Overview of the CLEF–2021 CheckThat! Lab on Detecting Check-Worthy Claims, Previously Fact-Checked Claims, and Fake News

Abstract

We describe the fourth edition of the CheckThat! Lab, part of the 2021 Conference and Labs of the Evaluation Forum (CLEF). The lab evaluates technology supporting tasks related to factuality, and covers Arabic, Bulgarian, English, Spanish, and Turkish. Task 1 asks to predict which posts in a Twitter stream are worth fact-checking, focusing on COVID-19 and politics (in all five languages). Task 2 asks to determine whether a claim in a tweet can be verified using a set of previously fact-checked claims (in Arabic and English). Task 3 asks to predict the veracity of a news article and its topical domain (in English). The evaluation is based on mean average precision or precision at rank k for the ranking tasks, and macro-F$_1$ for the classification tasks. This was the most popular CLEF-2021 lab in terms of team registrations: 132 teams. Nearly one-third of them participated: 15, 5, and 25 teams submitted official runs for tasks 1, 2, and 3, respectively.

B. Hamdan—Independent Researcher.

Appendices

Appendix

ASystems for Task 1

The positions in the task ranking appear after each team name. See Tables 5, 6 and 7 for further details.

Team Accenture [100] used BERT and RoBERTa with data augmentation. They further generated additional synthetic training data using lexical substitution. To find the most probable substitutions, they used BERT-based contextual embedding to create synthetic examples for the positive class. They further added a mean-pooling layer and a dropout layer on top of the model before the final classification layer.

Team Fight for 4230 [103] focused its efforts mostly on two fronts: the creation of a pre-processing module able to properly normalize the tweets and the augmentation of the data by means of machine translation and WordNet-based substitutions. The pre-processing included link removal and punctuation cleaning, as well as quantities and contractions expansion. All hashtags related to COVID-19 were normalized into one and the hashtags were expanded. Their best approach was based on BERTweet with a dropout layer and the above-mentioned pre-processing.

Team GPLSI [77] applied the RoBERTa and the BETO transformers together with different manually engineered features, such as the occurrence of dates and numbers or words from LIWC. A thorough exploration of parameters was made using weighting and bias techniques. They also tried to split the four-way classification into two binary classifications and one three-way classification. They further tried oversampling and undersampling.

Team iCompass used several prepossessing steps, including (i) English word removal, (ii) removing URLs and mentions, and (iii) data normalization, removing tashkeel and the letter madda from texts, as well as duplicates, and replacing some characters to prevent mixing. They proposed a simple ensemble of two BERT-based models, which include AraBERT and Arabic-ALBERT.

Team NLP&IR@UNED [49] used several transformer models, such as BERT, ALBERT, RoBERTa, DistilBERT, and Funnel-Transformer, for the experiments to compare the performance. For English, they obtained better results using BERT trained with tweets. For Spanish, they used Electra.

Team NLytics [74] used RoBERTa with a regression function in the final layer, approaching the problem as a ranking task.

Team QMUL-SDS [1] used the AraBERT preprocessing function to (i) replace URLs, email addressees, and user mentions with standard words, (ii) removed line breaks, HTML markup, repeated characters, and unwanted characters, such as emotion icons, and (iii) handled white spaces between words and digits (non-Arabic, or English), and/or a combination of both, and before and after two brackets, and also (iv) removed unnecessary punctuation. They addressed the task as a ranking problem, and fine-tuned an Arabic transformer (AraBERTv0.2-base) on a combination of the data from this year and the data from the CheckThat! lab 2020 (the CT20-AR dataset).

Team SCUoL [6] used typical pre-processing steps, including cleaning the text, segmentation, and tokenization. Their experiments consists of fine-tuning different AraBERT models, and their final results were obtained using AraBERTv2-base.

Team SU-NLP [22] also used several pre-possessing steps, including (i) removing emojis, hashtags, and (ii) replacing all mentions with a special token (@USER), and all URLs with the respective website’s domain. If the URL is for a tweet, they replaced the URL with TWITTER and the respective user account name. They reported that this URL expansion method improved the performance. Subsequently, they used an ensemble of BERTurk models fine-tuned using different seed values.

Team TOBB ETU [101] investigated different approaches to fine-tune transformer models including data augmentation using machine translation, weak supervision, and cross-lingual training. For their submission, they removed URLs and user mentions from the tweets, and fine-tuned a separate BERT-based models for each language. In particular, they fine-tuned BERTurk¹, AraBERT, BETO², and the BERT-base model for Turkish, Arabic, Spanish, and English, respectively. For Bulgarian, they fine-tune a RoBERTa model pre-trained with Bulgarian documents.³

Team UPV [14] used a multilingual sentence transformer representation (S-BERT) with knowledge distillation, originally intended for question answering. They further introduced an auxiliary language identification task, aside the downstream check-worthiness task.

Systems for Task 2

Team Aschern [25] used TF.IDF, fine-tuned pre-trained S-BERT, and the reranking LambdaMART model.

Team BeaSku [91] used triplet loss training to fine-tune S-BERT. Then, they used the scores predicted by the fine-tuned model along with BM25 scores as features to train a rankSVM re-ranker. They further discussed the impact of applying online mining of triplets. They also experimented with data augmentation.

Team DIPS [60] calculated S-BERT embeddings for all claims, then computed a cosine similarity for each pair of an input claim and a verified claim. The prediction is made by passing a sorted list of cosine similarities to a neural network.

Team NLytics approached the problem as a regression task, and used RoBERTa with a regression function in the final layer.

