=Paper=
{{Paper
|id=Vol-3180/paper-28
|storemode=property
|title=Overview of the CLEF-2022 CheckThat! Lab Task 1 on Identifying Relevant Claims in
Tweets
|pdfUrl=https://ceur-ws.org/Vol-3180/paper-28.pdf
|volume=Vol-3180
|authors=Preslav Nakov,Alberto Barrón-Cedeño,Giovanni Da San Martino,Firoj Alam,Rubén Míguez,Tommaso Caselli,Mucahid Kutlu,Wajdi Zaghouani,Chengkai Li,Shaden Shaar,Hamdy Mubarak,Alex Nikolov,Yavuz Selim Kartal
|dblpUrl=https://dblp.org/rec/conf/clef/NakovBMAMCKZLSM22
}}
==Overview of the CLEF-2022 CheckThat! Lab Task 1 on Identifying Relevant Claims in
Tweets==
https://ceur-ws.org/Vol-3180/paper-28.pdf
Overview of the CLEF-2022 CheckThat! Lab Task 1
on Identifying Relevant Claims in Tweets
Preslav Nakov1 , Alberto Barrón-Cedeño2 , Giovanni Da San Martino3 , Firoj Alam4 ,
Rubén Míguez5 , Tommaso Caselli6 , Mucahid Kutlu7 , Wajdi Zaghouani8 , Chengkai Li9 ,
Shaden Shaar10 , Hamdy Mubarak4 , Alex Nikolov11 and Yavuz Selim Kartal12
1
Mohamed bin Zayed University of Artificial Intelligence, UAE
2
DIT, Università di Bologna, Italy
3
University of Padova, Italy
4
Qatar Computing Research Institute, HBKU, Qatar
5
Newtral Media Audiovisual, Spain
6
University of Groningen, Netherland
7
TOBB University of Economics and Technology, Turkey
8
Hamad Bin Khalifa University, Qatar
9
University of Texas at Arlington, USA
10
Cornell University, USA
11
Sofia University, Bulgaria
12
GESIS – Leibniz Institute for the Social Sciences, Germany
Abstract
We present an overview of CheckThat! lab 2022 Task 1, part of the 2022 Conference and Labs of the
Evaluation Forum (CLEF). Task 1 asked to predict which posts in a Twitter stream are worth fact-checking,
focusing on COVID-19 and politics in six languages: Arabic, Bulgarian, Dutch, English, Spanish, and
Turkish. A total of 19 teams participated and most submissions managed to achieve sizable improvements
over the baselines using Transformer-based models such as BERT and GPT-3. Across the four subtasks,
approaches that targetted multiple languages (be it individually or in conjunction, in general obtained
the best performance. We describe the dataset and the task setup, including the evaluation settings, and
we give a brief overview of the participating systems. As usual in the CheckThat! lab, we release to
the research community all datasets from the lab as well as the evaluation scripts, which should enable
further research on finding relevant tweets that can help different stakeholders such as fact-checkers,
journalists, and policymakers.
Keywords
Check-Worthiness Estimation, Fact-Checking, Veracity, Social Media Verification, Computational Jour-
nalism, COVID-19.
CLEF 2022: Conference and Labs of the Evaluation Forum, September 5–8, 2022, Bologna, Italy
$ preslav.nakov@mbzuai.ac.ae (P. Nakov); a.barron@unibo.it (A. Barrón-Cedeño); dasan@math.unipd.it (G. Da
San Martino); fialam@hbku.edu.qa (F. Alam); ruben.miguez@newtral.es (R. Míguez); t.caselli@rug.nl (T. Caselli);
m.kutlu@etu.edu.tr (M. Kutlu); wzaghouani@hbku.edu.qa (W. Zaghouani); cli@uta.edu (C. Li); ss2753@cornell.edu
(S. Shaar); hmubarak@hbku.edu.qa (H. Mubarak); alexnickolow@gmail.com (A. Nikolov); ykartal@etu.edu.tr
(Y. S. Kartal)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
Figure 1: The CheckThat! lab verification pipeline. The 2022 edition of the lab covers three tasks:
(i) check-worthiness estimation, (ii) verified claim retrieval, and (iii) fake news detection. The grayed
tasks were addressed in previous editions of the lab [6, 7].
1. Introduction
The CheckThat! 2022 lab was held in the framework of CLEF 2022 [1]1 . Figure 1 shows the full
CheckThat! identification and verification pipeline, highlighting the three tasks targeted in
this fifth edition of the lab: Task 1 on detecting relevant claims in tweets (this paper), Task 2 on
retrieving relevant previously fact-checked tweets [2], and Task 3 on predicting the veracity of
news [3]. Task 1 asks to detect relevant tweets on the basis of different complementary criteria:
check-worthiness, verifiability, harmfulness, and attention-worthiness. We provided manually
annotated data for these four subtasks in five languages: Arabic, Bulgarian, Dutch, English,
and Turkish. For a sixth language, Spanish, we provided a larger-scale dataset annotated by
investigative journalists, but only for the subtask of check-worthiness identification.
The CheckThat! 2022 Task 1 framework enabled the experimentation on identifying relevant
tweets with different technologies, most of them built with transformer-based monolingual
and multilingual approaches, by 18 teams from around the world. Some of the most successful
approaches produced multilingual sequence-to-sequence models to take advantage of languages
with large amounts of training material to help processing tweets in less-resourced languages [4].
Another team explored training a feed-forward neural network with BERT embeddings and
Manifold Mixup regularization [5] to have smoother decision boundaries.
Among the different subtasks for finding relevant tweets, the check-worthiness one was
the most popular. English was the most popular target language for the participants. Across
the different submitted systems, transformer-based models were widely used. The top-ranked
systems also used data augmentation and additional prepossessing steps.
The remainder of the paper is organized as follows. Section 2 presents the different subtasks
offered this year. Section 3 describes the datasets and the evaluation measures. Section 5
discusses the system submissions and the evaluation results. Section 6 presents some related
work. Section 7 offers final remarks.
1
http://sites.google.com/view/clef2022-checkthat/
Table 1
The class labels for Subtasks 1A, 1B, 1C, and 1D.
Subtask 1A Subtask 1C Subtask 1D
1. No 1. No 1. No 6. Yes, contains advice
2. Yes 2. Yes 2. Yes, asks question 7. Yes, discusses action taken
Subtask 1B 3. Yes, blame authorities 8. Yes, discusses cure
1. No 4. Yes, calls for action 9. Yes, other
2. Yes 5. Yes, Harmful
2. Identifying Relevant Claims in Tweets
The aim of Task 1 is to determine whether a claim in a tweet is worth fact-checking. In order
to do that, we either resort to the judgments of professional fact-checkers (for Spanish) or
we ask non-expert annotators to answer several auxiliary questions [8, 9], such as “does the
tweet contain a verifiable factual claim?”, “is it harmful?”, and “is it of general interest?”, before
deciding on the final check-worthiness label. These questions opened the door to setting up
four subtasks:
Subtask 1A: Check-worthiness of tweets.
Given a tweet, predict whether it is worth fact-checking.
Subtask 1B: Verifiable factual claims detection.
Given a tweet, predict whether it contains a verifiable claim or not.
Subtask 1C: Harmful tweet detection.
Given a tweet, predict whether it is harmful to society.
Subtask 1D: Attention-worthy tweet detection.
Given a tweet, predict whether it should get the attention of policymakers and why.
Subtasks 1A, 1B, and 1C are all binary problems and the models are expected to establish
whether a tweet is relevant according to different criteria. Task 1D is a multi-class problem.
Table 1 shows the class labels for each task. Regarding languages, Arabic, Bulgarian, Dutch,
English, and Turkish are present in all four subtasks, whereas Spanish is included only in
Subtask 1A. We created and released an independently labeled dataset per language as explained
in the following section. The participants were free to work on any language(s) of their interest,
and they could also use multilingual approaches that make use of all datasets for training.
3. Datasets
For all subtasks (1A, 1B, 1C, and 1D) and for Arabic, Bulgarian, Dutch, and English languages
we used the datasets described in [9]. The datasets were developed based on a multi-question
annotation schema and the annotated tweets for the languages mentioned above [8]. Following
the same annotation schema, we further annotated a Turkish dataset.
Table 2
Statistics about the CT–CWT–22 corpus for all six languages and four subtasks. The bottom part of the
table shows the main topics covered.
Subtask Partition AR BG NL EN ES TR Total
Train 2,513 1,871 923 2,122 4,990 2,417 14,836
Dev 235 177 72 195 2,500 222 3,401
1A Dev-Test 691 519 252 574 2,500 660 5,196
Test 682 130 666 149 5,000 303 6,930
Total 4,121 2,697 1,913 3,040 14,990 3,602
Train 3,631 2,710 1,950 3,324 2,417 14,032
Dev 339 251 181 307 222 1,300
1B Dev-Test 996 736 534 911 660 3,837
Test 1,248 329 1,358 251 512 3,698
Total 6,214 4,026 4,023 4,793 3,811
Train 3,624 2,708 1,946 3,323 2,417 14,018
Dev 336 250 179 307 222 1,294
1C Dev-Test 994 735 531 910 660 3,830
Test 1,201 325 1,360 251 512 3,649
Total 6,155 4,018 4,016 4,791 3,811
Train 3,621 2,710 1,949 3,321 1,904 13,505
Dev 338 251 179 306 178 1,252
1D Dev-Test 995 736 533 909 533 3,706
Test 1,186 329 1,356 251 465 3,587
Total 6,140 4,026 4,017 4,787 3,080
Main topics
COVID-19 ■ ■ ■ ■ ■
Politics ■ ■ ■
For Spanish, we used a different approach. The Spanish tweets for subtask 1A were manually
annotated by journalists from Newtral —a Spanish fact-checking organization— and came from
the Twitter accounts of 300 Spanish politicians. Moreover, the Spanish dataset is the largest one
across the six languages, by a margin. Table 2 shows some statistics about the datasets, which
are split into training, development, and testing partitions.
