Misleading information in Spanish: a survey

Abstract

Misleading information spread on social networks is often supported by activists who promote this type of information and bots that amplify their visibility. The need for useful and timely mechanisms of credibility assessment in social media has become increasingly indispensable. Efforts to tackle this problem in Spanish are growing. The last years have witnessed many efforts to develop methods to detect fake news, rumors, stances, and bots on the Spanish social web. This work leads to a systematic review of the literature that relates the efforts to develop this area in the Spanish language. The work identifies pending tasks for this community and challenges that require coordination among the leading investigators on the subject.

Appendix

1.1 Review planning

The review planning step starts by defining search keywords, which will retrieve the first body of literature. These search keywords were defined using logical AND and OR connectors to control the coverage of documents matched by the search system. The search strings used for this process include the three variants of the problem that are the object of this study: stance, rumors, and bots. To locate the Spanish-speaking community’s results, we added the keyword Spanish to each of these terms. We also include Twitter as a keyword, the social media platform that concentrates the most cited studies in English (Zhang and Ghorbani 2020). To avoid restricting the search results to English publications, we also use these search strings in Spanish. The set of search strings used in the review is shown in Table 1.

Table 1 Search strings used in the SLR, A total of 4506 results were retrieved using these search strings

The search in Scopus was restricted to works published since 2009, ruling out works of rumors not related to this phenomenon’s explosion in social media. The works were restricted to two specific areas of knowledge: Computer Science and Engineering. In this way, the retrieved papers will include works on automatic detection methods, which is the focus of this study.

The review planning process also considers the definition of inclusion/exclusion criteria. These criteria are subsequently used in the literature screening stage, during which the content of the works retrieved during the search phase is reviewed. The works that meet the inclusion criteria and do not match any exclusion criteria are included within the literature’s definitive body. We first define a list of exclusion criteria with four items:

• Exclusion criteria 1 (ExCr1): When an article appears in more than one search, it will be considered only once. Accordingly, the articles repeated in the search results are eliminated, as well as versions of the same work published in different media (duplication by media).

• Exclusion criteria 2 (ExCr2): Articles written in a language other than Spanish or English are not considered.

• Exclusion criteria 3: Reviews (ExCr3-a), editorials (ExCr3-b), notes and erratum (ExCr3-c), and conference reviews ((ExCr3-d)) are not considered.

• Exclusion criteria 4 (ExCr4): Articles whose title or abstract do not refer to the study (semantic mismatch) are discarded.

The inclusion criteria consider two items:

• Inclusion criteria 1 (InCr1): Three sections of the work are reviewed. These are abstract, introduction, and conclusion. We verify if the work focuses on solving any of the tasks object of this study in Spanish.

• Inclusion criteria 2 (InCr2): If there is no conclusive evidence identified when applying inclusion criteria 1, the full article is read. If the work does not address any of the tasks in the Spanish language, the paper is discarded.

1.2 Literature search and screening

The search for papers was carried out during 2020. By applying the search strings to Scopus, we retrieved a total of 4506 documents. This first body of literature was examined, applying exclusion and inclusion criteria defined in this study. Figure 7 shows how many documents were deleted after applying the criteria. The reduction of the initial set is notorious. A total of 4360 documents were eliminated using the exclusion criteria. The remaining 146 documents were analyzed using inclusion criteria. The first inclusion criterion was validated in 102 documents, of which 67 also match the second inclusion criterion. As a result, the first body of literature records 67 documents.

Fig. 7

Exclusion/inclusion criteria applied to the documents detected in this SLR. The study considered two stages, the first based on the documents identified using search strings and the second based on the articles that cited the first body of literature. A total of 94 documents met the exclusion/inclusion criteria. Finally, after reviewing the selected documents’ references, 3 more papers were added to the survey validating the exclusion/inclusion criteria

The second body of literature was created by analyzing the works that cite the first body of literature. The citations include related work relevant to these articles, which provides an important source of papers connected to the survey subject that was not detected using search strings. A total of 194 documents were identified in this process, which was reduced to 69 after applying the exclusion criteria, and 27 after applying the inclusion criteria.

Both stages of the systematic review made it possible to identify a total of 94 documents. We conducted an exhaustive review of their references for these documents, looking for works related to this survey subject that had not been detected in the previous two stages. In this last process, three more papers were identified, which passed the exclusion criteria matching both inclusion criteria. In total, the SLR allowed the identification of 97 works related to the subject of this survey. The total number of papers per task is shown in Fig. 8.

Fig. 8

Papers per task

Table 2 Labels used to deploy the citation network (see Fig. 5) of the work surveyed in this study

1.3 Acronyms

• Systematic literature review: SLR

• Bag-of-Words: BOW

• Part-of-Speech: POS

• Term Frequency Inverted Document Frequency: TF-IDF

• Latent Semantic Analysis: LSA

• Universal Language Model Fine-Tuning: ULMFiT

• Singular Value Decomposition: SVD

• Recurrent Neural Network: RNN

• Bidirectional Encoder Representations based on the Transformer: BERT

• Supervised Autoencoder: SAE

• Pointwise Mutual Information: PMI

• Affective Norms for English Words: AFINN

• Linguistic Inquiry and Word Count: LIWC

• Named Entity Recognition: NER

• Global Vectors for word representation: GloVe

• Support Vector Machines: SVM

• Random Forests: RF

• Logistic Regression: LR

• Convolutional Neural Networks: CNN

• Long Short-Term Memory: LSTM

• Adaptive Boosting: ADABOOST

• Feed-Forward Neural Networks: FFNN

• Multinomial Naive Bayes: MNB

• Document frequency selection: DF

• Frequently co-occurring entropy: FCE

• Information Gain: IG

• Whale optimization: WO

• Genetic algorithms: GA

• Particle swarm optimization: PSO

• Hierarchical Attention Networks: HAN

• Bidirectional LSTM: Bi-LSTM

• Gated Recurrent Unit: GRU

• Spanish Billion Word Corpus and Embeddings: SBWCE

• The Catalonia Independence Corpus: CIC