Fake news detection: a survey of evaluation datasets

Characteristics of evaluation datasets used in fake news detection

To clearly understand and discriminate among the evaluation datasets for fake news detection, different types of features can be extracted from the descriptions of the 27 surveyed datasets that are described in the Supplemental Article S1. To identify the relevant features that characterize the datasets, we started from the fake news characterization introduced by Zhang & Ghorbani (2019). Their work is of special interest to us since they introduced a clear characterization of online fake news by identifying relevant features related to the users, content, and context that can be adapted to characterize also the datasets used in fake news detection. Specifically, Zhang & Ghorbani (2019) defined four major components to characterize fake news (see Fig. 2A), which are:

(a) the creator/spreader, who can be either a real human (benign authors and publishers who publish fake news unintentionally, and fake news creators) or non-human (cyborgs, social bots);

(b) the target victims, who can be users of online social media or other mainstreaming platforms;

(c) the news content, which can be either physical (e.g., headline, body text, video/image) or non-physical (e.g., main purpose, sentiment, news topics);

(d) the social context, which can be related both to the platforms used to spread the news (social media or mainstreaming platforms) and to the distribution pattern (community of users or broadcast pattern).

Having analyzed the fake news characterization by Zhang & Ghorbani (2019), we selected some of their components and extended this with other ones, to fit the purpose of our work to analyze the features of datasets instead of fake news in general.

Specifically, we checked the descriptions of the datasets provided in the 27 papers, retrieved as illustrated in the “Survey methodology” section, with the aim of searching for whether these descriptions include some of the components defined by Zhang & Ghorbani (2019). This analysis showed that the following three components were explicitly mentioned in the descriptions of the 27 datasets (depicted in Fig. 2A with a colored background): the news content, the social context, and the creator/spreader. On the contrary, the target victims component (Zhang & Ghorbani, 2019) cannot be extracted from the descriptions of the datasets as it is never explicitly mentioned.

Specifically, considering the news content (Zhang & Ghorbani, 2019), the descriptions of the 27 surveyed datasets provide information about both the physical content and the non-physical content (main purpose and news topics), while very few datasets collect information about the sentiment polarity of the news (i.e., positive, negative, neutral). Consequently, our FNDD characterization will include the following three features in the news content component: News content type, type of disinformation, and news domain (depicted in Fig. 2B with a colored background). The type of disinformation and the news domain correspond, respectively, to the main purpose and news topics in the characterization of Zhang & Ghorbani (2019). We changed their names to better represent what the authors of the 27 surveyed datasets refer to. In addition to these characteristics, we added two more features to this component (i.e., the language of the news and the rating scale used to assess its truthfulness) since the authors of the 27 datasets describe the datasets also according to these features and we considered it relevant to characterize them.

Turning to the social context (Zhang & Ghorbani, 2019), the aforementioned descriptions provide information about the media platforms (e.g., Facebook, Twitter) used to share and spread the news collected in the datasets, while no information about how the news is distributed is given. Therefore, we considered only one feature in this component referring to the media platform.

Finally, referring to the creator/spreader (Zhang & Ghorbani, 2019), the authors of the 27 datasets provide information about the spontaneity of the fake news editing (i.e., if fake news items are collected without editing them or have been produced by manipulating real ones). Therefore, we considered only one feature in this component referring to the spontaneity.

In addition to that information, the authors of the 27 datasets mentioned four further features of the datasets that are not addressed by Zhang & Ghorbani (2019); these are the size of the dataset, the application purpose in building the dataset, the online availability of the dataset and the extraction period when the data have been collected. Consequently, we added a new component to our FNDD characterization, named general information, which refers to the datasets' general characteristics, including the size, the application purpose, the availability, and the extraction period.

Based on these considerations, we evolve the characteristics proposed by Zhang & Ghorbani (2019) further by proposing the characterization of the datasets of fake and trustful news depicted in Fig. 2B and described below. The re-arranged characteristics are depicted with a colored background, while the newly defined ones are indicated with a white background.

Specifically, our FNDD characterization includes the eleven dataset characteristics described below:

- News domain: the dataset can contain fake news items that target certain news domains, such as health, education, tourism, sport, economy, security, science, IT, and political election.

