MASARYK U N I V E R S I T Y FACULTY OF INFORMATICS Propaganda Detection using Stylometric Text Analysis Master's Thesis BC. RADOSLAV SABOL Brno, Fall 2022 MASARYK U N I V E R S I T Y FACULTY OF INFORMATICS Propaganda Detection using Stylometric Text Analysis Master's Thesis BC. RADOSLAV SABOL Advisor: doc. RNDr. Aleš Horák, Ph.D. Department of Machine Learning and Data Processing Brno, Fall 2022 Declaration Hereby I declare that this paper is my original authorial work, which I have worked out on my own. A l l sources, references, and literature used or excerpted during elaboration of this work are properly cited and listed in complete reference to the due source. Be. Radoslav Sabol Advisor: doc. RNDr. Aleš Horák, Ph.D. iii Acknowledgements I would like to express my gratitude to my supervisor, doc. Ales Horak and other NLP lab members for their guidance and insightful suggestions that made it achievable for me to write this thesis. M y special thanks also belongs to my family and friends for their constant support during the entire course of my academic journey. iv Abstract Propaganda campaigns aim to spread disinformation and manipulate the target audience. The accessibility and the viral nature of the Internet provide an ideal environment for spreading content that affects the population's mindsets to a specific agenda. This content is cleverly disguised as a reliable information source, requiring significant focus and background knowledge to uncover its manipulative nature. That is why developing automated text analysis tools is essential in protecting or warning users from manipulative content. The main goal of this thesis is to automatically detect the manipulative style of Czech documents by utilizing stylometric text features. The work is focused around detecting various manipulative techniques by exploiting a feature vector of approximately 20,000 numerical features that describe the writing style of a document. The presented approaches are benchmarked using the Czech Propaganda benchmark dataset, where the stylometric techniques slightly outperform other approaches in most dataset attributes. Keywords deep learning, experimental, implementation, machine learning, NLP, manipulative style recognition v Contents Introduction 1 1 Propaganda Detection Approaches 3 1.1 Propaganda Detection Task 3 1.1.1 Detection Methods 3 1.1.2 Manipulative Style Recognition Datasets . . . . 5 1.1.3 Related Works 6 1.2 Stylometry 7 1.2.1 Assumptions 7 1.2.2 Stylome Construction 8 1.2.3 Related Tasks 12 2 The Propaganda Dataset 14 2.1 Annotation Description 14 2.1.1 Manipulative Techniques 14 2.1.2 Document-Level Attributes 16 2.1.3 Other Atrributes 17 3 Stylometric Feature Extraction 19 3.1 Document Preprocessing 19 3.2 Proposed Features 19 3.2.1 Word and Sentence Length Distribution Analysis 21 3.2.2 Morphological Features 22 3.2.3 Word Repetition 23 3.2.4 Word Richness Metrics 23 3.2.5 Punctuation and Typography 24 3.2.6 Other Features 25 4 Manipulative Style Detection 27 4.1 Classical Machine Learning Methods 27 4.1.1 Data Augmentation for unbalanced Domains . . 28 4.2 Deep learning Methods 29 4.2.1 Architecture Description 29 4.2.2 Training Description 30 4.3 Feature Selection for Deep Neural Approaches 32 vi 5 Evaluation 34 5.1 Experiment Setup 34 5.2 Feature Importance 35 5.3 Results 36 5.3.1 Manipulative Techniques 37 5.3.2 Document-Level Properties 40 5.3.3 Other Attributes 41 5.3.4 Comparison with the Previous Results 43 5.4 Discussion 45 6 Conclusion and Future Work 48 Bibliography 49 A The Complete Table of Results 55 vii List of Tables 1.1 Description of selected manipulative devices 6 2.1 Example annotations for attributes at the span level. Range annotations are higlighted in bold 17 3.1 Overview of proposed stylometric text features 20 4.1 Gradient accumulation steps (GA) and batch size (BS) values for different pretrained models and graphics cards 31 5.1 Results for measurement of feature importances via Random Forests. AMIFS and ALIFS refer to difference in weighted F l score between the setup with the omitted feature set and the setup with all features 36 5.2 A n overview of the best-performing methods for manipulative techniques. The first row refers to the bestperforming technique without secondary features for each attribute, whereas the second row uses stylometric features, r. style refers to stylometry with features discarded using feature selection 38 5.3 A A n overview of the best-performing methods for documentlevel properties. The first row refers to the best-performing technique without secondary features for each attribute, whereas the second row uses stylometric features, r. style refers to stylometry with features discarded using feature selection 41 5.4 A n overview of the best-performing methods for other attributes. The first row refers to the best-performing technique without secondary features for each attribute, whereas the second row uses stylometric features, r. style refers to stylometry with features discarded using feature selection 43 5.5 A comparison of the best results in this thesis with the previous, unsubmitted ones 44 A . l The complete table of results 55 V l l l List of Figures 1.1 A n example of a part of a stylome that visualizes topical features of authors. [21] 9 15figure .caption. 18 3.1 A n example of prepared document. Sentence boundaries labeled as are omitted for clarity. 19 3.2 A n example of Majka morphological tag chunks. The chunks highlighted in green were included in the final features, while the red ones were discarded 22 3.3 Fixed typography feature example 25 4.1 Deep neural architecture for manipulative style detection 29 4.2 Performance benchmark of the deep learning models for various GPUs 31 4.3 Feature categories organized into a tree structure 32 5.1 Visual overview of results for manipulative techniques. Experiments with results that are 5% below the median are omitted from the visualization to highlight more promising approaches 37 5.2 Range length distribution in tokens for the attribute Labelling 39 5.3 Visual overview of results for document-level properties. Experiments with results that are 5% below the median are omitted from the visualization to highlight more promising approaches 40 5.4 Visual overview of results for other attributes. Experiments with results that are 5% below the median are omitted from the visualization to highlight more promising approaches 42 5.5 A visualization of training and validation metrics for the stylometric Slavic Bert experiment on the attribute Blaming 46 ix Introduction With the rise of the Internet and social media platforms, reaching out to a broader audience has become more straightforward. The ease is generally viewed as a positive development, allowing practically any user to express themselves publicly. However, an unforeseen consequence is that social media users are exposed to manipulative content from various sources. The term computational propaganda is a relatively recent phenomenon that rose to attention after the online political manipulation in the 2016 US Presidental elections [1]. Later on, in 2020, during the COVID-19 pandemic, there was a global outbreak of misinformative content. Information warfare is waged in tandem with the massively medialized ongoing conflict in Ukraine. Typical characteristics of such warfare consist of sharing false narratives about the ongoing war, demonization of the opponent to justify the aggression, or spreading public unrest. Unfortunately, propaganda disguised as news and politically oriented social media content is deemed by many users as a source of reliable content. The technical aspect of computational propaganda allows for the creation of algorithms that distribute manipulative content on a mass scale to a suitable audience, making it more efficient at influencing other users' mindsets to a specific agenda [2]. Due to the enormous scale and dynamic nature of the World Wide Web, manual detection of manipulative content is challenging to grasp. The development of automated tools can provide immense support in protecting the public from deceptive content. The goal of this thesis is to design, implement, and evaluate an automated method for the detection of manipulative content. Several additional criteria are given for the implementation. First, the module should be optimized for the Czech language. Second, the solution should utilize stylometric text features to recognize the writing style specific to various manipulative techniques. Chapter 1 is dedicated to the recent trends in the propaganda detection task and stylometric text analysis. The content of the chapter serves as a theoretical foundation upon which the system designed in this thesis is built. 