Although all languages tackled major topics such as COVID-19 and politics, the crawling and
the annotation were done differently across the languages due to different resources available.
Below, we provide more detail about the crawling and the annotation for each language.
3.1. Arabic, Bulgarian, Dutch and English Datasets
We collected tweets by specifying the target language, a set of COVID-19 keywords (shown on
Figure 2), and time frames from January 2020 till March 2021. We removed retweets, replies,
duplicates (using a similarity-based approach) [10], as well as tweets with less than five words.
Finally, we selected the most frequently liked and retweeted tweets for annotation.
Figure 2: The keywords used to collect the Arabic, the Bulgarian, the Dutch, and the English tweets.
For the data, we considered a number of factors, including tweet popularity in terms of
retweets, which is already taken into account as part of the data collection process. We further
asked the annotators to answer four questions for each instance:2
• Verifiable factual claims: Does the tweet contain a verifiable factual claim? This
is an objective question, and it proved easy to annotate. Influenced by [11], positive
examples include tweets that state a definition, mention a quantity in the present or in
the past, make a verifiable prediction of the future, reference laws, procedures, and rules
of operation, discuss images or videos, and state correlation or causation, among others.
• Check-worthiness: Do you think that a professional fact-checker should verify
the claim in the tweet? This question asks for a subjective judgment. Yet, its answer
should be based on whether the claim is likely to be false, is of public interest, and/or
appears to be harmful. Note that we stress the fact that a professional fact-checker should
verify the claim, ruling out claims that are easy to fact-check by a layperson.
• Harmful tweet detection: Is the tweet harmful to the society and why? This is an
objective question. It further asks to categorize the nature of the harm if any. (even if we
do not ask this to the participating models).
• Attention-worthy tweet: Do you think that this tweet should get the attention
of a government entity? This question asks for a subjective judgment about whether
the target tweet should get the attention of a government entity or of policymakers in
general. The answers to this question are categorical and are not on an ordinal scale.
The annotations were performed by 2–5 people independently, and consolidated for the cases
of disagreement. The annotation setup was part of a broader initiative; see [9] for details.
2
We used the following MicroMappers setup for the annotations:
http://micromappers.qcri.org/project/covid19-tweet-labelling/.
3.2. Spanish Dataset
The Spanish dataset is an extended version from the one used in the 2021 edition of the lab [12].
A total of 5,000 new tweets have been added as test set, sampled from recent tweets from 350
well-known Spanish politicians. As in the previous year, professional journalists with expertise
in fact-checking determined the level of check-worthiness of the tweets on the basis of diverse
editorial criteria, such as actuality, public relevance of the character behind the claim, and
potential impact on the general audience. All tweets in Spanish were annotated independently
by three experts and the final decision was made by majority voting.
3.3. Turkish Dataset
We crawled Turkish tweets tracking keywords related to COVID-19 using Twitter API from 24
June, 2021 to September 29, 2021, yielding 10.87 M tweets in total. Keywords we used include
“covid”, “corona”, “kovid”, “korona”, “aşı”, “asi”, “pfizer”, “biontech”, “sinovac”, “astrazeneca’,
“moderna”, “turkovac”, “salgin”, “pandemi”, and “salgın”. Subsequently, we deduplicated our
crawl and eliminated tweets which (i) quote another tweet, (ii) are shorter than five words,
(iii) start with a URL, or (iv) are posted as a reply to other messages. Subsequently, we sorted
tweets based on their popularity in terms of number of retweets and likes. We picked the most
popular 4K tweets whose cosine similarity score was less than 0.75 for the annotation.
Each tweet was annotated by three people independently using the same platform utilized for
creating datasets for Arabic, Bulgarian, Dutch, and English. Disagreements on the annotations
were solved by discussions among the annotators. However, if disagreements for a particular
tweet continued even after the discussion, that tweet is discarded.
4. Evaluation Settings
For the lab, we provided the training, development dev-test set to enable the participants to
validate their systems internally, while they could use the dev set for parameter tuning. For
each language and subtask, we have annotated new instances, using three or four annotators
per instance. Class label has been assigned by majority voting and disagreements have been
solved by a consolidator or discussions among annotators. The test set has been used for final
evaluation and system ranking. Participants were allowed to submit as many runs as they
wanted on the test set, but only the last one was considered as official.
The evaluation framework is completed by the evaluation metrics for each subtask. For
subtasks 1A and 1C, we used the F1 -measure with respect to the positive class (yes), to
account for class imbalance. For subtask 1B, we used accuracy, as the data is fairly balanced.
For subtask 1D, we used weighted-F1 , as there are multiple classes and we wanted them
appropriately weighted. The data and the evaluation scripts are available online.3
3
https://gitlab.com/checkthat_lab/clef2022-checkthat-lab/
5. Overview of the Systems and Evaluation Results
Eighteen teams took part in this task, with English being the most popular language. Across
the different subtasks, some teams paid special attention to multiple languages, mostly through
three different strategies. MT-based data augmentation. Team TOBB-ETU [13] (the only team
that targeted all four subtasks in almost all languages available) applied both translation and
back-translation to increase the amount of training data in the different languages. Language-
specific models were then trained. Multilingual transformer. Team NUS-IDS [4] adopted
a multilingual mT5 transformer to train one single model and apply it to multiple languages.
Zero-shot. Team PoliMi-FlatEarthers [14] fine-tuned a GPT-3 model for each of the four
subtasks feeding only instances in English and applied them to other languages during testing.
In the rest of the section we zoom into each of the subtasks, by looking into the models and
resources explored and the performance of the official runs. Tables 3, 5, 7 and 9 offer a bird’s
eye overview of the participants’ approaches to each of the four subtasks. Tables 4, 6, 8 and 10
show the official evaluation scores for each of the four subtasks. Appendix A includes a brief
description of the approaches for every participating team.
5.1. Subtask 1A. Check-Worthiness Estimation
A total of 18 teams took part in this task, with English, Bulgarian, and Dutch being the most
popular languages. Three teams —NUS-IDS [4]), PoliMi-FlatEarthers [14] and TOBB ETU [13]—
participated in five out of the six languages offered. Table 3 shows an overview of the approaches,
whereas Table 4 shows the performance of the official submissions, ranked on the basis of F1
with respect to the positive class. The baseline consists of a random system. We now zoom into
the different languages.
Arabic Four teams participated. As observed for other languages, the multilingual approach
of team NUS-IDS [4] allowed them to effectively take advantage of the supervised data in
other languages, resulting in the top-performing approach. It is based on mT5, a multilingual
sequence-to-sequence transformer pretrained on the mC4 corpus, which covers 101 languages.
The second best system, TOBB ETU [13], used fine-tuned AraBERT.
Bulgarian Five teams took part. Once again NUS-IDS [4] was the top-ranked team, followed
by Team TOBB ETU [13], using the same approaches as for Arabic.
Dutch Five teams participated. The multilingual approach by team NUS-IDS [4] outperformed
the other systems again. This time, team AI Rational [15] arrived second with a system built
with RoBERTa, after a data augmentation process based on back-translation.
English A total of 13 teams took part. The top-ranked team was AI Rational [15], with a
similar approach to the one they applied for Dutch. Team Zorros [23] submitted the second-best
system, based on an ensemble approach combining BERT and RoBERTa. Two other aspects are
worth highlighting. On the one hand, PoliMi-FlatEarthers [14] fine-tuned a unique GPT-3
model with all instances in English. When applying it on the English test set, they obtained
Table 3
Overview of the approaches to subtask 1A. The numbers in the language box refer to the position of
the team in the official ranking; marks under multilingual flag models that have been trained on more
than one language at once; ✓“part of the official submission; Ë“considered in internal experiments.
Team Languages Transformers Repr. Classifiers Miscellaneous
Data normalization
Data augmentation
Multi-task learn.
Semi-supervised
Random forest
XLM RoBERTa
Quantum NLP
word 𝑛-grams
Multilingual
DistilBERT
Bulgarian
Ensemble
RoBERTa
XGBoost
mT5-XL
Spanish
Turkish
English
Electra
Arabic
XLNet
Dutch
GPT-3
ELMo
LIWC
BERT
CNN
RNN
SVM
AI Rational [15] 3 2 1 2 ËË ✓ ✓ ✓ ✓
ARC-NLP [16] 3 Ë ✓ ✓ ✓ ✓ ✓
Asatya [17] 10 ✓ ✓ ✓ Ë
Fraunhofer SIT [18] 5 ✓ ✓ ✓ ✓ ✓ Ë ✓
iCompass [19] 3 ✓ ✓ ✓ Ë
NUS-IDS [4] 1 1 1 8 1 ✓ ✓ Ë Ë ✓
PoliMi-FlatE. [14] 5 5 4 3 2 ✓ ✓
RUB-DFL [20] 6 1 ✓ Ë Ë ✓ ✓ ✓ ✓
TOBB ETU [13] 2 2 3 4 4 ✓ Ë ✓ ✓
VTU BGM [21] 11 ✓ ✓ ✓
Z-Index [22] 5 12 3 ✓ ✓ Ë ✓ Ë Ë ✓ ✓
Zorros [23] 2 ✓ ✓ ✓ ✓
the third best performance. The zero-shot application of the model to other languages was
not as successful. On the other hand, teams that trained actual multilingual models, that is
NUS-IDS [4] and Z-index [22], struggled when dealing with English.