Systems for Task 3

Team Black Ops [92] performed data pre-processing by removing stop-words and punctuation marks. Then, they experimented with decision trees, random forest, and gradient boosting classifiers for Task 3A, and found the latter to perform best.

Team CIC [8] experimented with logistic regression, multi-layer perceptron, support vector machines, and random forest. Their experiments consisted of using stratified 5-fold cross-validation on the training data. Their best results were obtained using logistic regression for task 3A, and a multi-layer perceptron for task 3B.

Team CIC experimented with a decision tree, a random forest, and a gradient boosting algorithms. They found the latter to perform best.

Team CIVIC-UPM [48] participated in the two subtasks of task 3. They performed pre-processing, using a number of tools: (i) ftfy to repair Unicode and emoji errors, (ii) ekphrasis to perform lower-casing, normalizing percentages, time, dates, emails, phones, and numbers, (iii) contractions for abbreviation expansion, and (iv) NLTK for word tokenization, stop-words removal, punctuation removal and word lemmatization. Then, they combined doc2vec with transformer representations (Electra base, T5 small and T5 base, Longformer base, RoBERTa base and DistilRoBERTa base). They further used additional data from Kaggle’s Ag News task, Kaggle’s KDD2020, and Clickbait news detection competitions. Finally, they experimented with a number of classifiers such as Naïve Bayes, Random Forest, Logistic Regression with L1 and L2 regularization, Elastic Net, and SVMs. The best system for subtask 3A used DistilRoBERTa-base on the text body with oversampling and a sliding window for dealing with long texts. Their best system for task 3B used RoBERTa-base on the title+body text with oversampling but no sliding window.

Team DLRG experimented with a number of traditional approaches like Random Forest, Naïve Bayes and Logistic Regression as well as an online passive-aggressive classifier and different ensembles thereof. The best result was achieved by an ensemble of Naïve Bayes, Logistic Regression, and the Passive Aggressive classifier for task 3A. For task 3B, the Online Passive-Aggressive classifier outperformed all other approaches, including the considered ensembles.

Team GPLSI [77] applied the RoBERTa transformer together with different manually-engineered features, such as the occurrence of dates and numbers or words from LIWC. Both the title and the body were concatenated as a single sequence of words. Rather than going for a single multi-class setting, they used two binary models considering the most frequent classes: false vs. other, and true vs. other, followed by one three-class model.

Team MUCIC [12] used a majority voting ensemble with three BERT variants. They applied BERT, Distilbert, and RoBERTa, and fine-tuned the pre-trained models.

Team NITK_NLP [57] proposed an approach, that included pre-processing and tokenization of the news article, and then experimented with multiple transformer models. The final prediction was made by an ensemble.

Team NKovachevich [55] created lexical features. They extracted the 500 most frequent word stems in the dataset, and calculated the TF.IDF values, which they used in a multinomial Naïve Bayes classifier. A much better performance was achieved with an LSTM model that used GloVe embeddings. A little lower F1 value was achieved using BERT. They further found RoBERTa to perform worse than BERT.

Team NLP&IR@UNED [49] experimented with four transformer architectures and input sizes of 150 and 200 words. In the preliminary tests, the best performance was achieved by ALBERT with 200 words. They also experimented with combining TF.IDF values from the text, all the features provided by the LIWC tool, and the TF.IDF values from the first 20 domain names returned by a query to a search engine. Unlike what was obtained in the dev dataset, in the official competition, the best results were obtained with the approach based on TF.IDF, LIWC, and domain names.

Team NLytics fined-tuned RoBERTa on the dataset for each of the sub-tasks. Since the data is unbalanced, they used under-sampling. They also truncated the documents to 512 words to fit into the RoBERTa input size.

Team NoFake [56] applied BERT without fine-tuning, but used an extensive amount of additional data for training, downloaded from various fact-checking websites.

Team Pathfinder [96] participated in both tasks and used multinomial Naïve Bayes and random forest. The former performed better for both tasks. For task 3A, the they merged the classed false and partially false into one class, which boosted the model performance by 41% (a non-official score mentioned in the paper).

Team Probity addressed the multiclass fake news detection subtask, they used a simple LSTM architecture where they adopted word2vec embeddings to represent the news articles.

Team Qword [97] applied pre-processing techniques, which included stop-word removal, punctuation removal and lemmatization using a Porter stemmer. The TF.IDF values were calculated for the words. For these features, four classification algorithms were applied. The best result was given by Extreme Gradient Boosting.

Team SAUD used an SVM with TF.IDF. They tried Logistic Regression, Multinomial Naïve Bayes, and Random Forest, and found SVM to work best.

Team Sigmoid [76] experimented with different traditional machine learning approaches, with multinomial Naïve Bayes performing best, and one deep learning approach, namely an LSTM with the Adam optimizer. The latter outperformed the more traditional approaches.

Team Spider applies an LSTM, after a pre-processing consisting of stop-word removal and stemming.

Team UAICS [26] experimented with various models including BERT, LSTM, Bi-LSTM, and feature-based models. Their submitted model is a Gradient Boosting with a weighted combination of three feature groups: bi-grams, POS tags, and lexical categories of words.

Team University of Regensburg [41] used different fine-tuned variants of BERT with a linear layer on top and applied different approaches to address the maximum sequence length of BERT. Besides hierarchical transformer representations, they also experimented with different summarization techniques like extractive and abstractive summarization. They performed oversampling to address the class imbalance, as well as extractive (using DistilBERT) and abstractive summarization (using distil-BART-CNN-12-6), before performing classification using fine-tuned BERT with a hierarchical transformer representation.