1 INTRODUCTION In Chapter 2, the Propaganda benchmark dataset of manipulative news content is presented. The emphasis is given to the annotations of various manipulative techniques and their comparison with similar datasets. The proposed set of stylometric text features is described in Chapter 3. In Chapter 4, various machine learning methods are proposed that utilize stylometric text features. Other text representations are considered along with the proposed features for comparison purposes. The last chapter focuses on an in-depth evaluation of all proposed techniques and their comparison with other results. Overall, the stylometric feature set helps to improve the performance on the majority of present annotations, most notably on Source, Expert, and Emotions. In cases of significantly reduced performance, the annotation typically focuses on the semantics of the presented document, where features describing the writing style present noise to the learning process. 2 1 Propaganda Detection Approaches The following chapter gives an overview of concepts that are relevant to the propaganda detection task. The first section describes the propaganda task, available datasets, and the detection approaches used. The last section serves as a survey on stylometric analysis, as these techniques will be used throughout this thesis. It defines stylometry, assumptions that are based on it, stylome construction, and typical applications. 1.1 Propaganda Detection Task Research on the detection of manipulative techniques focuses on two primary aspects of the analyzed text. Disinformation, where the information presented is intentionally false and intended to deceive public opinion. Fake news can be considered as a related concept; however, here, the main focus is for the information to be false - without the necessity to influence the audience's opinion (satire, clickbait) [2]. The second aspect, where manipulation belongs, focuses on truthful events that were altered from their objective interpretation to affect the audience according to a specific agenda. 1.1.1 Detection Methods The immense volume and dynamic nature of the content on the Internet make it difficult to assess deceptive articles manually. For this reason, various automated methods have been proposed that can be divided into four categories [3], which will be discussed in the following paragraphs. Fact-checking based methods focus on verifying the document's content. The proposed method consists of two stages. The first stage deals with the extraction of facts from the candidate document. The second stage compares the extracted facts with a reliable and verified knowledge base in a machine-readable format. Currently deployed systems must be assisted by annotators who extract claims from documents which are then checked against knowledge graphs [3]. 3 l . PROPAGANDA DETECTION APPROACHES Source credibility verification attempts to identify unreliable subjects, as it is believed that unreliable subjects are more likely to post untrustworthy content. Most of these subjects consist of fake news websites pretending to be trustworthy news outlets. Currently proposed methods compute the source's trustworthiness using machine learning methods. The analysis is based mainly on content-based features as social popularity metrics can be manipulated [4]. Another approach, called CredRank [5], measures user credibility using his online behavior on social media. Information flow analysis studies how manipulative content is propagated through social media via interactions. The most prominent concept in this area is the propagation tree, where each node represents a post and each edge represents a non-symmetric relation between two nodes (for example, child is response to parent node). The root node represents the original post. The propagation patterns of deceptive content differ from the regular content, as manipulative content diffuses faster, farther, deeper, and more broadly the truthful one [6]. Manipulative style recognition, the method used in this thesis, is based on the knowledge that deceptive, manipulative content is formulated differently compared to mainstream media, which is theorized to be a subconscious decision [7]. This difference can be assessed using the surface features of the text. Unlike other techniques, this method does not need any external knowledge besides the presented document, which is one of the reasons why active research is still conducted in this area. Detection of manipulative style is generally considered a supervised classification task. Given an input document, the task is to determine whether the document is manipulative or not. A more accurate description of the task ranges from document classification to span detection of individual manipulative techniques, which depends on the annotations of the currently used dataset, which is further discussed in Section 1.1.2. This method builds on well-established methods for supervised classification in natural language processing. Notable examples include Long Short-term Memory Networks (LSTM) [8], Gated Recurrent Units (GRU) [9], and deep neural models based on transformer 4 l . PROPAGANDA DETECTION APPROACHES architecture like Bidirectional Encoder Representation from Transformers (BERT) [10]. In contrast, older methods train classifiers using hand-crafted stylometric text features of fixed size instead of operating on the text itself. Simple features range from word length, spelling errors, and punctuation to more complex ones derived from syntax and semantics. Section 1.2 of this chapter discusses a more in-depth description of stylometry principles. Combining both approaches is also possible - to merge machine learning methods that operate directly on the text and stylometric features as additional cues that extend the input [11]. 1.1.2 Manipulative Style Recognition Datasets A suitably large annotated dataset is one of the critical assumptions for successful supervised classification. As the propaganda detection task has a short research history, researchers had to cope with the need for annotated data [2]. A n early approach for creating an annotated dataset was by using distant supervision [12], which is a process of automatically creating weakly annotated data from an existing database. For example, Rashkin et al. [11] used distant supervision to assemble a balanced dataset TSHP-17 labeled with one of the four classes: trusted, hoax, satire, and propaganda. They used the US News & World Report database of manipulative news outlets to annotate the individual articles at the document level. This approach generates noisy datasets as manipulative news outlets may not always generate propagandistic content. Some may, for example, occasionally release objective articles to raise their credibility. More recent propaganda detection datasets divide propaganda into individual techniques commonly used in propaganda. These manipulative techniques are well-defined and well-established in the literature. Clyde R. Miller in 1937 [13] proposed the categorization of manipulation into seven devices that are still accepted by the scientific community to this date. Their description can be seen in Table 1.1. As an example, the Propaganda Techniques Corpus (PTC) dataset [14] annotates 18 different attributes ranging from techniques that leverage emotions (like loaded language, name calling, and exagarration) 5 l . PROPAGANDA DETECTION APPROACHES Table 1.1: Description of selected manipulative devices Technique Description Example Name Calling attack on a subject using an insulting label without the presentation of the actual meaning stereotyping the subject as a demagogue, dictator, communist, oligarch, etc. Card Stacking focus on one side of the perspective while suppressing the others cherry-picking and overemphasizing facts, manipulating statistics Glittering Generalities appeal to emotions using highly valued concepts without proper reasoning slogans, appeal to freedom, hope, patriotism, etc. The Band Wagon appeal on the general popularity of the idea phrases like "everyone thinks so", "it's a popular belief" Plain Folks the author presents himself as an ordinary person to better connect with the audience choice of clothing, referring to the crowd using pronouns like we, us Testimonial using a respected person to promote/demote certain idea or product celebrities in advertisements Transfer using a symbol that carries respect, prestige, or authority alongside the presented idea using the national flag, prestigious university seal, etc. to argumentation fallacies (like straw-man argument, bandwagon, and red herring). Instead of distant supervision, this dataset is manually annotated by domain experts. Another difference is that the individual techniques are labeled at the fragment level - instead of the entire document, a small text span is flagged instead. Classifiers built upon individual manipulative techniques may explain why the document is classified as deceptive, which is considered a positive development. 