Spanish Three teams took part. The multilingual approach by team NUS-IDS [4] was
the most successful by a margin over the top runner —the zero-shot approach by PoliMi-
FlatEarthers [14]. A first sight might suggest that such zero-shot approach hints to be working
for Spanish (their performance goes below the baseline in the other languages), but the distance
to the top model is still bigger than two points. It is worth observing that the performances
over the Spanish datasets are the lowest. Whether the reason behind is that this is the only one
annotated by expert journalists rather than by crowdsourcing remains an open topic.
Turkish Four teams participated. All participants used BERT-based models and GPT-3. Team
RUB-DFL [20] is the top-performing one. The system uses BERT with a combination of both
ELMo and LIWC features. The runner up team AI Rational applied standard pre-processing
and data augmentation with back translation.
5.2. Subtask 1B: Verifiable Factual Claims Detection
Thirteen teams took part in Subtask 1B, with English, Bulgarian and Arabic being the most
popular languages. Team TOBB ETU [13] participated in all five languages, whereas team
Table 4
Subtask 1A: Check-Worthiness estimation, results for the official submissions in all six languages. F1
with respect to the positive class. Baseline is the random baseline.
Team F1 Team F1 Team F1
Arabic English Spanish
1. NUS-IDS [4] 0.628 1. AI Rational [15] 0.698 1. NUS-IDS [4] 0.571
2. TOBB ETU [13] 0.495 2. Zorros [23] 0.667 2. PoliMi-FlatEarthers [14] 0.323
3. iCompass [19] 0.462 3. PoliMi-FlatEarthers [14] 0.626 3. Z-Index [22] 0.303
4. Baseline 0.347 4. TOBB ETU [13] 0.561 4. Baseline 0.139
5. PoliMi-FlatEarthers [14] 0.321 5. Fraunhofer SIT [18] 0.552 Turkish
Bulgarian 6. RUB-DFL [20] 0.525 1. RUB-DFL [20] 0.801
1. NUS-IDS [4] 0.617 7. hinokicrum˚ 0.522 2. AI Rational [15] 0.789
2. TOBB ETU [13] 0.542 8. NUS-IDS [4] 0.519 3. ARC-NLP [16] 0.760
3. AI Rational [15] 0.483 9. TonyTTTTT˚ 0.500 4. TOBB ETU [13] 0.729
4. Baseline 0.434 10. Asatya [17] 0.500 5. Baseline 0.496
5. PoliMi-FlatEarthers [14] 0.341 11. VTU_BGM [21] 0.482
6. pogs2022˚ 0.000 12. Z-Index [22] 0.478
Dutch 13. NLP&IR@UNED˚ 0.469
1. NUS-IDS [4] 0.642 14. Baseline 0.253
2. AI Rational [15] 0.620
3. TOBB ETU [13] 0.534
4. PoliMi-FlatEarthers [14] 0.532
5. Z-Index [22] 0.497
6. Baseline 0.451
˚
No working note submitted.
AI Rational participated in four. Table 5, shows an overview of the explored approaches and
Table 6 shows the performance of the official submissions on the test set, including the random
baseline. The table shows the runs ranked on the basis of the official accuracy measure.
Arabic Three teams participated. Team TOBB ETU [13] ranked best, being the only one
that surpassed the baseline. They used a four-layer feedforward network with Manifold Mixup
regularization and BERT embeddings. Arabic showed to be the most challenging language for
Task 1B, with a top performance shorter than 0.60 and ony one team beating the baseline.
Bulgarian Two teams stook part. Team AI Rational [15] topped using XLM-RoBERTa with
data augmentation, followed by Team TOBB ETU [13], which fine-tuned RoBERTa.
Dutch Two teams participated. Team AI Rational [15] and TOBB ETU [13] used similar
approaches as for Bulgarian.
English Nine teams participated. Team PoliMi-FlatEarthers [14] ranked as the best system
thanks to their GPT-3 fine-tuning. Team Asatya [17] arrived second by fine-tuning BERT.
Turkish Four teams participated. The top-ranked team is RUB-DFL [20], which used again a
combination of RoBERTa, Electra, and BERTurk. The second-best team is AI Rational [15],
Table 5
Overview of the approaches to subtask 1B. The numbers in the language box refer to the position of the
team in the official ranking; ✓“part of the official submission; Ë“considered in internal experiments.
Team Languages Transformers Repr. Misc
Data augmentation
XLM RoBERTa
word 𝑛-grams
Preprocessing
DistilBERT
Bulgarian
RoBERTa
Turkish
English
Electra
Arabic
XLNet
Dutch
GPT-3
ELMo
LIWC
BERT
AI Rational [15] 1 1 4 2 ËË✓✓ ✓ ✓
ARC-NLP [16] 3 ✓ Ë
Asatya [17] 2 3 ✓ ✓ ✓
PoliMi-FlatEarthers [14] 1 ✓
RUB-DFL [20] 6 1 ✓ ✓ ✓✓ ËË
TOBB ETU [13] 1 2 2 9 4 Ë ✓ ✓ ✓
VTU BGM [21] 7 ✓ ✓ ✓
Zorros [23] 5 ✓ ✓ ✓
Table 6
Subtask 1B: Verifiable Factual Claims Detection, results for the official submissions in all five languages.
Team Acc Team Acc Team Acc
Arabic English Turkish
1. TOBB ETU [13] 0.570 1. PoliMi-FlatEarthers [14] 0.761 1. RUB-DFL [20] 0.801
2. Baseline 0.531 2. Asatya [17] 0.749 3. AI Rational [15] 0.789
3. claeser˚ 0.454 3. NLP&IR@UNED˚ 0.725 3. ARC-NLP [16] 0.760
4. pogs2022˚ 0.454 4. AI Rational [15] 0.713 4. TOBB ETU [13] 0.729
Bulgarian 5. Zorros [23] 0.709 5. Baseline 0.496
1. AI Rational [15] 0.839 6. RUB-DFL [20] 0.709
2. TOBB ETU [13] 0.742 7. VTU_BGM [21] 0.709
3. Baseline 0.535 8. hinokicrum˚ 0.665
Dutch 9. TOBB ETU [13] 0.641
1. AI Rational [15] 0.736 10. Baseline 0.494
2. TOBB ETU [13] 0.658
3. Baseline 0.521
˚
No working note submitted.
which used BERT, RoBERTa, and DistilBERT.
5.3. Subtask 1C: Harmful Tweets Detection
Thirteen teams participated in subtask 1C, with English and Turkish being the most popular
languages. Teams TOBB ETU [13] and AI Rational [15] participated in five and four languages,
Table 7
Overview of the approaches to subtask 1C. The numbers in the language box refer to the position of the
team in the official ranking; ✓“part of the official submission; Ë“considered in internal experiments.
Team Languages Transformers Repr. Misc
Data augmentation
XLM RoBERTa
word 𝑛-grams
Preprocessing
DistilBERT
sentiment
Bulgarian
RoBERTa
Turkish
English
Electra
Arabic
XLNet
Dutch
GPT-3
ELMo
LIWC
BERT
AI Rational [15] 1 2 2 3 ËË✓✓ ✓ ✓
ARC-NLP [16] 6 1 ✓ ✓
Asatya [17] 3 ✓ ✓ ✓
COURAGE [24] 8 ✓ ✓
iCompass [19] 1 ✓ ✓ ✓
PoliMi-FlatEarthers [14] 10 ✓
RUB-DFL [20] 9 2 ✓ ✓✓ Ë Ë Ë
TOBB ETU [13] 2 2 1 5 4 Ë ✓ ✓ ✓
VTU BGM [21] 7 ✓ ✓ ✓
Zorros [23] 1 ✓ ✓ ✓
respectively. Table 7 overviews the teams’ approaches and Table 8 shows the performance of
the official submissions on the test set, together with the random baseline. The table shows the
runs ranked based on the official F1 with respect to the positive class.
Arabic Two teams participated. Team iCompass [19] obtained the top performance by
fine-tuning AraBERT and ARBERT. Team TOBB ETU [13] opted for a feedforward network
trained with Manifold Mixup regularization and represented tweets with AraBERT embeddings.
Notice that the top model doubles the performance of the second one.
Bulgarian Two teams participated. Team AI Rational [15] topped using XLM-RoBERTa.
TOBB ETU [13] arrived second with a fine-tuned RoBERTa. Both applied data augmentation
via back-translation. Bulgarian showed to be the most challenging language in task 1C.
Dutch Two teams participated. Team TOBB ETU [13] ranked on top by fine-tuning
BERTje [25] after data-augmentation. Team AI Rational [15] used XLM-RoBERTa.
English Eleven teams participated. Team Zorros [23] ranked as the best system, using an
ensemble of five transformers. Team ARC-NLP [16] ranked second. Besides transformer-based
models across all approaches, some teams have also used data augmentation.