- Application purpose: datasets can be built for different aims, such as fake detection, fact-checking, veracity classification, and rumor detection. The first consists of the prediction of the chances of a particular piece of information (news article, reviews, posts, etc.) being intentionally deceptive. Fact-checking refers to the process of vetting and verifying factual statements contained in a piece of information; unlike fake detection, fact-checking works at the level of a particular statement or claim. Veracity classification is very similar to fake detection but attempts to predict the actual truth of a given piece of information. Finally, rumor detection tries to distinguish between verified and unverified information (instead of true or false) and the unverified information may turn out to be true or false, or may remain unresolved.

- Type of disinformation: fake news or misleading news may be categorized as fake reviews, fake advertisements, or fake news articles according to the types of false information they contain. Fake news articles can be further classified into (a) hoaxes, considered to be false information deliberately fabricated to masquerade as the truth (Sharma et al., 2019); (b) rumors, which refers to unsubstantiated claims that are disseminated without any evidence to support them (Sharma et al., 2019); and (c) satire, which uses humor, irony, exaggeration, ridicule, and false information to present the news.

- Language: this concerns the language of the fake news contained in the dataset, which can be written in different languages according to the sources used to retrieve them. The dataset can be categorized as multi-lingual or monolingual.

- Size: the size of the dataset is commonly determined by the number of news items that it contains. It can also be measured in kilobytes/megabytes of the overall archive.

- News content type: this concerns linguistic and syntactic features, such as headlines and body text, as well as images and videos of the news. This survey considers the following four types of news content (as they characterize the 27 datasets): headlines, body text, images, and videos.

- Rating scale: this concerns the labels that are associated with the news contained in the datasets and used to rate the truthfulness of the news. Different rating scales can be used containing a different number of rating levels. For instance, a five-point rating scale can have the following labels: true, mostly true, mixture of true and false, mostly false, and false. An example of a three-point rating scale is true, false, or unverified.

- Media platform: this concerns the digital environment where the news collected in the dataset is shared and spread to the audience. Two main types of media platforms can be used to share and transfer fake news: (i) mainstream media, meaning the traditional media, such as newspapers, TV, and radio, and (ii) online social media, such as Twitter, Facebook, Instagram, and blogs. In this paper, mainstream media include also traditional media (e.g., NBC News, Washington Post, etc.) that have extended the way they spread information from mainstream platforms also to digital platforms.

- Spontaneity: fake news in the dataset may be spontaneous if it is automatically extracted from public web sources without editing, or artificial if it is manually generated by asking someone to produce items by manipulating real ones.

- Availability: this concerns free online availability of the data contained in the dataset.

-Extraction period: this concerns the definition of a specific time frame during which the data have been collected.

A summary of the datasets selected for further analysis in the present review, along with their main characteristics, is shown in Table 2. A brief description of each identified dataset is provided in the Supplemental Article S1.

Quantitative analysis of surveyed datasets

In this section, a quantitative analysis of the 27 surveyed datasets in terms of the eleven characteristics of our FNDD characterization (i.e., news domain, application purpose, type of disinformation, language, size, news content type, rating scale, media platform, spontaneity, availability, and extraction period), is provided by analyzing the frequency distribution of their occurrences.

The temporal distribution of dataset publications, shown in Fig. 3, underscores the increasing interest of the scientific community in the topic of fake news detection datasets, which started growing in 2016 and continues to grow in 2019.

Attending to the news domain (see Fig. 4A), the majority of the surveyed datasets collected fake news items concerning politics (55.6%) and society (44.4%), followed by technology (14.8%), economy (7.4%), science (7.4%) and crime (7.4%). Only one dataset collected news on security, health, tourism, sport, education, entertainment, fraud and scam, or fauxtography (3.7%). The high level of attention to the political topic is probably due to the recent studies that demonstrated the significant influence that fake news can exert on political debate (Flynn, Nyhan & Reifler, 2017; Allcott & Gentzkow, 2017), even affecting electoral outcomes during an election cycle. Moreover, a study conducted by Vosoughi, Roy & Aral (2018) founds that political fake news reaches more people and is more viral than false information in any other domain.