1.1.3 Related Works In 2019, a shared task was held in Workshop on NLP4IF: censorship, disinformation, and propaganda1 . The task consisted of two different problems based on the PTC dataset described in the previous subsection: 1. sentence-level binary classification on any of the 18 propaganda techniques 2. fragment-level multi-class multi-label classification, where each fragment needs a correct technique assigned 1. http://www.netcopia.net/nlp4if/2019/ 6 i . PROPAGANDA DETECTION APPROACHES Overall, 25 systems were evaluated, and the best-performing ones utilized BERT contextual embeddings. Other successful approaches exploited contextual embeddings of RoBERTa, ELMo, or context-independent representations based on lexical, sentiment, or TF-IDF features. Proppy is the first publicly available tool for document-level propaganda detection [15]. It is a pipeline of four separate modules that regularly monitor various news outlets and detect real-time events. The last module, Propaganda Index Computation, uses a maximum entropy classifier with L2 regularization to predict the probability of an article containing propagandistic content. The classifier is fed with various stylometric features split into four categories (word n-gram features, lexicon features, readability measures, and NEws LAndscape features (NELA)). The evaluation of proppy on [11] led to a significant improvement of 8.51% over the n-gram system proposed in their work. Baisa et al. [16] present the first results of style-based analysis on the Czech corpus Propaganda aimed at identifying manipulative techniques in Czech news outlets. They used ten machine learning methods to classify 18 different attributes, where 8 were manipulative techniques. Regularized linear regression and Support Vector Machines yielded the best results; however, the text representation needed to be enhanced by additional features to raise the accuracies above baselines. 1.2 Stylometry Computational stylometry is a linguistic discipline for automated analysis of an author's style through statistical analysis of texts. The most common applications include but are not limited to authorship attribution, author profiling, and plagiarism detection [17]. Stylometry was applied in multiple areas such as literary, historical, and forensic studies [18]. 1.2.1 Assumptions Many factors influence the overall style of the text. The document's topic is one of them, as it predetermines some of the used vocabu- 7 l . PROPAGANDA DETECTION APPROACHES lary. Document genre also modifies the text's overall presentation and writing style. However, some parts of the text are influenced by the author's sociological and psychological background. These may include the author's personality traits, mental health, age, gender, and education. The research in computer stylometry focuses on relationships between the author's properties and his writing style. Assume that the stylome is a set of features that describe the author's writing style. Then, human stylome hypothesis is based on the fact that each author's style is unique, like a fingerprint, so we can clearly distinguish between authors. However, some research questions still remained unanswered due to the lack of systematic studies. It is uncertain whether the style is invariant or it does change with the author's age and experience. Also, it is difficult to assess if the style is a collection of unconscious decisions or if it can be imitated to obfuscate the author's authentic style. If no human stylome that is robust enough for the problems mentioned above is found, the human stylome hypothesis should be rejected [19]. 1.2.2 Stylome Construction The choice of stylometric features is one of the most crucial tasks in computational stylometry. Unfortunately, the is little to no consensus on an optimal feature set, as it depends on currently solved problem and text domain. It is currently considered as one of the greatest problems in computational stylometry. Current research is focused on quantitative evaluation of feature quality, but little attention is given to interpretation of the results [20]. Stylome is a numerical vector constructed from thousands of features on different levels of analysis: lexical, syntactical, semantic, structural, subject-specific, or idiosyncratic levels [17]. The representation is extracted from the author's document using various linguistic analysis tools. The following pages will discuss the most commonly used features. A n example of such stylome can be seen on Figure 1.1. Lexical Features Lexical features can be observed at both character and word levels. At the character level, character n-grams are usually extracted from the 8 l . PROPAGANDA DETECTION APPROACHES topic 5 topic 16 topic 25 topic 43 topic 46 - gambling - commerce - food = electronic = fashion games communication game f 10 > r > r > c > game f 10 > food phone hair team store eat computer look games card chicken free little play item cream internet black win items cheese web wear poker price cake site dress casino money chocolate email bought V J I J ^ J ^ J poker_star gambling net andy telly A hcme_cookin LI t Ruby Red Vegan h 0 Culinarily Obsessed r Japan's Garage Sale i_love_to_shop :_L -I I c::: J=.LI 11:: i. Heave"!-. Figure 1.1: A n example of a part of a stylome that visualizes topical features of authors. [21] input document. The n itself varies between tasks and languages, but it is a common practice to include multiple n values inside the same feature set. The number of different n-grams can be overwhelming in larger corpora; that is why additional optimizations are required. These include filtering by the frequency of n-grams, not distinguishing between lower and upper case letters, and so on. The documents are usually viewed at the word level as a bag of word representation, omitting any information about the word order and context. Multiple features can be derived from words, for example, word length in characters and vocabulary richness [20]. The lexical level is the most commonly exploited level of stylometric analysis. This is due to language independence, straightforward interpretation of the features, and simple implementations that are not as computationally intensive. In the overview of PAN-2018 competition [22], it was shown that features consisting of character n-grams had outperformed more complex linguistic features. 9 5 10 15 20 25 30 35 40 45 50 Topic Index i . PROPAGANDA DETECTION APPROACHES Syntactic Features Syntactic features describe the sentence structure. The computation of syntactic features may vary in complexity. Simple surface features include word length distributions, punctiation presence, and punctuation placement, and are usually language independent. A domainspecific feature used in analyzing social media content like emoji presence is commonly used too [23]. More complex features require language-specific tools and resources to derive them. The quality of such features is highly dependent on the quality of syntactic analysis tools, as poor accuracy produces noisy feature values. Frequencies of function words and features obtained using part of speech tags belong in this category [24]. Feng et al. [25] used context-free grammar parse trees to encode both shallow and deep syntactic features for the deception detection task. Semantic Features At the semantic level of stylometric analysis, the goal is to capture the semantic meaning of words, phrases, or entire sentences [20]. Thanks to the current availability of tools that carry out complicated language understanding tasks like sentiment analysis, semantic analysis, and pragmatic analysis, the construction of semantic features is more ac- cessible. De Bruine et al. [24] used semantic features to detect autism spectrum disorder among Dutch adolescents. The first semantic feature used was Linguistic Inquiry and Word Count (LIWC) [26]. LIWC is a software standard for analyzing an author's psychological and social states based on word usage. The author's writeprint contained 68 features based on LIWC that were grouped into five more significant categories: linguistic dimensions, psychological processes, relativity, personal matters and other. The other semantic feature used was the number of ideas or propositions inside the text normalized by the number of sentences, also called idea density. The feature was detected using an automated tool for Dutch also used for the detection of Alzheimer's disease [27]. 