Turkish Four teams participated. Team ARC-NLP [16] ranked on top by approaching harm
as a contradiction detection problem. They extracted facts related to COVID-19 from reliable
Table 8
Subtask 1C: Harmful Tweet Detection, results for the official submissions in all five languages.
Team F1 Team F1 Team F1
Arabic English Turkish
1. iCompass [19] 0.557 1. Zorros [23] 0.397 1. ARC-NLP [16] 0.366
2. TOBB ETU [13] 0.268 2. AI Rational [15] 0.361 2. RUB-DFL [20] 0.353
3. Baseline 0.118 3. Asatya [17] 0.361 3. AI Rational [15] 0.346
Bulgarian 4. NLP&IR@UNED˚ 0.347 4. TOBB ETU [13] 0.262
1. AI Rational [15] 0.286 5. TOBB ETU [13] 0.329 5. Baseline 0.061
2. TOBB ETU [13] 0.054 6. ARC-NLP [16] 0.300
3. Baseline 0.000 7. hinokicrum˚ 0.281
Dutch 8. COURAGE [24] 0.280
1. TOBB ETU [13] 0.3589. RUB-DFL [20] 0.273
2. AI Rational [15] 0.14710. PoliMi-FlatEarthers [14] 0.270
3. Baseline 0.11411. Baseline 0.200
12 VTU_BGM [21] 0.000
˚
No working note submitted.
Table 9
Overview of the approaches to subtask 1D. The numbers in the language box refer to the position of the
team in the official ranking; ✓“part of the official submission; Ë“considered in internal experiments.
Team Languages Transformers Misc
Data augmentation
XLM RoBERTa
Preprocessing
DistilBERT
Bulgarian
RoBERTa
Turkish
English
Arabic
Dutch
GPT-3
BERT
AI Rational [15] 1 1 3 1 ËË✓✓ ✓ ✓
PoliMi-FlatEarthers [14] 7 ✓
TOBB ETU [13] 2 2 2 4 3 ✓ ✓ ✓ Ë
Zorros [23] 1 ✓ ✓ ✓
sources and associated them with tweets based on similarity. The pairs were used to fine-tune
BERTurk. Team RUB-DFL [20] arrived second with a fine-tuned ConvBert.
5.4. Subtask 1D: Attention-Worthy Tweet Detection
Seven teams participated in subtask 1D, with English being the most popular language. Again,
teams TOBB ETU [13] and AI Rational [15] participated in five and four languages. Table 9
overviews the approaches whereas Table 10 shows the performance of the official submissions
on the test, together with the random baseline. The ranking is based on the official weighted F1.
Table 10
Subtask 1D: Attention-Worthy Tweet Detection, results for the official submissions in all five languages.
Performance is reported as weighted F1.
Team F1 Team F1 Team F1
Arabic English Turkish
1. Baseline 0.206 1. Zorros [23] 0.725 1. AI Rational [15] 0.895
2. TOBB ETU [13] 0.184 2. Baseline 0.695 2. Baseline 0.853
Bulgarian 3. AI Rational [15] 0.684 3. TOBB ETU [13] 0.806
1. AI Rational [15] 0.915 4. TOBB ETU [13] 0.670 Dutch
˚
2. TOBB ETU [13] 0.8775. NLP&IR@UNED 0.650 1. AI Rational [15] 0.715
3. Baseline 0.8756. hinokicrum˚ 0.643 2. TOBB ETU [13] 0.694
7. PoliMi-FlatEarthers [14] 0.636 3. Baseline 0.641
˚
No working note submitted.
Arabic Only one team participated. Team TOBB ETU [13] ran short and did not manage to
pass the random baseline. They tried with a dual approach: the combination of a binary model
attention-worthy vs not and a multi-class model with the original classes.
Bulgarian Two teams participated. Both Teams AI Rational [15] and TOBB ETU [13] used
similar models as the ones for subtask 1C.
Dutch Two teams participated, resulting in a similar picture as for Bulgarian.
English Six teams participated. Team Zorros [23] ranked first, by fine-tuning a COVID
Twitter BERT pre-trained model. The random baseline ranked second.
Turkish Two teams participated. Team AI Rational [15] was the only team to beat the
baseline with a fine-tuned XLM-RoBERTa model.
6. Related Work
There has been a significant research interest in recent years in identifying disinformation,
misinformation, and “fake news”, which thrive in social media, political debates and speeches.
Several recent works highlighted how information is disseminated and consumed in social
media [26], fact-checking perspective on “fake news” and related problems [27], truth discovery
[28], stance towards misinformation and disinformation detection [29], automatic fact-checking
to assist human fact-checkers [30], predicting the factuality and the bias of entire news outlets
[31], multimodal disinformation detection [32], and on abusive language in social media [33].
Within the scope in identifying disinformation, misinformation and the “fake news” in general,
the research interests have focused on more specific problems such as automatic identification
and verification of claims [34, 35, 36, 37, 7, 38, 39], to identifying check-worthy claims [40, 41,
42, 43, 44], detecting whether a claim has been previously fact-checked [45, 46, 47], retrieving
evidence to accept or to reject a claim [48, 49], checking whether the evidence supports or denies
the claim [50, 51], and inferring the veracity of the claim [52, 53, 54, 55, 56, 57, 49, 58, 59, 60].
Such specific tasks can help fact-checkers and/or journalists.
Among these tasks check-worthiness estimation received an wider attention since the pio-
neering work proposed by [41], where the idea is to detect whether a sentence in a political
debate is non-factual, unimportant factual, or check-worthy factual. The proposed system later
extended with more data and to cover Arabic content [42]. Most of the earlier work on check-
worthiness estimation was mainly focused political debates [61, 43] and lately attention has
been focused on social media [9, 8, 62].
Major research attention emerged due to the CheckThat! lab initiatives in CLEF 2018, 2019,
2020, and 2021 where the focus was once again on political debates and speeches, from a single
fact-checking organization. In the 2018 edition of the task, a total of seven teams submitted runs
for Task 1 (which corresponds to Subtask 1B in 2021), with systems based on word embeddings
and RNNs [63, 64, 65, 66]. In the 2019 edition of the task, eleven teams submitted runs for the
corresponding Task 1, again using word embeddings and RNNs, and further trying a number of
interesting representations [67, 68, 69, 70, 71, 72, 73, 74]. In the 2020 edition of the task, three
teams submitted runs for the corresponding Task 5 with systems based on word embeddings and
BiLSTM [75], TF.IDF representation with Naïve Bayes, logistic regression, decision trees [76],
BERT prediction scores, and word embeddings with logistic regression [77]. Several teams fine-
tuned pre-trained models such as AraBERT and multilingual BERT [78, 79, 77]. Other approaches
relied on pre-trained models such as GloVe and Word2vec [80, 75] to obtain embeddings for the
tweets, which were fed into a neural network or an SVM. In addition to text representations,
some teams used other features, namely morphological and syntactic, part-of-speech (POS) tags,
named entities, and sentiment features [81, 82]. As for the English task, we also observed the
popularity of pre-trained Transformers, namely BERT and RoBERTa [78, 80, 83, 84, 77]. In the
2021 edition [12], check-worthiness estimation has been offered for political debates/speeches
and tweets. Top ranked system used transformer based models [85, 86].
A large body of work has been devoted to identifying the factuality of claims, which are often
expressed and disseminated through social networks [87]. The studies include fact-checking on
news media [88], fact-checking such as fact-checked URL recommendation model [89] , fact-
checking with stance detection [57], factuality of media outlets [90], generating justifications
for verdicts on claims [91], and fact-checking claims from Wikipedia [92].
For harmfulness detection, research mainly focused on offensive and hateful content on social
media that can harm an individual, organization, and society [93, 94, 95].
Attention-worthiness is a relatively new area, which has recently been proposed in [9, 8]
and has not been explored yet in the current literature, and the CheckThat! lab 2022 initiative
opened up a new research interest on this topic.
7. Conclusions
We have presented an overview of task 1 of the CLEF-2022 CheckThat! lab. The lab featured
tasks that span the full verification pipeline: from spotting check-worthy claims to checking
whether they have been fact-checked before. Task 1 asked to identify relevant claims in tweets
in terms of check-worthiness, verifiability, harmfulness, and attention-worthiness.
Inline with the general mission of CLEF, we promoted multilinguality by offering the task
in six different languages. This edition of the task has attracted diverse approaches in terms
of model (e.g., kind of transformer), representations (e.g., embeddings, 𝑛-grams) and data
augmentation (e.g., back-translation). Among the most innovative ones, we highlight the use of
quantum NLP [4], GPT-3 [14], and mT5 [4]. Teams targeting multiple languages tended to rank
at the top positions across tasks and languages. The exception is English. Teams excelling in
multiple languages tended to rank relatively low in this language.
The general problem of check-worthiness estimation remains open in general and further
efforts could be paid on considering external evidence when assessing whether a tweet is calling
for the attention of a verifier. In this edition of the lab, we have observed already efforts in this
direction and the results are promising.
Acknowledgments
Part of this work is made within the Tanbih mega-project,4 developed at the Qatar Computing
Research Institute, HBKU, which aims to limit the impact of “fake news”, propaganda, and
media bias by making users aware of what they are reading, thus promoting media literacy and
critical thinking.