The most frequent application purpose of the surveyed datasets is fake detection (55.6%), followed by fact-checking (18.5%), veracity classification (11.1%) and rumor detection (11.1%), and clickbait detection (3.7%) as depicted in Fig. 4B. Fake detection is the most general application purpose and includes the whole process of classifying any types of false information as true or false; this generality makes it the most common area when collecting datasets of fake information. Fact-checking is also quite widely applied, although it is more specific than fake detection as it focuses on the specific step of the fake detection process that consists of collecting the relevant combination of facts related to a claim from different web sources to assign a truth value to the claim. Rumor detection and veracity classification have a more specific and limited task than fake detection, as the former is devoted to determining information that is spreading but yet to be verified (rumor) and the latter aims to determine whether a detected rumor can be classified as true, false, or still unverified. Their specificity and limited scope are the main reasons for which rumor detection and veracity classification are applied less than fake detection. Similarly, clickbait detection has a specific and limited task since it is devoted to determining the use of sensational headlines that attract users to click but are misleading and provocative.

The type of disinformation contained in the datasets (see Fig. 4C) consists mainly of fake news articles (63%), followed by rumors (14.8%), fake reviews and hoaxes (7.4%), satire and clickbait (3.7%). Fake news articles are the most general type of disinformation that is used mainly to spread false information through news outlets for political or financial gain (Zubiaga et al., 2018). This generality makes fake news articles the most common type of disinformation in the surveyed datasets. As suggested by Wardle (2017), the term fake news is being used to refer to different types of disinformation, including also satire, hoaxes, and rumors, which are, by their nature, more specific and, consequently, less frequently addressed by fake news detectors.

Considering the language, 22 of the 27 described datasets are monolingual with news in English (81.5%), while the Spanish fake news corpus is monolingual in Spanish, and Zheng et al. (2018) dataset is monolingual with news in Chinese. Only three datasets (11.1%) are multi-lingual. English, indeed, is the primary language of the Web, and a lot of linguistic tools and fake detection methods have been developed for the English language. Consequently, also the data collected for fake news detection are predominantly written in English.

Considering the size of the surveyed datasets, the total number of news articles or reviews they contain has been taken into account. The graph depicted in Fig. 5A shows this value in ascending order for 22 of the 27 datasets. Five datasets (CREDBANK, Tam et al. (2019) dataset, NELA-GT-2018, CNN/Daily Mail summarization dataset, and FEVER), which have the highest number of collected articles, are depicted in Fig. 5B since we have used a larger scale for the y axis. The size of the dataset plays an important role in ensuring a high accuracy of the fake detection process. In particular, if the dataset is used to train a fake news detection method that is based on machine learning, it is fundamental to have a large dataset because the performance of this kind of method improves as the training dataset size increases (Torabi & Taboada, 2019). The negative aspect is that very large datasets are less reliable using manual annotation due to time consumption and misclassification (Ghiassi & Lee, 2018). This is why the majority of the surveyed datasets are relatively small, with less than 15,000 news articles.

As concerns the type of news content, 24 of the 27 datasets collected only the text of the news (88.9%), while FakeNewsNet collects also the images included in the news, FVC-2018 collects also the videos, and Verification Corpus collects both images and videos in addition to the text. This fact is probably due to the widespread use of deception detection methods that use natural language processing techniques that depend heavily on text data only (refer to Oshikawa, Qian & Wang (2018) and Su et al. (2020) for a comprehensive survey on these techniques).Only recently have researchers started incorporating images by developing multimodal fake news detection methods, although these still suffer from the scarcity of labeled data due to the labor-intensive process of annotating them.

Considering the rating scale (see Fig. 4D), the majority (51.9%) of the datasets used a two-point rating scale to rate the truthfulness of the fake news items, followed by a four-point rating scale (22.2%), three-point rating scale (14.8%), five-point rating scale (7.4%), and six-point rating scale (3.7%). The predominant use of the binary classification is due to the higher accuracy obtained by two-point prediction compared to six-point classification, which is over 90% for the former and less than 30% for the latter, as found by Oshikawa, Qian & Wang (2018). This is because classifying False news as Mostly False is considered a mistake like classifying True news as False, so a six-point classification method has a higher probability of making mistakes.

The media platforms used to share and transfer fake news items collected in the surveyed datasets are mainly mainstream media (48.1%), followed by online social media (37%), mainly using Facebook or Twitter, and the integration of these two platforms (14.8%), as depicted in Fig. 4E.