10 i . PROPAGANDA DETECTION APPROACHES Structural Features The following group of features attempts to characterize the layout of the document. Cristani et al. [23], in authorship attribution of short texts and e-mails, structural features at the message level were presented like the presence of greetings, farewells, and signatures. Other promising features compute the average paragraph length and the number of paragraphs per document. The main downside of structural features is that they are sensitive to the current text domain. For example, the structure of the research paper differs heavily from the structure of informal e-mail. This sensitivity can be problematic in some tasks, as stylometry aims to model the author, not the specifics of the currently used text genre [20]. Content-specific Features Content-specific features generally match specific collections of keywords in documents. The patterns require prior knowledge about the language vocabulary that is usually computed using large text corpora. Standard features cover stopword statistics, frequency of abbreviations, slang word presence, and gender-based words [23]. Topical features, as shown in Figure 1.1 [21], can be used to determine the author's preferred topics if a large enough collection of the author's texts is presented. Idiosyncratic Features Idiosyncratic features in the stylometric analysis are typically related to phenomena that deviate from the correct language use. The most commonly used features include mispells [23] and typography rules violation. Similarly to syntactic features, robust grammar checkers should be utilized to reduce the features' noisiness. It is believed that each author's errors are consistent between different writings, and such mistakes are typically not intentional. That is why these features are robust to authors that intend to obfuscate their authentic style and are commonly used in authorship attribution tasks. 11 l . PROPAGANDA DETECTION APPROACHES Feature Selection Many stylometric tasks require large combinations of various feature types used together. However, combining certain features may lead to the increased complexity of text representation, which may harm the ability of the classifier to learn relevant concepts. Feature selection is an efficient technique used in text classification to reduce overall dimensionality, remove irrelevant features, and find the best combination of features to increase the classifier's performance. Commonly used methods for features selection use measures from information theory such as entropy [28], information gain, and ChiSquared [29]. However, more straightforward methods like term frequency are also used [30]. Another such method for feature selection uses a decision tree where the root node preserves all feature groups. Every branch splits the parent node into distinct subsets of features where a single feature group is missing. Finally, the decision tree is cropped to nodes that achieve the highest classification accuracy [31]. 1.2.3 Related Tasks The following subsection describes commonly used applications for stylometric text analysis. Authorship Attribution (AA) is a domain centered around determining the author's characteristics using the characteristics of the document [32]. The A A is divided into three specific problems. The first one, called Closed Set Author Attribution (CSAA), is a task of determining the document's author from the set of known authors. Authorship Verification, or Open Set Author Attribution is a generally more difficult version of CSAA. The goal is to verify whether the candidate is the document's author, where the candidate may not be present in the known author set. The last problem is author profiling, where the task is to determine the author's specific traits based on his writing. PAN is a series of scientific competitions centered around text forensics and stylometry. The events propose various shared tasks based on authorship analysis, computational ethics, and plagiarism detection [22]. 12 i . PROPAGANDA DETECTION APPROACHES Although stylometry is centered around meta-information not explicitly present in the provided texts, i.e., author's characteristics, it is also used to enhance the performance of other text classification tasks where writing style helps to discriminate between classes. Succesful applications include e-mail spam filtering [33], sentiment analysis [34], hate speech detection [35], and propaganda detection [15]. 13 2 The Propaganda Dataset Propaganda [36] corpus is a joint work between FSS M U , FI M U , a PrF M U under the project Manipulative techniques of propaganda in the age of Internet1 . In its current state, the dataset contains 8,644 documents that were extracted from four Czech news websites 2 . These websites were previously investigated for distributing Russian propaganda. The resulting corpus is tokenized using unit ok [37] and morphologically annotated using majka [38] and desamb [39] tools. 2.1 Annotation Description The dataset is annotated with a wide range of attributes. In this thesis, only 17 are considered (Politician 1-3, Video, and Image are left out), as they are not relevant for the stylometric text analysis. A l l 17 annotations are present at the level of the entire document, whereas 12 attributes also have annotations present at the span level. These spans label-specific sections of the text where the instance of the category is present. Figure 2.1 shows the dataset proportion for binary attributes. Frequencies for non-binary attributes can be seen via the link provided in the footnote. 2.1.1 Manipulative Techniques The following subsection describes eight manipulative techniques used in the dataset. A l l techniques are annotated at both the document and the span level. The techniques are then compared to either propaganda devices of thelnstitute of Propaganda Analysis (IPA) described in Table 1.1, or the attributes of the PTC dataset. Examples of manipulative techniques from the dataset are presented in Table 2.1. Argumentation (yes, no) carries content that may resemble either logical or emotional argument. The logic of the argument may be apparent but only sometimes complete. Uneducated readers, however, 1. https://www.muni.cz/en/research/projects/37145 2. parlamentnilisty.cz , svetkolemnas.info (currently down), sputnik.cz (currently down), and ac24. cz 14 2. T H E PROPAGANDA DATASET Figure 2.1: Relative frequency distributions for the attributes of Propaganda dataseta a. http://asteria04.fi.muni.cz:10008/graphs may consider this as a solid argumentation. Compared to the IPA propaganda devices, argumentation can contain a variety of logical fallacies such as The Band Wagon. Blaming (yes, no) intends to blame a subject for an event. The usual target of blaming is the opponent, but events can be blamed too. For example, Casual Oversimplification (assuming a single cause to the problem where there may be multiple) from the PTC dataset can be used to achieve blaming [14]. Fabulation (yes/no) consists of exaggerating problems, spreading false rumors, and presenting unsubstantiated claims. Overaxaggeration is a technique also used in Card Stacking of the IPA devices. Labelling (yes/no) contains short, impactful phrases that describe certain subjects or objects. It may be pejorative and based on stereotypes intended to decrease the subject's credibility, but positive labels are also present. Labeling is just a different name for Name Calling described in IPA's propaganda devices. Demonizing (yes/no) can be understood as emotionally heightened Labelling. The demonization victim is displayed as an evil entity that is the source of the described problems. 15 2. T H E PROPAGANDA DATASET Emotions is an annotation that describes specific emotions in the text. Emotive writing tends to distort the view of the reader, who is then less able to assess the objective nature of the topic rationally. Possible devices used include Loaded Language (using words and phrases with solid emotional implications) from the PTC dataset. The possible categories are fear, anger, indignation, compassion, other, or missing. Fear mongering (yes/no) is a technique that deliberately attempts to cause fear or distress in the reader. The writer usually reaches out into the future and presents some uncertainty or threat that affects the reader, where the opposing entity causes the threat. It is a similar attribute to Appeal to fear/prejudice in the PTC dataset. Relativizing (yes/no) is used to relativize the negative aspects of the writer's beliefs or the positive deeds of the opposing entity. Commonly used techniques to achieve this are Exaggeration (presenting facts in an excessive manner) and Minimization (making facts seem less important) present in the PTC dataset. 