References
[1] P. Nakov, A. Barrón-Cedeño, G. Da San Martino, F. Alam, J. M. Struß, T. Mandl, R. Míguez,
T. Caselli, M. Kutlu, W. Zaghouani, C. Li, S. Shaar, G. K. Shahi, H. Mubarak, A. Nikolov,
N. Babulkov, Y. S. Kartal, J. Beltrán, M. Wiegand, M. Siegel, J. Köhler, Overview of
the CLEF-2022 CheckThat! lab on fighting the COVID-19 infodemic and fake news
detection, in: Proceedings of the 13th International Conference of the CLEF Association:
Information Access Evaluation meets Multilinguality, Multimodality, and Visualization,
CLEF ’2022, Bologna, Italy, 2022.
[2] P. Nakov, G. Da San Martino, F. Alam, S. Shaar, H. Mubarak, N. Babulkov, Overview of
the CLEF-2022 CheckThat! lab task 2 on detecting previously fact-checked claims, in:
Working Notes of CLEF 2022—Conference and Labs of the Evaluation Forum, CLEF ’2022,
Bologna, Italy, 2022.
[3] J. Köhler, G. K. Shahi, J. M. Struß, M. Wiegand, M. Siegel, T. Mandl, M. Schütz, Overview
of the CLEF-2022 CheckThat! lab task 3 on fake news detection, in: Working Notes of
CLEF 2022—Conference and Labs of the Evaluation Forum, CLEF ’2022, Bologna, Italy,
2022.
[4] S. K. N. Mingzhe Du, Sujatha Das Gollapalli, NUS-IDS at CheckThat! 2022: identify-
ing check-worthiness of tweets using CheckthaT5, in: Working Notes of CLEF 2022 -
Conference and Labs of the Evaluation Forum, CLEF ’2022, Bologna, Italy, 2022.
4
http://tanbih.qcri.org
[5] V. Verma, A. Lamb, C. Beckham, A. Najafi, I. Mitliagkas, D. Lopez-Paz, Y. Bengio, Manifold
mixup: Better representations by interpolating hidden states, in: International Conference
on Machine Learning, PMLR, 2019, pp. 6438–6447.
[6] A. Barrón-Cedeño, T. Elsayed, P. Nakov, G. Da San Martino, M. Hasanain, R. Suwaileh,
F. Haouari, N. Babulkov, B. Hamdan, A. Nikolov, S. Shaar, Z. Sheikh Ali, Overview of
CheckThat! 2020: Automatic identification and verification of claims in social media,
LNCS (12260), Springer, 2020, pp. 215–236.
[7] T. Elsayed, P. Nakov, A. Barrón-Cedeño, M. Hasanain, R. Suwaileh, G. Da San Martino,
P. Atanasova, Overview of the CLEF-2019 CheckThat!: Automatic identification and
verification of claims, in: Experimental IR Meets Multilinguality, Multimodality, and
Interaction, LNCS, 2019, pp. 301–321.
[8] F. Alam, F. Dalvi, S. Shaar, N. Durrani, H. Mubarak, A. Nikolov, G. Da San Martino,
A. Abdelali, H. Sajjad, K. Darwish, P. Nakov, Fighting the COVID-19 infodemic in social
media: A holistic perspective and a call to arms, in: Proceedings of the International
AAAI Conference on Web and Social Media, ICWSM ’21, 2021, pp. 913–922.
[9] F. Alam, S. Shaar, F. Dalvi, H. Sajjad, A. Nikolov, H. Mubarak, G. D. S. Martino, A. Abdelali,
N. Durrani, K. Darwish, A. Al-Homaid, W. Zaghouani, T. Caselli, G. Danoe, F. Stolk,
B. Bruntink, P. Nakov, Fighting the COVID-19 infodemic: Modeling the perspective of
journalists, fact-checkers, social media platforms, policy makers, and the society, in:
Findings of EMNLP 2021, 2021, pp. 611–649.
[10] F. Alam, H. Sajjad, M. Imran, F. Ofli, CrisisBench: Benchmarking crisis-related social media
datasets for humanitarian information processing, in: Proceedings of the International
AAAI Conference on Web and Social Media, ICWSM ’21, 2021, pp. 923–932. URL: https:
//ojs.aaai.org/index.php/ICWSM/article/view/18115.
[11] L. Konstantinovskiy, O. Price, M. Babakar, A. Zubiaga, Towards automated factcheck-
ing: Developing an annotation schema and benchmark for consistent automated claim
detection, arXiv:1809.08193 (2018).
[12] S. Shaar, M. Hasanain, B. Hamdan, Z. S. Ali, F. Haouari, A. Nikolov, M. Kutlu, Y. S. Kartal,
F. Alam, G. Da San Martino, A. Barrón-Cedeño, R. Míguez, J. Beltrán, T. Elsayed, P. Nakov,
Overview of the CLEF-2021 CheckThat! lab task 1 on check-worthiness estimation in
tweets and political debates, in: [96], 2021.
[13] A. B. Eyuboglu, M. B. Arslan, E. Sonmezer, M. Kutlu, TOBB ETU at CheckThat! 2022:
detecting attention-worthy and harmful tweets and check-worthy claims, in: Working
Notes of CLEF 2022 - Conference and Labs of the Evaluation Forum, CLEF ’2022, Bologna,
Italy, 2022.
[14] S. Agrestia, A. S. Hashemianb, M. J. Carmanc, PoliMi-FlatEarthers at CheckThat! 2022:
GPT-3 applied to claim detection, in: Working Notes of CLEF 2022 - Conference and Labs
of the Evaluation Forum, CLEF ’2022, Bologna, Italy, 2022.
[15] A. Savchev, AI Rational at CheckThat! 2022: using transformer models for tweet classifi-
cation, in: Working Notes of CLEF 2022 - Conference and Labs of the Evaluation Forum,
CLEF ’2022, Bologna, Italy, 2022.
[16] C. Toraman, O. Ozcelik, F. Şahinuç, U. Sahin, ARC-NLP at CheckThat! 2022: contradiction
for harmful tweet detection, in: Working Notes of CLEF 2022 - Conference and Labs of
the Evaluation Forum, CLEF ’2022, Bologna, Italy, 2022.
[17] P. K. Manan Suri, S. Dudeja, Asatya at CheckThat! 2022: multimodal BERT for identifying
claims in tweets, in: Working Notes of CLEF 2022 - Conference and Labs of the Evaluation
Forum, CLEF ’2022, Bologna, Italy, 2022.
[18] R. A. Frick, I. Vogel, I. Nunes Grieser, Fraunhofer SIT at CheckThat! 2022: semi-supervised
ensemble classification for detecting check-worthy tweets, in: Working Notes of CLEF
2022 - Conference and Labs of the Evaluation Forum, CLEF ’2022, Bologna, Italy, 2022.
[19] T. Bilel, B. N. Mohamed Aziz, H. Haddad, iCompass at CheckThat! 2022: ARBERT and
AraBERT for Arabic checkworthy tweet identification, in: Working Notes of CLEF 2022 -
Conference and Labs of the Evaluation Forum, CLEF ’2022, Bologna, Italy, 2022.
[20] Z. M. Hüsünbeyi, O. Deck, T. Scheffler, RUB-DFL at CheckThat! 2022: transformer models
and linguistic features for identifying relevant claims, in: Working Notes of CLEF 2022 -
Conference and Labs of the Evaluation Forum, CLEF ’2022, Bologna, Italy, 2022.
[21] S. Kavatagi, R. Rachh, M. Mulimani, VTU_BGM at Check That! 2022: an autoregressive
encoding model for verifying check-worthy claims, in: Working Notes of CLEF 2022 -
Conference and Labs of the Evaluation Forum, CLEF ’2022, Bologna, Italy, 2022.
[22] P. Tarannum, H. Md. Arid, F. Alam, S. R. H. Noori, Z-Index at CheckThat! Lab 2022: check-
worthiness identification on tweet text, in: Working Notes of CLEF 2022 - Conference
and Labs of the Evaluation Forum, CLEF ’2022, Bologna, Italy, 2022.
[23] R. M. Buliga Nicu, Zorros at CheckThat! 2022: ensemble model for identifying relevant
claims in tweets, in: Working Notes of CLEF 2022 - Conference and Labs of the Evaluation
Forum, CLEF ’2022, Bologna, Italy, 2022.
[24] F. Lomonaco, G. Donabauer, M. Siino, COURAGE at CheckThat! 2022: harmful tweet
detection using graph neural networks and ELECTRA, in: Working Notes of CLEF 2022 -
Conference and Labs of the Evaluation Forum, CLEF ’2022, Bologna, Italy, 2022.
[25] W. de Vries, A. van Cranenburgh, A. Bisazza, T. Caselli, G. van Noord, M. Nissim, Bertje:
A dutch bert model, arXiv preprint arXiv:1912.09582 (2019).
[26] K. Shu, A. Sliva, S. Wang, J. Tang, H. Liu, Fake news detection on social media: A data
mining perspective, SIGKDD Explor. Newsl. 19 (2017) 22–36. URL: http://doi.acm.org/10.
1145/3137597.3137600. doi:10.1145/3137597.3137600.
[27] J. Thorne, A. Vlachos, Automated fact checking: Task formulations, methods and future
directions, in: Proceedings of the 27th International Conference on Computational
Linguistics, COLING ’18, Association for Computational Linguistics, Santa Fe, NM, USA,
2018, pp. 3346–3359. URL: http://www.aclweb.org/anthology/C18-1283.
[28] Y. Li, J. Gao, C. Meng, Q. Li, L. Su, B. Zhao, W. Fan, J. Han, A survey on truth discovery,
SIGKDD Explor. Newsl. 17 (2016) 1–16. URL: http://doi.acm.org/10.1145/2897350.2897352.
doi:10.1145/2897350.2897352.
