lie detector: explorations in the automatic recognition of deceptive languageThe lie detector: explorations in the automatic recognition of deceptive language. In Proceedings of the ACL-IJCNLP 2009 Conference Short Papers, USA, pp. 309 – 312. Cited by: §1, §2.
  28. Inference in an authorship problem. Journal of the American Statistical Association 58 (302), pp. 275–309. External Links: ISSN 01621459, Link Cited by: §2.
  29. Language-aware truth assessment of fact candidates. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Baltimore, Maryland, pp. 1009–1019. External Links: Link, Document Cited by: §2.
  30. Surveying stylometry techniques and applications. ACM Comput. Surv. 50, pp. 86:1–86:36. Cited by: §1, §2.
  31. Finding deceptive opinion spam by any stretch of the imagination. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, Portland, Oregon, USA, pp. 309–319. External Links: Link Cited by: §1, §1, §2.
  32. Automatic detection of fake news. In Proceedings of the 27th International Conference on Computational Linguistics, pp. 3391–3401. Cited by: §1, §2.
  33. Deception detection in news reports in the Russian language: lexics and discourse. In Proceedings of the 2017 EMNLP Workshop: Natural Language Processing meets Journalism, Copenhagen, Denmark, pp. 74–79. External Links: Link, Document Cited by: §2.
  34. DeClarE: debunking fake news and false claims using evidence-aware deep learning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, pp. 22–32. External Links: Link, Document Cited by: §2.
  35. A stylometric inquiry into hyperpartisan and fake news. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia, pp. 231–240. External Links: Link, Document Cited by: §2.
  36. Language models are unsupervised multitask learners. OpenAI Blog 1 (8). Cited by: §1, §4, §5.
  37. Truth of varying shades: analyzing language in fake news and political fact-checking. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 2931–2937. Cited by: §1, §2.
  38. Towards a description of negatively conditioned subject operator inversion in english. English Studies 63 (4), pp. 348–359. Cited by: §4.
  39. Continuous n-gram representations for authorship attribution. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, Valencia, Spain, pp. 267–273. External Links: Link Cited by: §1.
  40. Towards debiasing fact verification models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, pp. 3419–3425. External Links: Document, Link Cited by: §2.
  41. Fake news detection on social media: a data mining perspective. ACM SIGKDD Explorations Newsletter 19 (1), pp. 22–36. Cited by: §2.
  42. All the news dataset. Note: Kaggle\urlhttps://www.kaggle.com/snapcrack/all-the-news Cited by: §4.
  43. Evaluating adversarial attacks against multiple fact verification systems. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, pp. 2944–2953. External Links: Link, Document Cited by: §2.
  44. The fact extraction and verification (fever) shared task. In Proceedings of the First Workshop on Fact Extraction and VERification (FEVER), pp. 1–9. Cited by: §2.
  45. NewsQA: a machine comprehension dataset. In Proceedings of the 2nd Workshop on Representation Learning for NLP, Vancouver, Canada, pp. 191–200. External Links: Link, Document Cited by: Figure 1, §4.
  46. Fake news detection in social networks via crowd signals. In Companion Proceedings of the The Web Conference 2018, pp. 517–524. Cited by: §2.
  47. Neural network applications in stylometry: the federalist papers. Computers and the Humanities 30, pp. 1–10. Cited by: §1, §2.
  48. Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008. Cited by: §2.
  49. Separating facts from fiction: linguistic models to classify suspicious and trusted news posts on twitter. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), Vancouver, Canada, pp. 647–653. External Links: Link, Document Cited by: §2.
  50. Training optimus prime, md: generating medical certification items by fine-tuning openai’s gpt2 transformer model. arXiv preprint arXiv:1908.08594. Cited by: §4.
  51. The spread of true and false news online. Science 359 (6380), pp. 1146–1151. External Links: Document, ISSN 0036-8075, Link, https://science.sciencemag.org/content/359/6380/1146.full.pdf Cited by: §1.
  52. “liar, liar pants on fire”: a new benchmark dataset for fake news detection. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), Vancouver, Canada, pp. 422–426. External Links: Link, Document Cited by: §1, §2, §3.
  53. Information disorder: toward an interdisciplinary framework for research and policy making. Cited by: §3.
  54. HuggingFace’s transformers: state-of-the-art natural language processing. ArXiv abs/1910.03771. Cited by: §1.
  55. Defending against neural fake news. In Advances in Neural Information Processing Systems 32, H. Wallach, H. Larochelle, A. Beygelzimer, F. d\textquotesingleAlché-Buc, E. Fox and R. Garnett (Eds.), pp. 9054–9065. External Links: Link Cited by: §1, §2, §3, §3, §3.
  56. Enquiring minds: early detection of rumors in social media from enquiry posts. In Proceedings of the 24th International Conference on World Wide Web, pp. 1395–1405. Cited by: §2.
  57. Fake news: a survey of research, detection methods, and opportunities. arXiv preprint arXiv:1812.00315. Cited by: §3.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
408876
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description