2.1.2 Document-Level Attributes The following attributes are annotated only on the document level, describing the document's properties and structure. Genre classifies the document into one of three categories (news, interview, comment). Topic orders the documents via its subject. Possible categories are commonly used topics in manipulative news outlets. There are 13 categories overall: notable ones include domestic/foreign policy, the migration crisis, and the conflict in Ukraine. Location (8 classes) describes where the article is set in. The first six classes refer to the locations that are commonly discussed in manipulative articles (NATO, Russia, USA, Russia and USA, EU). The last two classes are reserved for cases where other locations are discussed or the location cannot be determined. Scope describes if the scope of the article is within the domestic, foreign events or both (last class unknown is reserved for cases where it is impossible to assess). Overall Sentiment summarizes emotions contained in the article into positive, negative, and neutral. Fear Mongering and Emotions may 16 2. T H E PROPAGANDA DATASET Table 2.1: Example annotations for attributes at the span level. Range annotations are higlighted in bold Technique Example Argumentation CS: žena byla příliš vylekaná, a proto se neobrátila na policii EN: the woman was too scared, and that's why she didn't contact the police Blaming CS: Podle jejich názoru je tento časopis pornografický a ponižuje ženy. EN: In their opinion, this magazine is pornographic and degrades women. Demonization CS: ...na západě je všechno hrozné a že zemřeme, když tam pojedeme... EN: ...everything is terrible in the west and we will die if we go there... Emotions (anger) CS: Jaká xenofóbie? Kdyby se nechovali jak kreténi, nikdo si jich nevšimne EN: What xenophobia? If they didn't act like morons, no one would notice Fabulation CS: byla sesazena z vlády poté, co přijala rozkaz Bruselu EN: she was ousted from the government after accepting orders from Brussels Fear Mongering CS: Severokorejská hrozba klepe na dveře střední Evropy. EN: The North Korean threat is knocking on the doors of central Europe. Labeling CS: Putin potvrdil novou zbraň: nepřemožitelná jaderná hlavice EN: Putin has confirmed a new weapon: an unstoppable nuclear warhead Relativizing CS: Už byl do toho hrozně emočně zatažený, že už to psychicky neustál EN: he was so emotionally overwhelmed that he couldn't take it anymore Expert CS: Sociolog Jan Hartl upozorňuje, že nyní Zeman... EN: Sociologist Jan Hartl points out that now Zeman... Location CS: ...se v první části sjezdu v Hradci Králové ČSSD rozhodla... EN: ...ČSSD has decided in the first part of the convention in Hradec Králové... Russia (victim) CS: Ať se stane cokoliv, vždy „provokuje" Rusko. EN: Whatever happens, Russia is the one "provoking" here. Opinion CS: Je přijatelné, že se takto prezident k veřejnoprávnímu médiu vyjadřuje? EN: Is it acceptable for the president to speak about the public media like this? Source CS: Britské BBC přinesl informace o řádění fotbalových fanoušků... EN: The British BBC brought information about the rampage of soccer fans... contribute to the overall sentiment, but they are individual manipulative techniques within specific sections of the article. 2.1.3 Other Atrributes The following attributes are on the span level of the document; however, they are not considered manipulative. Expert (yes/no) is positively labeled in articles where the claims presented are supported by an expert or anyone other influential from the third party. Although it is not considered a manipulative technique in this dataset, it can be used to reinforce the author's point of view by introducing claims backed up by an expert. That is why this category shares some similarities with the Testimonial device of IPA. Russia category specifies how Russia is depicted in the article. This annotation can help to determine the target audience of the article. 17 2. T H E PROPAGANDA DATASET The possible classes of this category are hero, positive/negative example, victim, or neutral. Source (yes/no) indicates the presence of references to other sources, possibly other web pages, media, or documents. Similarly to Expert, sources can be used to raise the seeming credibility of the content, especially when facts from a well-known mainstream media are referenced. In these cases, the category is similar to the Transfer device on Table 1.1. Opinion (yes/no) indicates the presence of the author expressing his personal opinion on the discussed topic. This attribute can also be a comment or suggestive question in an interview. 18 3 Stylometric Feature Extraction The following chapter presents a processing pipeline for stylometric text analysis. The first section describes the preparation of the document for feature extraction. The remainder of the chapter is dedicated to describing stylometric text features used to detect manipulative style. 3.1 Document Preprocessing The input document is presented in plain text format. Its preprocessing is split into two separate stages. In the first stage, irrelevant text content is removed to avoid creating noisy features. Removing the remaining H T M L tags is essential as the documents are sourced from online news outlets. In the second stage, unitok, majka, and desamb tools are used to annotate the input document morphologically. The resulting tokens, lemmas, and morphological tags are provided alongside the presented plaintext document for the incoming stylometric analysis. The resulting format is presented as a sequence of pairs (token, lemma) for lemmas and (token, tag) for morphology, as can be seen in Figure 3.1. {"text": "Rusko varovalo USA před využitím obvinění p r o t i . . . " , "lemmas": [("Rusko", "Rusko"), ("varovalo", "varovat"), . . . ] , "tags": [("Rusko", " k l g N S c l " ) , ("varovalo", "k5eAaImAgNnS"), ...] } Figure 3.1: A n example of prepared document. Sentence boundaries labeled as are omitted for clarity. 3.2 Proposed Features The following section presents a detailed overview of stylometric features implemented as part of this thesis. It is a set of around 20,000 numerical features originally proposed by Rygl et al. [40]. A n overview of various feature types can be seen in Table 3.1. A non-exhaustive list of modifications performed to the original features is present below: 19 3. STYLOMETRIC FEATURE EXTRACTION 1. morphological features were adapted to the desamb tagset 2. fine-tuning of required relative frequencies to generate a more compact set of features 3. adjustment of n for n-gram features to support wider range of n-grams 4. automated tests to ensure the correctness of the feature extrac- tion Table 3.1: Overview of proposed stylometric text features Feature Type Feature Subtype # features Language Independent naive 30 / Word Length improved 77 / n-grams 30 / naive 25 / Sentence Length improved 127 / n-gram 25 / avg. repetition per sent. 1 / avg. repetition per doc. 1 / Word Repetition word class repetition prob. word class repetition 13 13 word repetition distance 12 / bag of words repetition 100 / Word Class N-Grams 1 to 4-grams 514 Morphological Tags N-Grams full simplified tags 10,000 200 Letter Casing 1 to 3-grams indexed 1 to 3-grams 77 417 / / Word Suffixes stemmed 100 / Word Suffixes parametrized n-grams 325 / Word Richness richness metrics 6 / Stop words for lemmas 300 / Stop words for tokens 300 / frequency 11 / Punctuation position frequency 60 / n-gram frequency 76 / Typography fixed rules 11 / Typography dynamic 100 / Character N-Gram Distribution 1 to 5-grams 6,550 / Emoticons Presence n-grams 28 / Total 19,529 20 3. STYLOMETRIC FEATURE EXTRACTION 3.2.1 Word and Sentence Length Distribution Analysis Word Length Analysis is a lexical feature that measures the distribution of word lengths in letters. It is measured at both the token and lemma levels. Generally, tokens that contain numbers or nonalphanumerical characters are omitted. Sentence Length Analysis is computed on the syntactic level, and it is a distribution of sentence lengths w.r.t. the number of tokens. Both features are implemented using three distinct methods: naive, improved, and n-gram. The naive version creates a n-dimensional vector where z-th position (1 < i < n) represents the relative frequency of words/sentences with length i. The last dimension is reserved for words/sentences that are of length n and longer. The value of n is set to 15 for words and 25 for sentences. Improved analysis creates variable length vectors derived from the referential corpus. The content of the vector is the following: • n + 1 dimensions representing the relative frequency distribution, similarly to the naive approach - (n + l)th element is reserved for word/sentences longer than n • n — 1 elements as a relative frequency histogram with a bin size of 2 • n — 2 elements as a relative frequency histogram with the bin size of 3 The n is determined by finding the maximum length in the referential corpus with a relative frequency of at least 0.01%. The N-Gram approach extracts relative frequencies of sums of 3-grams present in the text. The construction of such features remains similar to the naive approach, where z-th element (1 < i < n) represents sum 3-gram lengths in the range (3z, 3z + 2), and the n-th element is reserved for longer 3-grams. 