[29] M. Hardalov, A. Arora, P. Nakov, I. Augenstein, A survey on stance detection for mis-
and disinformation identification, arXiv/2103.00242 (2021).
[30] P. Nakov, D. Corney, M. Hasanain, F. Alam, T. Elsayed, A. Barrón-Cedeño, P. Papotti,
S. Shaar, G. Da San Martino, Automated fact-checking for assisting human fact-checkers,
in: Proceedings of the 30th International Joint Conference on Artificial Intelligence,
IJCAI ’21, 2021, pp. 4551–4558.
[31] P. Nakov, H. T. Sencar, J. An, H. Kwak, A survey on predicting the factuality and the bias
of news media, arXiv/2103.12506 (2021).
[32] F. Alam, S. Cresci, T. Chakraborty, F. Silvestri, D. Dimitrov, G. D. S. Martino, S. Shaar,
H. Firooz, P. Nakov, A survey on multimodal disinformation detection, arXiv/2103.12541
(2021).
[33] P. Nakov, V. Nayak, K. Dent, A. Bhatawdekar, S. M. Sarwar, M. Hardalov, Y. Dinkov,
D. Zlatkova, G. Bouchard, I. Augenstein, Detecting abusive language on online platforms:
A critical analysis, arXiv/2103.00153 (2021).
[34] P. Atanasova, L. Màrquez, A. Barrón-Cedeño, T. Elsayed, R. Suwaileh, W. Zaghouani,
S. Kyuchukov, G. Da San Martino, P. Nakov, Overview of the CLEF-2018 CheckThat! lab
on automatic identification and verification of political claims, task 1: Check-worthiness,
in: CLEF 2018 Working Notes. Working Notes of CLEF 2018 - Conference and Labs of the
Evaluation Forum, CEUR Workshop Proceedings, 2018.
[35] P. Atanasova, P. Nakov, G. Karadzhov, M. Mohtarami, G. Da San Martino, Overview of
the CLEF-2019 CheckThat! lab on automatic identification and verification of claims.
Task 1: Check-worthiness, in: [97], 2019.
[36] A. Barrón-Cedeño, T. Elsayed, R. Suwaileh, L. Marquez, P. Atanasova, W. Zaghouani,
S. Kyuchukov, G. Da San Martino, P. Nakov, Overview of the CLEF-2018 CheckThat!
lab on automatic identification and verification of political claims. Task 2: Factuality, in:
[98], 2018.
[37] T. Elsayed, P. Nakov, A. Barrón-Cedeño, M. Hasanain, R. Suwaileh, G. Da San Martino,
P. Atanasova, CheckThat! at CLEF 2019: Automatic identification and verification of
claims, in: L. Azzopardi, B. Stein, N. Fuhr, P. Mayr, C. Hauff, D. Hiemstra (Eds.), Advances
in Information Retrieval, ECIR ’19, 2019, pp. 309–315.
[38] M. Hasanain, R. Suwaileh, T. Elsayed, A. Barrón-Cedeño, P. Nakov, Overview of the
CLEF-2019 CheckThat! lab on automatic identification and verification of claims. Task 2:
Evidence and factuality, in: [97], 2019.
[39] P. Nakov, A. Barrón-Cedeño, T. Elsayed, R. Suwaileh, L. Màrquez, W. Zaghouani,
P. Atanasova, S. Kyuchukov, G. Da San Martino, Overview of the CLEF-2018 CheckThat!
lab on automatic identification and verification of political claims, in: Proceedings of
the Ninth International Conference of the CLEF Association: Experimental IR Meets
Multilinguality, Multimodality, and Interaction, Lecture Notes in Computer Science, 2018,
pp. 372–387.
[40] P. Gencheva, P. Nakov, L. Màrquez, A. Barrón-Cedeño, I. Koychev, A context-aware
approach for detecting worth-checking claims in political debates, in: Proceedings of
the International Conference Recent Advances in Natural Language Processing, RANLP
2017, 2017, pp. 267–276.
[41] N. Hassan, C. Li, M. Tremayne, Detecting check-worthy factual claims in presidential
debates, in: Proceedings of the 24th ACM International on Conference on Information
and Knowledge Management, CIKM ’15, 2015, pp. 1835–1838.
[42] I. Jaradat, P. Gencheva, A. Barrón-Cedeño, L. Màrquez, P. Nakov, ClaimRank: Detecting
check-worthy claims in Arabic and English, in: Proceedings of the 2018 Conference of
the North American Chapter of the Association for Computational Linguistics: Demon-
strations, 2018, pp. 26–30.
[43] S. Vasileva, P. Atanasova, L. Màrquez, A. Barrón-Cedeño, P. Nakov, It takes nine to
smell a rat: Neural multi-task learning for check-worthiness prediction, in: Proceedings
of the International Conference on Recent Advances in Natural Language Processing,
RANLP ’19, 2019, pp. 1229–1239.
[44] Y. S. Kartal, M. Kutlu, Re-think before you share: A comprehensive study on prioritizing
check-worthy claims, IEEE Transactions on Computational Social Systems (2022).
[45] S. Shaar, N. Babulkov, G. Da San Martino, P. Nakov, That is a known lie: Detecting
previously fact-checked claims, in: Proceedings of the 58th Annual Meeting of the
Association for Computational Linguistics, ACL ’20, 2020, pp. 3607–3618.
[46] P. Nakov, D. Corney, M. Hasanain, F. Alam, T. Elsayed, A. Barrón-Cedeño, P. Papotti,
S. Shaar, G. D. S. Martino, Automated fact-checking for assisting human fact-checkers
(2021).
[47] S. Shaar, F. Alam, G. D. S. Martino, P. Nakov, The role of context in detecting previously
fact-checked claims, arXiv:2104.07423 (2021).
[48] I. Augenstein, C. Lioma, D. Wang, L. Chaves Lima, C. Hansen, C. Hansen, J. G. Simonsen,
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 2019, 2019, pp. 4685–4697.
[49] G. Karadzhov, P. Nakov, L. Màrquez, A. Barrón-Cedeño, I. Koychev, Fully automated fact
checking using external sources, in: Proceedings of the International Conference Recent
Advances in Natural Language Processing, RANLP 2017, 2017, pp. 344–353.
[50] M. Mohtarami, R. Baly, J. Glass, P. Nakov, L. Màrquez, A. Moschitti, Automatic stance
detection using end-to-end memory networks, in: Proceedings of the 2018 Conference of
the North American Chapter of the Association for Computational Linguistics: Human
Language Technologies, NAACL-HLT 2018, 2018, pp. 767–776.
[51] M. Mohtarami, J. Glass, P. Nakov, Contrastive language adaptation for cross-lingual
stance detection, 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 2019, 2019, pp. 4442–4452.
[52] P. Atanasova, P. Nakov, L. Màrquez, A. Barrón-Cedeño, G. Karadzhov, T. Mihaylova,
M. Mohtarami, J. Glass, Automatic fact-checking using context and discourse information,
Journal of Data and Information Quality (JDIQ) 11 (2019) 12.
[53] C. Castillo, M. Mendoza, B. Poblete, Information credibility on twitter, in: S. Srinivasan,
K. Ramamritham, A. Kumar, M. P. Ravindra, E. Bertino, R. Kumar (Eds.), Proceedings of
the 20th International Conference on World Wide Web, WWW 2011, 2011, pp. 675–684.
[54] D. Kopev, A. Ali, I. Koychev, P. Nakov, Detecting deception in political debates using
acoustic and textual features, in: Proceedings of the IEEE Automatic Speech Recognition
and Understanding Workshop, ASRU ’19, 2019, pp. 652–659.
[55] H. Rashkin, E. Choi, J. Y. Jang, S. Volkova, Y. Choi, 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, 2017, pp. 2931–2937.
[56] R. Baly, G. Karadzhov, D. Alexandrov, J. Glass, P. Nakov, Predicting factuality of reporting
and bias of news media sources, in: Proceedings of the 2018 Conference on Empirical
Methods in Natural Language Processing, 2018, pp. 3528–3539.
[57] R. Baly, M. Mohtarami, J. Glass, L. Màrquez, A. Moschitti, P. Nakov, Integrating stance
detection and fact checking in a unified corpus, in: Proceedings of the 2018 Conference
of the North American Chapter of the Association for Computational Linguistics: Human
Language Technologies, Volume 2 (Short Papers), 2018, pp. 21–27.
[58] V. Nguyen, K. Sugiyama, P. Nakov, M. Kan, FANG: leveraging social context for fake
news detection using graph representation, in: M. d’Aquin, S. Dietze, C. Hauff, E. Curry,
P. Cudré-Mauroux (Eds.), CIKM ’20: The 29th ACM International Conference on Infor-
mation and Knowledge Management, Virtual Event, Ireland, October 19-23, 2020, 2020,
pp. 1165–1174.
[59] K. Popat, S. Mukherjee, J. Strötgen, G. Weikum, Where the truth lies: Explaining the
credibility of emerging claims on the web and social media, in: Proceedings of the
26th International Conference on World Wide Web Companion, WWW ’17, 2017, pp.
1003–1012.
[60] J. Thorne, A. Vlachos, C. Christodoulopoulos, A. Mittal, FEVER: a large-scale dataset for
fact extraction and VERification, in: Proceedings of the 2018 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language
Technologies, Volume 1 (Long Papers), 2018, pp. 809–819.