Considering the spontaneity, all the described datasets are spontaneous, except FEVER and Ott et al. (2011) dataset, which are annotated by human annotators and the AMT crowdsourcing service, respectively, to manipulate real news items for generating deceptive ones.

As concerns the availability, the survey included only datasets that were either publicly available or explicitly available for research purposes upon request. In the eligibility step of the PRISMA methodology, described in the “Survey methodology” section, 34 publications were excluded because they described datasets that are not publicly available.

Finally, about half of the datasets (48.1%) were extracted within defined extraction period. This allows the extraction of more homogeneous news with less differentiation and more consistent linguistic features, as Kwon, Cha & Jung (2017) found. To summarize, the datasets can be classified according to the characteristics introduced above. Tables 3–12 provide the clusters obtained by applying these criteria and they can be useful when searching for a dataset having specific characteristics.

Comparative analysis of surveyed datasets

As evidenced in the systematic literature search, many datasets exist and are suitable for evaluating fake news detection methods. It is necessary for researchers to make a comparative analysis of those datasets in order to decide which one to use according to the purpose of their research. To provide this analysis, we need to find parameters on which to compare the characteristics (defined in our FNDD characterization in “Characteristics of evaluation datasets used in fake news detection”) of the datasets described in the Supplemental Article S1.

The main objective of this section is therefore to provide a set of dataset requirements for comparing dataset characteristics and to show how each of the surveyed datasets scores against them. To do this, we need to establish which types of requirements are treated by each dataset.

With this aim, we started from a set of nine requirements defined by Rubin, Chen & Conroy (2015) (see Fig. 6A) that fake news detection datasets should satisfy if they are to be fruitfully applied for modeling and testing fake news detection algorithms. We extended and re-arranged these requirements to fit all the characteristics defined in our FNDD characterization in the “Characteristics of evaluation datasets used in fake news detection” section. As shown in Table 13, indeed, the matching between Rubin et al.’s requirements and our dataset characteristics resulted in several characteristics that are not addressed in the requirements (i.e., type of disinformation, news domain, application purpose). We therefore added new requirements that address these missing features, as shown in Table 14.

We introduced the following four categories of requirements for classifying the ten requirements considered in our study:

• Homogeneity requirements: a dataset is homogeneous if it is made up of news items that are equal or similar to each other according to some characteristics. For example, in the news domain, a dataset is homogeneous if all the news items collected are chosen because they are frogman equal or similar news domain.

• Availability requirements: these require that some characteristics associated with the dataset are available. For instance, considering the language, a dataset is multilingually available if the news is collected in more than one language;

• Verifiability requirements: these allow the veracity of the news collected in the dataset to be verified;

• Temporal requirements: they allow some temporal restrictions over the collected news to be specified.

For each of these categories and each of the characteristics introduced in the “Characteristics of evaluation datasets used in fake news detection” section, we have re-arranged the requirements proposed by Rubin et al. by proposing the set of requirements depicted in Fig. 6B.

A brief description of the ten dataset requirements is provided in the remainder of this section.

1. homogeneity of news length: this implies having datasets with news items that are comparable in length;

2. homogeneity in news domain: this implies having a corpus with texts that are aligned with the topics (e.g., science, government, technology);

3. homogeneity in type of disinformation: this implies having a corpus with texts that are aligned with the type of disinformation (e.g. fake reviews, rumors, fake articles, etc.);

4. homogeneity in application purpose: the aim for which the dataset has been created (e.g., rumor detection, fake deception, fact-checking, etc.) supports interpretation of the situation;

5. availability of both fake and real news: the availability of both fake and real news allows algorithms to find patterns and regularities and, therefore, to improve the identification rate;

6. textual format availability: the availability of texts and textual transcriptions of audio and video is preferred when using algorithms based on natural language processing;

7. public availability: the corpus is publicly available;

8. multi-lingual availability: the availability of news written in multiple languages;

9. verifiability of ground truth: the ability to verify if the news is clearly genuine or fabricated;

10. belonging to a predefined time frame: this implies that the dataset has been collected within a set timeframe.

In Table 15, the public availability requirement is the criterion for selecting the surveyed datasets, while the temporal requirement concerns the period when the data of the datasets are collected; these are relevant in giving more information about the context and the historical period of the considered dataset.

According to the requirements described above, we have checked the fulfillment of those requirements for each dataset. Table 15 provides the results of our comparison.