21 3- STYLOMETRIC FEATURE EXTRACTION 3.2.2 Morphological Features The following set of syntactic features make heavy use of morphological tags extracted during the preprocessing stage. For this reason, the following features are language dependent, and their implementation can vary depending on the currently used tagset. Word Class N-Gram only extracts relative frequencies of part of speech tags. The N parameter ranges between 1 and 4, and only the n-grams with a relative frequency above 0.1% are considered. Morphological Tags N-gram frequencies consider more information than just part of speech. The full version uses the entire morphological tag. Due to the morphological richness of the Czech language, the limit of featured n-grams is set as 100 * n 3 , where n stands for n-gram length. The value of n is set from 1 to 4. Simplified version omits the infrequent parts of the tag. In the first step, the morphological tag is split into chunks, each representing a separate piece of information. A n example chunk split of the currently used tagset can be seen in Figure 3.2 1 . Unikátní Iverdikt schválen: Čínští poslanci. klglnSd Part of speech: noun Gender: masculine inanimate Number: singular Case: nominative Figure 3.2: A n example of Majka morphological tag chunks. The chunks highlighted in green were included in the final features, while the red ones were discarded. Out of all the possible tag chunks, only the ones with their relative frequency in the reference corpora larger than the median are selected. The final, simplified tags are composed of the selected tag chunks. 1. Full tagset reference: https: //nip. f i .muni. cz/projekty/ajka/tags .pdf 22 3. STYLOMETRIC FEATURE EXTRACTION The final features are filtered using preferred relative frequency and a hard limit on total number of features 3.2.3 Word Repetition Word Repetition metrics are typically lexical features analyzed using multiple methods. Average repetition per document features compute frequency histogram for each unique token/lemma in the document, where the mean relative frequency is the sole, resulting feature. Average repetition per sentence computes the frequency histogram for each word repeated at least once in the sentence (for example, three repetitions of the same word count as two in the histogram). The histograms are then summed up, normalized, and the mean frequency is returned. Bag of Words repetition turns documents into a normalized bag of lemmas representation where stopwords and words with too low relative frequency (0.05%) are omitted. The referential corpus determines the stopwords by their relative frequencies, where 4% is the minimum threshold. Out of the resulting candidates, the top 100 words are selected as the resulting features. Word Class Repetition is a normalized word class histogram where each token's corresponding word class count is incremented for each sentence where the token is repeated. Probabilistic Word Class Repetition computes the theoretical probability of word class being repeated from the referential corpus and returns the difference between the referential probabilities and the probabilities from the provided document. 3.2.4 Word Richness Metrics Word Richness considers two methods of computing vocabulary richness as lexical features. Inverse Simpson's Diversity Index [41] is computed on all alphanumerical and alpha tokens. 1 1 1 + A " l + X £ i p ? Where R stands for richness (the number of unique words in the presented document, and pi represents the relative frequency of word 23 3. STYLOMETRIC FEATURE EXTRACTION i, also known as the proportional abundance of word i. The one added in the denominator is not present in the original literature and is added to enforce the possible values to be < 1 for better properties in machine learning algorithms. Yule's K is a measure based on text repetitiveness [42], and is computed as follows: 1 1 l + K " 1 + C [ - i + E S V(m,N)($)2] Where N is the total number of words in the document, mmax is the maximal frequency of any word in the document, and V(m,N) is the number of words appearing m times in the text. C is a numerical constant set to 10,100,1,000, and 10,000 to extract four features. 3.2.5 Punctuation and Typography Punctuation frequency is a syntactic feature that examines the presence of various punctuation marks in the document. The list of considered punctuation marks is .,:, ,, ;, ?, !, -, (,),/, and The basic version computes the relative frequency of punctuation marks in the input document. Punctuation N-grams computes at most 100 most frequent n-grams with relative frequency above 0.2%. The position frequency version also considers the placement of punctuation marks. Four different positions are examined: • start - first two tokens in the sentence, or the first three if the sentence is longer than eight tokens • end - two last tokens in the sentence, or the last three if the sentence is longer than eight tokens • middle - if neither of two conditions above hold • anywhere - if any of the conditions above hold Fixed Typography is an idiosyncratic feature that checks for typography rules violations and various patterns related to typography that are checked using 11 regular expressions. The resulting feature vector is the proportion of typography violations w.r.t. to all possible 24 3. STYLOMETRIC FEATURE EXTRACTION positions where the rule can be violated. A n example can be seen in Figure 3.3. Dynamic Typography observes the n-gram frequencies of nonalphanumeric tokens. Minimum and maximum n-gram lengths are set to 2 and 8, respectively, where the 100 most frequent n-grams are chosen from the referential texts. |THE END OF THE AMERICAN DOLLARfrchina has ^just released petro-juan,so it can face the superiority of US dollar. 0.11 0.06 0 0.26 0 0 0 0.66 0.33 0 0 f / / to Figure 3.3: Fixed typography feature example 3.2.6 Other Features Character N-gram frequencies is a lexical feature extracted for at most 1,000 unique items with preferred document frequency around 50%. Emoticon N-gram Presence checks for the presence of pre-defined emojis in the presented document. The emojis are grouped into three categories according to their sentiment - posit ive, neutral, and negat ive. Possible n-grams are determined from the referential corpus. Letter Casing features are composed of two different methods. The first method computes n-grams of capital letters according to their position in the word and sentence using fixed rules. The indexed version also considers the exact position of the token in the sentence. The parametrized version of word suffixes computes relative frequencies of the last n characters of each token document-wide. The stemmed version attempts to guess the word suffix based on the provided word and its corresponding lemma. 25 3. STYLOMETRIC FEATURE EXTRACTION Stopword presence is a content-specific feature that attempts to learn stopwords from the referential texts by extracting the most common words with the minimal relative frequency of 5%. The extraction works in three different modes. Presence extracts a one-hot vector indicating the presence of individual stopwords in the presented document. Frequency returns a vector of relative frequencies of stopwords. Deviation computes vector of standard deviations w.r.t. to other stop- words. 26 4 Manipulative Style Detection The following chapter describes the machine learning methods used in this thesis. Stylometric features are employed in various classical machine learning methods and deep neural transformer models with modified classification heads that can accept extra features. At the end of the chapter, a feature selection approach is proposed to derive the most prominent feature set for the given attribute. 