[61] A. Patwari, D. Goldwasser, S. Bagchi, TATHYA: A multi-classifier system for detecting
check-worthy statements in political debates, in: E. Lim, M. Winslett, M. Sanderson,
A. W. Fu, J. Sun, J. S. Culpepper, E. Lo, J. C. Ho, D. Donato, R. Agrawal, Y. Zheng,
C. Castillo, A. Sun, V. S. Tseng, C. Li (Eds.), Proceedings of the 2017 ACM on Conference
on Information and Knowledge Management, CIKM, 2017, pp. 2259–2262.
[62] S. Shaar, F. Alam, G. Da San Martino, A. Nikolov, W. Zaghouani, P. Nakov, A. Feldman,
Findings of the nlp4if-2021 shared tasks on fighting the covid-19 infodemic and censor-
ship detection, in: Proceedings of the Fourth Workshop on NLP for Internet Freedom:
Censorship, Disinformation, and Propaganda, 2021, pp. 82–92.
[63] R. Agez, C. Bosc, C. Lespagnol, J. Mothe, N. Petitcol, IRIT at CheckThat! 2018, in: [98],
2018.
[64] B. Ghanem, M. Montes-y Gómez, F. Rangel, P. Rosso, UPV-INAOE-Autoritas - Check
That: Preliminary approach for checking worthiness of claims, in: [98], 2018.
[65] C. Hansen, C. Hansen, J. Simonsen, C. Lioma, The Copenhagen team participation in the
check-worthiness task of the competition of automatic identification and verification of
claims in political debates of the CLEF-2018 fact checking lab, in: [98], 2018.
[66] C. Zuo, A. Karakas, R. Banerjee, A hybrid recognition system for check-worthy claims
using heuristics and supervised learning, in: [98], 2018.
[67] B. Altun, M. Kutlu, TOBB-ETU at CLEF 2019: Prioritizing claims based on check-
worthiness, in: CLEF 2019 Working Notes. Working Notes of CLEF 2019 - Conference
and Labs of the Evaluation Forum, CEUR Workshop Proceedings, 2019.
[68] L. Coca, C.-G. Cusmuliuc, A. Iftene, CheckThat! 2019 UAICS, in: CLEF 2019 Working
Notes. Working Notes of CLEF 2019 - Conference and Labs of the Evaluation Forum,
CEUR Workshop Proceedings, 2019.
[69] R. Dhar, S. Dutta, D. Das, A hybrid model to rank sentences for check-worthiness, in:
CLEF 2019 Working Notes. Working Notes of CLEF 2019 - Conference and Labs of the
Evaluation Forum, CEUR Workshop Proceedings, 2019.
[70] L. Favano, M. Carman, P. Lanzi, TheEarthIsFlat’s submission to CLEF’19 CheckThat!
challenge, in: CLEF 2019 Working Notes. Working Notes of CLEF 2019 - Conference and
Labs of the Evaluation Forum, CEUR Workshop Proceedings, 2019.
[71] J. Gasior, P. Przybyła, The IPIPAN team participation in the check-worthiness task of the
CLEF2019 CheckThat! lab, in: CLEF 2019 Working Notes. Working Notes of CLEF 2019 -
Conference and Labs of the Evaluation Forum, CEUR Workshop Proceedings, 2019.
[72] C. Hansen, C. Hansen, J. Simonsen, C. Lioma, Neural weakly supervised fact check-
worthiness detection with contrastive sampling-based ranking loss, in: CLEF 2019
Working Notes. Working Notes of CLEF 2019 - Conference and Labs of the Evaluation
Forum, CEUR Workshop Proceedings, 2019.
[73] S. Mohtaj, T. Himmelsbach, V. Woloszyn, S. Möller, The TU-Berlin team participation
in the check-worthiness task of the CLEF-2019 CheckThat! lab, in: CLEF 2019 Working
Notes. Working Notes of CLEF 2019 - Conference and Labs of the Evaluation Forum,
CEUR Workshop Proceedings, 2019.
[74] T. Su, C. Macdonald, I. Ounis, Entity detection for check-worthiness prediction: Glasgow
Terrier at CLEF CheckThat! 2019, in: CLEF 2019 Working Notes. Working Notes of CLEF
2019 - Conference and Labs of the Evaluation Forum, CEUR Workshop Proceedings, 2019.
[75] J. Martinez-Rico, L. Araujo, J. Martinez-Romo, NLP&IR@UNED at CheckThat! 2020:
A preliminary approach for check-worthiness and claim retrieval tasks using neural
networks and graphs, in: [99], 2020.
[76] T. McDonald, Z. Dong, Y. Zhang, R. Hampson, J. Young, Q. Cao, J. Leidner, M. Stevenson,
The University of Sheffield at CheckThat! 2020: Claim identification and verification on
Twitter, in: [99], 2020.
[77] Y. S. Kartal, M. Kutlu, TOBB ETU at CheckThat! 2020: Prioritizing English and Arabic
claims based on check-worthiness, in: [99], 2020.
[78] E. Williams, P. Rodrigues, V. Novak, Accenture at CheckThat! 2020: If you say so:
Post-hoc fact-checking of claims using transformer-based models, in: [99], 2020.
[79] M. Hasanain, T. Elsayed, bigIR at CheckThat! 2020: Multilingual BERT for ranking Arabic
tweets by check-worthiness, in: [99], 2020.
[80] G. S. Cheema, S. Hakimov, R. Ewerth, Check_square at CheckThat! 2020: Claim detection
in social media via fusion of transformer and syntactic features, in: [99], 2020.
[81] A. Hussein, A. Hussein, N. Ghneim, A. Joukhadar, DamascusTeam at CheckThat! 2020:
Check worthiness on Twitter with hybrid CNN and RNN models, in: [99], 2020.
[82] I. Touahri, A. Mazroui, EvolutionTeam at CheckThat! 2020: Integration of linguistic and
sentimental features in a fake news detection approach, in: [99], 2020.
[83] R. Alkhalifa, T. Yoong, E. Kochkina, A. Zubiaga, M. Liakata, QMUL-SDS at CheckThat!
2020: Determining COVID-19 tweet check-worthiness using an enhanced CT-BERT with
numeric expressions, in: [99], 2020.
[84] A. Nikolov, G. Da San Martino, I. Koychev, P. Nakov, Team_Alex at CheckThat! 2020:
Identifying check-worthy tweets with transformer models, in: [99], 2020.
[85] E. Williams, P. Rodrigues, S. Tran, Accenture at CheckThat! 2021: Interesting claim
identification and ranking with contextually sensitive lexical training data augmentation,
in: [96], 2021.
[86] X. Zhou, B. Wu, P. Fung, Fight for 4230 at CLEF CheckThat! 2021: Domain-specific
preprocessing and pretrained model for ranking claims by check-worthiness, in: [96],
2021.
[87] F. Alam, S. Cresci, T. Chakraborty, F. Silvestri, D. Dimitrov, G. D. S. Martino, S. Shaar,
H. Firooz, P. Nakov, A survey on multimodal disinformation detection, arXiv:2103.12541
(2021).
[88] H. Rashkin, E. Choi, J. Y. Jang, S. Volkova, Y. Choi, 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, Association for Computational
Linguistics, Copenhagen, Denmark, 2017, pp. 2931–2937. URL: https://www.aclweb.org/
anthology/D17-1317. doi:10.18653/v1/D17-1317.
[89] N. Vo, K. Lee, The rise of guardians: Fact-checking URL recommendation to combat
fake news, in: K. Collins-Thompson, Q. Mei, B. D. Davison, Y. Liu, E. Yilmaz (Eds.), The
41st International ACM SIGIR Conference on Research & Development in Information
Retrieval, SIGIR 2018, Ann Arbor, MI, USA, July 08-12, 2018, ACM, 2018, pp. 275–284.
URL: https://doi.org/10.1145/3209978.3210037. doi:10.1145/3209978.3210037.
[90] R. Baly, G. Karadzhov, J. An, H. Kwak, Y. Dinkov, A. Ali, J. Glass, P. Nakov, What was writ-
ten vs. who read it: News media profiling using text analysis and social media context, in:
Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics,
ACL ’20, Association for Computational Linguistics, Online, 2020, pp. 3364–3374. URL:
https://aclanthology.org/2020.acl-main.308. doi:10.18653/v1/2020.acl-main.308.
[91] P. Atanasova, J. G. Simonsen, C. Lioma, I. Augenstein, Generating fact checking expla-
nations, in: Proceedings of the 58th Annual Meeting of the Association for Computa-
tional Linguistics, ACL ’20, Association for Computational Linguistics, Online, 2020, pp.
7352–7364. URL: https://aclanthology.org/2020.acl-main.656. doi:10.18653/v1/2020.
acl-main.656.
[92] A. Sathe, S. Ather, T. M. Le, N. Perry, J. Park, Automated fact-checking of claims from
Wikipedia, in: Proceedings of the 12th Language Resources and Evaluation Conference,
ACL ’20, European Language Resources Association, Marseille, France, 2020, pp. 6874–
6882. URL: https://aclanthology.org/2020.lrec-1.849.
[93] S. Sharma, F. Alam, M. Akhtar, D. Dimitrov, G. D. S. Martino, H. Firooz, A. Halevy,
F. Silvestri, P. Nakov, T. Chakraborty, et al., Detecting and understanding harmful memes:
A survey, arXiv preprint arXiv:2205.04274 (2022).