As Table 15 underlines, none of the surveyed datasets satisfies all the considered requirements. The BuzzFace, LIAR, Benjamin Political News, BuzzFeed News, NELA-GT-2018 and TW_info datasets satisfy all the cited requirements except for multilinguality. PHEME, FVC-2018, and Verification Corpus are the only three multi-lingual databases since they include: conversational threads in English and German; in English, Russian, Spanish, Arabic, and German; and in English, Spanish, Dutch, French, respectively. PHEME satisfies all the considered requirements except for the timeframe because the corpus has not been collected within a defined timeframe. Meanwhile, FCV-2018 and Verification Corpus satisfy all the considered requirements except for the homogeneity in news length because they contain tweets, images, and videos.

Six of the ten requirements are satisfied by all the analyzed datasets, namely: availability of both fake and real news, textual format availability, verifiability of ground-truth, homogeneity in type of disinformation, homogeneity in application purpose, and public availability.

Table 15 underlines that 13 of the 27 surveyed datasets cover a predefined time frame, while the others have been extracted according to the application purpose, as Table 16 specifies.

Analyzing the requirement forth verifiability of ground-truth in Table 17, the majority of the surveyed datasets are fact-checked by researchers (41%), followed by journalists (22%), editors (7%), trained annotators, both human (7%) and artificial (4%), editors and journalists (4%), workers (4%), assessment sites (4%), assessment sites and human annotators (4%), and by journalists and crowdsourcing (3%). This underlines that most of the authors of the surveyed databases directly verified if the news is genuine or fabricated, as Fig. 7A underlines.

Considering the homogeneity in news length requirement in Table 18, the majority of the surveyed datasets are composed of articles (41%), while 15% of them are composed of Tweets and 11% by Facebook (FB) posts, followed by short claims (7%), statements (7%), reviews (4%), and headlines (4%). 11% of the surveyed datasets do not contain data with homogeneous news length, as Fig. 7B shows.

Figure 7C makes clear that most of the surveyed datasets do not satisfy the requirements of homogeneity in the news domain (41%), or are collected for a political purpose (26%), followed by societal (22%), tourism (4%), scientific (4%), and technology (3%) purposes.

Considering the homogeneity in the type of disinformation requirement in Table 19, Fig. 7D shows that most of the surveyed datasets have been created with fake articles (63%), followed by rumors (15%), fake reviews (7%), misinformation (4%), satire (4%), clickbait (4%), and hoaxes (3%).

Finally, the datasets are compared considering the homogeneity in the application purpose. Fig. 7E shows that most of the surveyed datasets have been created for fake detection (56%), followed by fact-checking (18%), veracity classification (11%) and rumor detection (11%), and clickbait detection (4%) purposes.

Open challenges

The quantitative and comparative analyses we have discussed in the previous sections allow us to identify several challenges faced by researchers when defining and developing new datasets for fake news detection.

- multimedia datasets: To develop multimedia datasets for fake news detection is one of the major open challenges. Few attempts to incorporate non-textual components of the news, like images and videos, have yet been made. As mentioned before, the only surveyed dataset that collects both images and text is FakeNewsNet. However, it suffers from the scarcity of labeled data (having less than 500 samples) due to the laborious process of annotating them, and this limits its contribution to fake news research. More recently, Nakamura, Levy & Wang (2020) proposed Fakeddit, a large-scale multimodal fake news dataset consisting of over 1 million samples containing text, image, metadata, and comments. Jindal, Vatsa & Singh (2019) also provided a benchmark dataset containing text and images for training and testing models for fake news detection, named NewsBag. These two recent datasets testify to the emerging need to cope with multimedia data, going beyond natural language processing based-fake news detection that depends on text data only. However, further efforts should be made to expand the modality set when building fake news datasets by adding, for example, also audio and video.

- multi-lingual datasets: As emerged from the quantitative analysis, most of the developed datasets are only in the English language, limiting the efficacy of fact-checking in different languages. The opportunity to have multi-lingual fake news datasets helps to detect fake news in languages with lacking resources, broadening the datasets’ applicability to detection methods that are not based on specific languages. There is a lack of multi-lingual and cross-domain datasets collected from multiple sources. As we mentioned before, the only surveyed dataset that is multi-lingual (English and German) is PHEME. More recently, Shahi & Nandini (2020) proposed FakeCovid, a multi-lingual dataset containing 5,182 fact-checked news articles on COVID-19 written in 40 different languages. Another multi-lingual dataset has been developed by Abonizio et al. (2020) and contains 9,930 news items written in American English, Brazilian Portuguese, and Spanish. These datasets represent the first attempts to provide a more generic fake news detection approach that is not restricted to any specific language.