4.1 Classical Machine Learning Methods Various classical machine learning methods are used to compare the viability of the proposed stylometric features with more mainstream text representations. Due to well-established and fast implementations, these methods serve as a solid baseline for further experimentation. This thesis uses standard s c i k i t - l e a r n [43] implementations with their default parameters. TF-IDF normalized bag of words is used to describe the non-secondary features of the text. Multinomial Naive Bayes is a Naive Bayes variant derived from Bayes's Theorem with the assumption of conditional independence of every featured pair given the target class. The used version assumes multinomial distribution of the used data. It is used in this thesis as it is considered a solid baseline for text classification. Unfortunately, it cannot be used for the stylometric features as it has negative values incompatible with the distribution assumption. Random Forests is an ensemble of decision trees used for classification and regression tasks. Each tree in the ensemble draws a random sample from the training data (with replacement). The resulting prediction is computed by majority voting of the individual trees. For reproduction purposes, the random seed of this classifier is set to 42. It is an ideal candidate due to the dataset bootstrapping process, which helps with learning from imbalanced domains. Support Vector Machines use hyperplanes to separate instances of different classes. The hyperplane is chosen to maximize the distance between the closest instances of different classes - also known as support vectors. It is a popular machine learning method for classification due to straightforward interpretability and reasonable accuracy, and 27 4. MANIPULATIVE STYLE DETECTION for this reason, it is used in this thesis. The LI penalty is used to train the model, as it allows for the sparsity of the model (feature selection). Perceptron is a feedforward neural network without any hidden layers. It is a linear model where each neuron is interpreted as a separating hyperplane of the feature space. It is trained using the Stochastic Gradient Descend algorithm. The introduction of perceptrons in this thesis is due to the similarity of the representation and learning algorithm with other, more complex deep learning methods. 4.1.1 Data Augmentation for Imbalanced Domains Due to the class imbalance in the Propaganda dataset, many learning algorithms may create a bias towards the majority classes, potentially ignoring the minority classes completely. For this reason, multiple techniques can cope with the class imbalance. Majority class undersampling randomly discards data points from the majority classes until the desired balance is achieved. Unfortunately, this means discarding valuable data, which is problematic for small datasets (like the Propaganda dataset). Minority class oversampling creates copies of random data points in the minority classes until the desired proportion is reached. It is more widely used in smaller datasets; however, in cases of extreme imbalance (for example, the Demonization class in the Propaganda dataset), the classifier may overfit on the majority class due to the lack of diverse examples. Synthetic Minority Oversampling Technique (SMOTE) [44] is a method for creating artificial examples for the minority classes. The synthesis process is performed as follows: 1. select a random data point x from the minority class 2. compute k nearest neighbors of x belonging to the same class (typically, k — 5) 3. select a random nearest neighbor y 4. create a new example that is located randomly along the line between x and y 28 4. MANIPULATIVE STYLE DETECTION Output Class Scores [ Feedforward Layer | Dropout Layer ] Feedforward Layer X Dropout Layer I Feature Vector (19,529 features) Stylometric Feature Extraction Lemmas Hidden Representation Final Representation [CLS] Token Vector (2,000 features) (concatenated) (768 features) Deep Transformer Model Tags Tokenization and Morphological Annotation Document Text Tokens Tokenizer Figure 4.1: Deep neural architecture for manipulative style detection 4.2 Deep learning Methods The following section proposes an approach for manipulative style recognition using both stylometric text analysis and deep neural transformer models. A visual schema of this architecture is depicted in Figure 4.1. 4.2.1 Architecture Description The input document is tokenized using unit ok and morphologically annotated using maj ka and desamb tools. The resulting tokens, lemmas, and morphological tags are used to extract a feature vector using the stylometric analysis described in Section 3. The representation is then passed to a single feedforward layer. The resulting, more dense representation highlights the essential features for classification and represents the writing style used in the input document. 29 4. MANIPULATIVE STYLE DETECTION In tandem with the previous paragraph, the input text document is processed using deep transformer models [10]. Considered models are Czert [45] and Slavic Bert [46] based on BERT, Robezcech [47] based on RoBERTa , and X L M Roberta Large [48] pretrained on 100 languages including Czech. The priorities behind the choice of models were the ability to process the Czech language, public availability, and solid reported performance on classification tasks. The models were pretrained on sequences with a maximum token length of 512, so parts of the input document that exceed this limit are removed. The CLS token vector is extracted from the first item in the resulting sequence and it is then concatenated to the hidden stylometric representation. In the final step, the concatenated representation is passed through the final feedforward layer, which models the predictions for each class in the attribute. The final prediction is selected using the argmax function. To compare this approach, another architecture is proposed that does not use the stylometric features. The method uses a classification head composed of a single feedforward layer that maps the [CLS] token into the output classes. 4.2.2 Training Description The proposed model is fine-tuned using the HuggingFace framework on 20 epochs. Hyperparameter values for the training were empirically determined. We use the learning rate of 3 x 1 0 - 6 and the linear warmup ratio of 0.1 for the AdamW optimizer. The cross-entropy loss is used as the objective to minimize. A n early stopping mechanism is used to prevent further overfitting. Every 500 optimization steps (also called checkpoint), the current model is evaluated on the validation set. If the evaluation metric of the validation set does not improve three times in a row, the training process is terminated prematurely and the best checkpoint is used for the final evaluation. Due to the lack of training examples and label unbalance in some classes, more aggressive values for generalization were chosen. We use the dropout probability p = 0.5 for each presented feedforward layer, and a weight decay value of 0.01. The training computations were accelerated using G P U with a batch size of 64 and gradient 30 4. MANIPULATIVE STYLE DETECTION Table 4.1: Gradient accumulation steps (GA) and batch size ( B S ) values for different pretrained models and graphics cards Pretrained Model Tesla T4 A40 Ainn Pretrained Model G A B S G A B S G A B S Czert 2 32 1 64 1 64 Slavic Bert 2 32 1 64 1 64 Robeczech 4 16 2 32 1 64 X L M Roberta Large 32 2 4 16 2 32 accumulation to fit the G P U memory adequately. A n exact setup of gradient accumalation and batch size can be observed in Table 4.1. A performance benchmark for various graphics cards is presented in Figure 4.2. The batch sizes and gradient accumulation steps correspond to the ones presented in Table 4.1 to efficiently utilize the computing capabilities of the graphics cards. The addition of stylometric features slightly increases the required training times. However, when taking into the account the complexity of the deep transformer models, the addition of other manually engineered features does not change the required training times dramatically. Tesla T4 XLM Roberta Large Slavic Bert Robeczech Czert XLM Roberta Large Slavic Bert Robeczech Czert XLM Roberta Large Slavic Bert Robeczech Czert 200 300 A40 200 300 A100 200 300 Running time (m) Text only Text + stylometry 400 500 Figure 4.2: Performance benchmark of the deep learning models for various GPUs 31 4. MANIPULATIVE STYLE DETECTION 4.3 Feature Selection for Deep Neural Approaches With the proposed stylometric text features, the task is to discard the noisy subsets of features that hurt the system's overall performance. The process has to be applied separately for each attribute of the Propaganda dataset. A common approach is to retrain models with different subsets of features and select the one that produces the best accuracy metric. Unfortunately, this approach is not practical from the time perspective