[94] B. Haidar, M. Chamoun, F. Yamout, Cyberbullying detection: A survey on multilingual
techniques, in: EMS, 2016, pp. 165–171. doi:10.1109/EMS.2016.037.
[95] F. Husain, O. Uzuner, A survey of offensive language detection for the Arabic language,
TALLIP 20 (2021) 1–44.
[96] G. Faggioli, N. Ferro, A. Joly, M. Maistro, F. Piroi (Eds.), CLEF 2021 Working Notes.
Working Notes of CLEF 2021–Conference and Labs of the Evaluation Forum, 2021.
[97] L. Cappellato, N. Ferro, D. Losada, H. Müller (Eds.), Working Notes of CLEF 2019 Confer-
ence and Labs of the Evaluation Forum, CEUR Workshop Proceedings, 2019.
[98] L. Cappellato, N. Ferro, J.-Y. Nie, L. Soulier (Eds.), Working Notes of CLEF 2018–Conference
and Labs of the Evaluation Forum, CEUR Workshop Proceedings, 2018.
[99] L. Cappellato, C. Eickhoff, N. Ferro, A. Névéol (Eds.), CLEF 2020 Working Notes, CEUR
Workshop Proceedings, 2020.
[100] M. Müller, M. Salathé, P. E. Kummervold, Covid-twitter-bert: A natural language pro-
cessing model to analyse covid-19 content on twitter, arXiv preprint arXiv:2005.07503
(2020).
A. Summary of the Approaches
Here we give a brief description of the approaches explored by the different teams for all four
subtasks.
Team AI Rational [15] (1A-bg:3 1B-bg:1 1C-bg:1 1D-bg:1 1A-nl:2 1B-nl:1 1C-nl:2 1D-nl:1 1A-en:1
1B-en:4 1B-en:2 1C-en:2 1D-en:3 1B-tr:2 1C-tr:3 1D-tr:1) experimented with different transformer
models: DistilBERT, BERT, RoBERTa. It used DistilBERT to find the best parameters for the
model and in the final the model they used RoBERTa large for English and XLM-RoBERTa large
for Bulgarian, Dutch and Turkish. For data transformation, all links from tweets were replaced
with “@link”. The proposed system used data augmentation using back translation for English
and Bulgarian. For the English tweets texts were translated to French then back translated
to English and combined to the training dataset. The Bulgarian text are back translated with
English, then combined with the training set.
Team ARC-NLP [16] (1C-en:6 1A-tr:3 1B-tr:3 1C-tr:1) proposed an approach called ARC-NLP-
contra, in which the idea is that harmful tweets contradict with the real-life facts in the scope of
COVID-19 pandemic. In addition, authors explore two other models. The first model, called
ARC-NLP-hc, which utilizes hand-crafted tweet and user features. The second model, called
ARC-NLP-pretrain, which pretrains a transformer-based model by using COVID-related Turkish
tweets.
Team Asatya [17] (1A-en:10 1B-en:2 1C-en:3) proposes a methodology for the subtasks 1A, 1B
and 1C, which includes data augmentation to increase the size of our training dataset, followed
by preprocessing of the tweets and feature extraction for the tweet. Authors used a multimodal
model, which uses numerical and categorical features in addition to the textual data from tweets.
The multimodal network is combined with BERT for training the model.
Team COURAGE [24] (1C-en:8) proposed a deep learning model based on graph machine
learning (i.e. Graph Attention Convolution) and a pretrained transformer-based model (i.e.
ELECTRA). The representation of each tweet into a graph starts with text preprocessing (i.e.,
lowercasing, removing url, user mention and hashtag symbol), and Part Of Speech (POS) tagging.
The POS-tagged tweet is then transformed into an undirected and attributed graph using a
window equal to 3 to populate the adjacency matrix and using ELECTRA word embedding
as nodes’ attribute. The model reaches average performance on the English test set, but it
improves F1-Score for positive class with respect to both the baseline and the simple ELECTRA
embedding.
Team iCompass [19] (1A-ar:3) finetuned pre-trained language models such as AraBERT and
ARBERT. The first model consisted of adding stacked gated recurrent units and one-dimensional
convolutional neural networks to ARBERT and finding the optimal configuration, dropout rates,
and training strategy to classify the tweet as harmful or normal. The second model was
composed of a gated recurrent network layer, a dense layer, and a dropout layer on the top of
the pre-trained AraBERT (V1) model to predict whether a tweet is worth fact-checking.
Fraunhofer SIT [18] (1A-en:5) used an ensemble classification method that took advantage of
state-of-the-art transformer networks and semi-supervised learning using GAN-BERT as well as
data augmentation and data preprocessing. The ensemble classifier consisted of fine-tuned BERT-
base-cased, BERTweet and RoBERTa-base5 models that were trained using cross-validation on
the training and validation split of the released dataset. Similarly, the BERT-base-cased and
RoBERTa-base models were fine-tuned using GAN-BERT, while include additional unlabeled
training data. Using a meta-classifier, the classification system was able to rank fifth best in
the competition. Early experiments with quantum natural language processing (QNLP)6 were
used. However, the current state of the technique (i.e., QNLP) posed some problems and was
therefore not included in the final model.
Team NUS-IDS [4] (1A-ar:1 1A-bg:1 1A-nl:1 1A-en:8 1A-es:1) describes the system CheckthaT5,
which was designed in the context of the CheckThat! lab 2022 competition at CLEF. CheckthaT5
explores the feasibility of adapting sequence-to-sequence models for detecting check-worthy
social media content in a multilingual texts (Arabic, Bulgarian, Dutch, English, Spanish and
Turkish) provided in the competition. CheckThaT5 system takes all languages as input uniformly,
thus enabling knowledge transfer from high-resource languages to low-resource languages.
Empirically, CheckthaT5 outperforms strong baselines in all low-resource languages. In addition,
the system incorporates tasks based on non-textual features that complement tweets and other
related CheckThat! 2022 tasks through multitask learning further improving the average
classification performance by 3%.
Team PoliMi-FlatEarthers [14] (1A-ar:5 1A-bg:5 1A-nl:4 1A-en:3 1B-en:1 1C-en:10 1D-en:7 1A-es:2)
propose a system which is based on GPT-3, which outperforms previous language models on the
task of finding relevant tweets. Though GPT-3 model is originally trained on English, however,
it shows competitive performances on other languages as well.
Team RUB-DFL [20] (1A-en:6 1B-en:6 1C-en:9 1A-tr:1 1B-tr:1 1C-tr:2) used transformer-based
pre-trained language models, as well as ELMo embeddings, which are combined with a range
of linguistic features in attention networks. They tried to include some forms of pre-processing
with LIWC features and URL resolution. In the end, the best results were achieved with fine-
tuned transformer-based models for both English and Turkish in subtasks 1A, 1B, and 1C.
Team TOBB ETU [13] (1A-ar:2 1B-ar:1 1C-ar:2 1D-ar:2 1A-bg:2 1B-bg:2 1C-bg:2 1D-bg:2 1A-nl:3
1B-nl:2 1C-nl:1 1D-nl:2 1A-en:4 1B-en:9 1C-en:5 1D-en:4 1A-tr:3 1B-tr:4 1C-tr:4 1D-tr:3) participated
in all subtasks for Arabic, Bulgarian, Dutch, English, and Turkish, yielding 20 submissions
in total. They investigated fine-tuning transformer models pre-trained on various texts. In
5
https://huggingface.co/cardiffnlp/twitter-roberta-base-2021-124m
6
https://github.com/CQCL/lambeq
addition, they explored two data augmentation methods including i) machine translating labeled
datasets in other languages to the corresponding language and ii) back-translation. Furthermore,
as another approach, they represent tweets using BERT embeddings and train a feedforward
neural network with Manifold Mixup regularization [5]. Based on their experiments on the test
development dataset for each subtask and language, they submitted one of the three methods
accordingly: i) a transformer model fine-tuned with the original data, ii) a transformer model
fine-tuned with the dataset augmented by back-translation, iii) a feedforward neural network
trained with Manifold Mixup regularization.
Team VTU BGM [21] (1A-en:11 1B-en:7 1C-en:12) used autoregressive model XLNet for feature
extraction and SVM for the classification of tweets.
Team Z-Index [22] (1A-nl:5 1A-en:12 1A-es:3 ) preprocessed the claims data by removing user-
name, URLs, non-ASCII characters, and stopwords. They experimented with both deep learning
and traditional approaches. For deep learning approach, transformer based models BERT multi-
lingual and XLM-RoBERTa base were used for training. For Dutch and English they obtained
better results using transformer based model and for Spanish they obtained better results using
SVM and Random Forest.
Team Zorros [23] (1A-en:2) used fine-tuned and pre-trained transformer-based models like
BERT and RoBERTa. They built ensemble models which combine the fine-tuned transformer
models. The preprocessing step in their system include removing URLs, hashtags, numbers and
other symbols. For subtask 1A, check-worthiness of tweets, authors used an ensemble of ten
transformer-based models, pre-trained on tweets about COVID-19. A classification header and
a dropout layer is used to avoid over-fitting. For subtask 1B (verifiable factual claims detection)
and 1C (Harmful tweet detection) they used an ensemble of five transformer-base models. The
ensemble models obtained a better performance than simple fine-tuned transformer models.
For subtask 1D (attention-worthy tweet detection) they used COVID Twitter BERT v2[100].