- cross-domain datasets: The development of cross-domain datasets is another open challenge that emerged from the quantitative analysis discussed in the “Quantitative analysis of surveyed datasets” section. The majority of the surveyed datasets (55%), indeed, collect fake news items concerning only one news domain, which limits the efficacy of news detection methods that suffer from data dependence and are easily affected by noise. Six surveyed datasets (30%) consider only two news domains, while only one dataset (i.e. Spanish fake news corpus) considers nine news domains. More recently, Wang et al. (2020) collected a Weibo dataset containing 7,300 news articles in Chinese across eight domains (health, economic, technology, entertainment, society, military, political, and education). Similarly, Amjad et al. (2020) proposed a benchmark dataset in the Urdu language that contains 900 news articles in 5 different domains (business, health, showbiz, sports, and technology). Therefore, to overcome the limitations on the diversity of fact-checking, future datasets should be collected from a variety of news domains.

- COVID-19 datasets: A further consideration regarding the news domain is that after the outbreak of the COVID-19 pandemic there was a proliferation of datasets focused on the health domain. As emerged from the quantitative analysis, indeed, the majority of the surveyed datasets, which were selected before the pandemic, focused mainly on politics and society, following the necessities that emerged during the period in which the news was collected. From March 2020 to date, however, there was a spread of misinformation related to COVID-19 that caused serious social disruption; consequently, several researchers started to collect datasets of fake medical news about the effect of the COVID-19 pandemic. For instance, Shahi & Nandini (2020) proposed FakeCovid, a dataset containing 5,182 fact-checked news articles for COVID-19. Cui & Lee (2020) collected CoAID, a Covid-19 healthcare misinformation dataset including 3,235 fake and true news articles and 851 social platform posts about COVID-19. In addition, Qazi, Imran & Ofli (2020) collected GeoCoV19, a large-scale geolocated Twitter dataset containing more than 524 million multi-lingual tweets posted over 90 days since February 1, 2020. Another Twitter dataset, named ReCOvery and containing a total of 2,029 news articles and 140,820 tweets on coronavirus published from January to May 2020, was collected by Zhou et al. (2020). The repository provides multimodal information on news articles on coronavirus, including textual, visual, temporal, and network information. Besides, COVID-19-FAKES is a publicly available (https://github.com/mohaddad/COVID-FAKES) and automatically annotated bilingual (Arabic/English) COVID-19 Twitter dataset (Elhadad, Li & Gebali, 2021) that was collected from February 04 to March 10, 2020. The dataset has been annotated using the shared information on the official websites and the official Twitter accounts of the WHO, UNICEF, and UN as a source of reliable information, and the collected COVID-19 pre-checked facts from different fact-checking websites to build a ground-truth database. A further example is COV19Tweets Dataset (Lamsal, 2020), a publicly available dataset (https://ieee-dataport.org/open-access/coronavirus-covid-19-tweets-dataset), that contains more than 310 million COVID-19-specific English language tweets and their sentiment scores. Tweets were collected from 20 March to 17 April 2020 and the number of tweets captured in that period per day, on average, was around 893 k. Therefore, in the last months, the need to detect misinformation related to COVID-19 has rapidly risen, making it urgent to collect specific datasets to evaluate the COVID-19 fake news detection methods.

- Big datasets: Building a large-scale fake news benchmark dataset is another open challenge that should be pursued to facilitate further research in this area. The quantitative analysis illustrated above showed that only one dataset (the CREDBANK dataset) collected millions of articles, while the majority of the surveyed datasets are relatively small collections with less than 10,000 news articles. A big dataset is fundamental for achieving a highly accurate fake detection process, mainly for fake news detection methods based on deep neural network models, which need a large dataset because their performance improves as the training dataset size increases. Torabi & Taboada (2019) discussed the necessity to use big data for fake news detection and encouraged researchers in this field to share their datasets and to work together towards a standardized large-scale fake news benchmark dataset.