This paper describes ongoing work on the creation of a multilingual rumour dataset on football transfer news, FTR-18. Transfer rumours are continuously published by sports media. They can both harm the image of player or a club or increase the player's market value. The proposed dataset includes transfer articles written in English, Spanish and Portuguese. It also comprises Twitter reactions related to the transfer rumours. FTR-18 is suited for rumour classi cation tasks and allows the research on the linguistic patterns used in sports journalism.
In the last years, the way news are spread has
changed. Social networks, blogs/micro-blogs and other
untrusted online news sources have gained popularity,
thus allowing any user to produce unveri ed content.
This modern kind of media enables real-time
proliferation of news stories, and, as consequence, increases the
di usion of rumours, hoaxes and misinformation to a
global audience [
The football transfer market o ers a fertile ground
for rumour dissemination because rumours and
misCopyright © CIKM 2018 for the individual papers by the papers'
authors. Copyright © CIKM 2018 for the volume as a collection
by its editors. This volume and its papers are published under
information spreads may be motivated by personal
pro t or public harm. Releasing false or misleading
news about alleged moves of players between clubs
emerges as a strategy to increase a player's market
value and transfer fees. Negligent transfer
announcements are assigned to high sums directly related to
some transfers [
According to the Union of European Football Associations' (UEFA) annual report1, the European transfer market reached a record revenue of e5.6 billion during the Summer of 2017, a 6% increase in spending during this period relatively to the highest value recorded in the last ten years. Economical side e ects of rumours on football clubs shares are also reported, like the rumour of Cristiano Ronaldo joining Juventus FC, which made the club shares jump almost 10% on a single day2 before any o cial announcement.
Motivated by the extensive attention given by sports media to transfer news and the high amounts involved, we propose the creation of Football Transfer Rumours 2018 (FTR-18), a transfer rumours dataset for researching the linguistic patterns in their text and propagation mechanisms. FTR-18 is designed as a multilingual collection of articles published by the relevant news organisations either in English, Spanish or Portuguese languages. Besides news articles, our proposed dataset also comprises Twitter posts associated to the transfer rumours. The present work describes the creation process of FTR-18 dataset and discusses 1https://goo.gl/FoAv2g 2https://goo.gl/3wKhj3
Evidence: Real Madrid o cial announcementb Extract: Given the information published in the last few hours regarding an alleged agreement between Real Madrid C.F. and PSG for the player Kylian Mbappe, Real Madrid would like to state that it is completely false.
Real Madrid has not made any o er to PSG or the player and condemns the spreading of this type of information that has not been proven by the parties concerned. a https://goo.gl/HhRafF b https://goo.gl/7oCqBy the intended usage of the collection. 2
Collections of rumours from di erent natures have
been assembled before. However, these datasets
mainly focused on political and social issues [
The rst large-scale dataset on rumour tracking and
classi cation was proposed by Qazvinian et al. [
Silverman developed a dataset for the analysis of
how online media handles rumours and unveri ed
information [
LIAR is another dataset that could be employed
in fact-checking and stance classi cation tasks [
Thorne et al. developed FEVER (Fact Extraction
and VERi cation), a large-scale manually annotated
dataset focusing on veri cation of claims against
textual sources [
The PHEME-RSD and FEVER datasets were
employed in stance detection and veracity prediction
shared tasks. The PHEME-RSD was employed in
the Semantic Evaluation (SemEval) 2017
competition, Task 8 RumourEval5 [
Despite the variety of existing datasets for rumour analysis, none of them addresses football rumours. Besides, rumours in multilingual environments like the European Football market are yet to be explored by the academic research community. For our purpose, however, it is extremely important to track how news about transfer rumours are covered by sports media in di erent languages, as UEFA transfers generally occur between distinct European countries. Hence, our proposal di ers from the previous work as it introduces a new dataset comprising football transfer rumours in three di erent languages: English, Spanish and Portuguese. 3
The data included in FTR-18 was harvested during the 2018 Summer Transfer Window (STW18). Transfer windows are periods during the year in which football clubs can make international transfers. Every league has the right to choose two annual transfer windows, with the opening and closing dates being de ned by the football associations of the league.
The collected material includes both transfer news and rumours and the corresponding reactions to these rumours on Twitter. As we develop a multilingual dataset containing news and comments in English, Spanish and Portuguese involving UEFA clubs, the harvesting is performed during a common period including the transfer windows of England, Spain and Portugal. Accordingly, data is being collected from June 24 until August 31, 2018.
FTR-18 is built in three stages. First, we browsed the media and social networks for an initial assessment of transfer rumours involving top UEFA clubs. Next, we performed both a semi-automated news articles harvesting and a Twitter crawling, selecting news and tweets related to the rumours, respectively. Finally, as transfer moves are con rmed (or not), we annotate rumour veracity.
5http://alt.qcri.org/semeval2017/task8/ 6http://fever.ai/task.html 3.1
The rst step in the creation of the FTR-18 is the selection of the clubs to follow during the Summer Transfer Window of 2018. We picked English, Spanish and Portuguese clubs that played the group stage in 2017-2018 UEFA Champions League. This decision was made based on the languages we are more familiar with. With this restriction, we decided to monitor the following clubs:
England: Chelsea, Liverpool, Manchester City, Manchester United, Tottenham Hotspur Spain: Atletico Madrid, Barcelona, Real Madrid, Sevilla
Portugal: Ben ca, Porto, Sporting CP 3.2
The next stage of the FTR-18 generation is the collection of transfer news. We selected transfer rumours involving the monitored clubs. The rumours included narratives of permanence or transfer moves concerning football players, that is, whether players are leaving or being hired by the monitored clubs. For consistency reasons, we restricted the rumour selection to players a liated to clubs from countries having English, Spanish or Portuguese as their majority language. The rumours were manually tracked and extracted from many sports news sources.
For each rumour, we annotated the target player, the target's club by the opening of the 2018 Summer Transfer Window (denoted as source), and the rumoured destination club (See Figure 1). We considered rumours either about the transference of a target to another club (source 6= destination) or about the permanence of the player in the current club (source = destination). 3.3
News harvesting involved searching and collecting news articles, published by di erent news organisations, related to the identi ed rumours in the lan
Transfer rumour and news articles atguages of the dataset. It was performed on a daily basis during the STW18. Each selected news article is associated with a single transfer move, and it might report a rumour or a veri ed information. We restricted our harvesting to news containing headlines with clear transfer reports about a single target, even if they report secondary moves in the body text (e.g. another player's alleged move, or interest manifested by other clubs). News with headlines pointing potential transfer of a player to multiple clubs were discarded.
For the news collection, we extracted content and meta-data from each article. Content information includes headline, subhead, and body text, while for the meta-data we registered the article source, language, url and the publication date (See Figure 1). The news harvesting task is semi-automated, as we could not build scrapers suited for all news sources due to their poor CSS structure. 3.4
Once a rumour is identi ed, we build a crawler using the Twitter's streaming API7 to collect posts related to the rumour. In compliance to the established language restrictions, we only harvest tweets written in the languages of the FTR-18 dataset. We record both the message content, tweet creation date and the data associated to the user who posted the message (id, screen name, description, location, friends and followers count, veri cation account status ). If the post is a retweet, we also register the retweet status, the data associated to the user who posted the original message and the retweet and favourite counts.
Besides collecting tweets related to rumours, we also track Twitter account meta-data and followers statistics either for the players, for the monitored clubs and for the clubs involved in the transfer rumours. 3.5
As the rumour unfolds, we annotate the rumour veracity by adding to the transfer rumour meta-data the evidences that support or refute the rumour and the publication date of these evidences. Rumour veracity is to be inferred until the end of 2018 Summer Transfer Window. Thus, for the case when source 6= destination, the rumour veracity is stated as True, if the transfer is con rmed, or is assigned as False, if the target remains in the source club or if the target is transferred to a di erent destination during STW18. Similarly, for rumours in which source = destination, the rumour veracity is inferred as False if the player is transferred, or True otherwise.
7https://developer.twitter.com/en/docs/tweets/ lterrealtime/overview.html
Despite the challenges of manually determining the
credibility of a rumour [
Currently, the FTR-18 dataset comprises 3,045 news articles and more than 2,064K tweets (See Table 2). The news articles subset considers transfer news covered by online sports media written in English (1,517), Spanish (747) or Portuguese (781) by 96 di erent news organisations. This collection includes 304 transfer moves associated with 175 target football players. The Twitter subset contains original messages and retweets written in English (1,130K), Spanish (677K) or Portuguese (257K). The Twitter posts are related to 112 claimed transfer moves involving 84 di erent football players. The FTR-18 dataset meta-data and corresponding news source scrapers are available at https://github.com/dcaled/FTR-18.
In a preliminary analysis of the collected news
articles, we noticed many cross-references between
different news sources. This pattern has already been
identi ed by Silverman [
Another interesting observation concerning transfer news is the cascading move chain associated to a rumour. For example, news sources condition the acquisition of a new player on the sale of another player by the same club (\Cristiano Ronaldo open to Juventus move as Real Madrid consider selling to fund move for 8https://www.theguardian.com/football/2018/jul/03/footballtransfer-rumours-cristiano-ronaldo-real-madrid-juventusneymar Kylian Mbappe or Neymar 9"). This pattern drives the reader to expect a continuation of the rumourous story and its unfoldings, thus keeping audience's attention for follow-up stories. 5
FTR-18 is suited for most of the steps involved in
a rumour classi cation process, as presented by
Zubiaga et al. [
In a second annotation phase, we will manually label our news articles subset, identifying which articles report rumours and which report veri ed information. This annotation will make FTR-18 suited for a rumour detection task. Stance detection is another possible application for the FTR-18 dataset. The collected content allows the classi cation of how the news headlines orient towards a given transfer rumour. Besides news' stances, we can evaluate the attitudes manifested by the audience with respect to a transfer move using reactions to Twitter posts. Both news articles labelling and stance annotation depend on human judgement.
The collected data also encourages analysis of
rumour tracking. This research topic is scarcely explored
and consists of identifying content associated with a
rumour that is currently monitored [
Available
lingual
teriori as related or unrelated to the rumour,
following Qazvinian et al's approach [
Finally, the FTR-18 dataset will also allow the analysis on how football transfer news are reported by the sports media. The news articles collection will serve as input for identifying the linguistics structures employed by journalists in transfer coverage. We expect to conduct an investigative work on the propagation patterns present in football transfer news, such as the recurrent conditional transfer moves and the echo effect caused by repetitions of unveri ed stories published by third-party news organisations. Further improvements in the FTR-18 dataset include the addition of new football leagues (e.g. CONMEBOL), the expansion of monitored clubs set and the collection of data from future transfer windows.
This work was supported by FCT, under grant No. UID/CEC/50021/ 2013.