due to the training complexity of deep transformer models. The proposed approach organizes all features into a tree structure, where the leaf nodes represent a small subset of semantically similar features. Each non-leaf node unifies feature subsets presented in child nodes. The root node represents the original feature set. The feature tree is presented in Figure 4.3. SURFACE FEATURES WORD/SENT. LENGTHS ALL FEATURES MORPHOLOGY AND CORPORA STATISTICS SURFACE WORD CHARACTERISTICS NON-ALPHANUMERIC STATISTICS I — STOPWORDS MORPHOLOGY REPETITION improved word suffixes char, n-grams capital letters punctuation typography presence frequency std. deviation P O S n-grams full morphology simplified morpho. word repetition P O S repetition Figure 4.3: Feature categories organized into a tree structure In order to select the most promising feature set for the large neural models, Random Forest classifiers are trained on different subsets of features. No data augmentation is used through this process. The tree is traversed in a f reorder fashion, and for each traversed node, a 32 4. MANIPULATIVE STYLE DETECTION new classifier is trained and evaluated with all features present in the subtree discarded. The most insignificant feature is removed from the training of the deep learning models as an alternative approach to the complete feature set. The highest performance increase when specific feature subsets are absent implies the irrelevance of the mentioned features. Also, the highest performance decrease helps to pinpoint the essential feature sets in the classification of a particular attribute, adding some explainability to the improvement of the model. 33 5 Evaluation The final chapter presents results on all of this thesis's suggested approaches and features. After describing the performed experiments, a detailed description and interpretation of the results is presented for each attribute. Then, the accomplished results are compared with the other results measured on the Propaganda dataset. Finally, an error analysis and a preview of possible improvements for the described methods are presented. 5.1 Experiment Setup Each technique described in Chapter 3 is evaluated on all attributes of the Propaganda dataset. Overall, 918 models were computed when considering classical machine learning and deep learning methods. For the classical machine learning methods, the data is augmented using SMOTE. If the method is parallelizable, it is run using 40 CPUs. The non-stylometric approach uses the TF-IDF representation of the dataset, while the stylometric approaches use the complete feature set (except for the Random Forests). No data augmentation is used for the deep learning approaches. Using SMOTE, the stylometric features are augmented from multiple nearest neighbors, so the synthetic features do not correspond with the original texts. Unfortunately, the proposed approach uses pairs (text, stylometric_f eatures) as an input, which is why such augmentation cannot be applied here. The non-stylometric approach accepts the plain text representation that is processed using the corresponding tokenizer. In contrast, the stylometric approaches use text and stylometric features, where the complete feature set is used by default. If a feature subset exists, the feature selection process declares that as the least significant, an alternative training (from now on referred to as reduced stylometry) discards the feature subset. The used evaluation metric is a weighted F l measure, where the classic F l measure is a harmonic mean of precision and recall in binary 34 5- EVALUATION classification and is expressed as follows: F l presicion * recall precision + recall To adapt this measure for multi-class classification and to factor in for label imbalance, the weighted F l measure is expressed in the following way: Where the C stands for the number of classes, n is the number of evaluated examples, supp(i) represents the number of true instances for the class i, and Fl,- is the Fl-measure computed separately for class i. Each experiment is repeated three times, and the resulting metrics are averaged. Finally, to compare this thesis's results with recent but yet unpublished ones, the approach has to be altered for non-binary attributes. For each attribute class, a separate binary classifier has to be trained with a single output neuron with a sigmoid function applied to represent logistic regression. The loss function is replaced with mean squared error. A standard cutoff value of 0.5 is used to determine the correctness of the classification. The weighted F l score is computed for each classifier separately, and the final score is the average of all classifiers. 5.2 Feature Importance Table 5.1 shows results for the feature selection process described in Section 4.3. Feature importances might differ between different proposed approaches, so the feature set choice may not always be optimal for other methods. The Random Forests were chosen to compromise fast training times and reasonable accuracy. Considerable importance is measured for character n-grams in Russia and Location attributes. It is one of the few features that can also describe the document's semantics, which is one of the core assumptions for successfully classifying these attributes. The same assumption 35 5- EVALUATION applies to Topic, with 13.39% of drop in weighted F l when omitting surface features. The most problematic feature set proves to be subsets that include morphological information. In the most extreme cases, removal of full morphological tags n-grams increases the performance by 2.63% of weighted F l . The main problem might be the morphological richness of the Czech language. When not bound by occurrence frequencies, millions of distinct morphological tags are present in the referential corpus, so it is difficult to adequately capture the ones that are significant for classification. Simplified morphology on the other hand, seems to be less noisy - showing that retaining only the frequent parts of morphological tags may be more helpful. Table 5.1: Results for measurement of feature importances via Random Forests. AMIFS and ALIFS refer to difference in weighted F l score between the setup with the omitted feature set and the setup with all features Attribute Most Important FS AMIFS Least Important FS ALIFS Argumentation capital letters 0.31 mor. and cor. stats. -1.82 Blaming surface features 0.91 morphology -0.53 Demonization all features 0 all features 0 Emotions surface features 0.01 full morpho. tags -0.08 Fabulation surface word char. 0.49 word/sent, lengths -0.58 Fear Mongering typography 0.02 non-alnum char. -0.16 Labelling surface word char. 0.88 full morpho. tags -0.51 Relativizing mor. and cor. stats. 0.9 non-alnum char. -0.04 Genre surface features 1.22 mor. and cor. stats. -0.25 Location char, n-grams 12.34 full morpho. tags -2.63 Scope surface word char. 4.67 mor. and cor. stats -1.26 Overall Sentiment surface word char. 0.82 morphology -0.67 Topic surface features 13.39 morphology -2.33 Expert surface features 2.08 punctuation -0.65 Source indexed cap. letters 0.76 non-alnum char. -1.36 Russia char, n-grams 7.21 typography -0.1 5.3 Results The following section presents the results for all attributes of the Propaganda dataset. For a better overview, the presented results are split 36 5- EVALUATION into three categories according to the attribute type. A n emphasis is given to the comparison of stylometric and non-stylometric techniques. Appendix A presents the complete results table, including the ones missed in this chapter. 5.3.1 Manipulative Techniques A n overview of results for manipulative techniques can be seen in Table 5.2 and in Figure 5.1. In the vast majority of cases, the usage of stylometric text techniques helped to improve the overall performance of the proposed methods. • U s e s s t y l o m e t r y MWM N a i v e B a y e s S u p p o r t V e c t o r M a c h i n e MMM C z e r t MMM R o b e c z e c h • U s e s r e d u c e d s t y l o m e t r y P e r c e p t ron MMM R a n d o m F o r e s t MUM S l a v i c Bert MUM X L M R o b e r t a L a r g e * U s e s t e x t o n l y • # • •• * * • • 6 7 . 0 6 7 . 5 68.0 6 8 . 5 69.0 6 9 . 5 70.0 70.5 • • • • «• • • • 71.0 71.5 7 2 . 0 72.5 7 3 . 0 73.5 74.0 7 4 . 5 • • • # .ft • 95.7 95.8 9 5 . 9 96.0 96.1 • • • • • ft •» * 78 79 8 0 81 82 • • mm • • •• • • • •* 75 76 77 7 8 79 8 0 • • • • • • • • • • * 89.0 89.5 9 0 . 0 9 0 . 5 91.0 91.5 • # • • • •• • • * • 8 1 . 0 81.5 8 2 . 0 82.5 8 3 . 0 83.5 • • ti|ft**^fr m 91.5 9 2 . 0 W e i g h t e d F l Figure 5.1: Visual overview of results for manipulative techniques. Experiments with results that are 5% below the median are omitted from the visualization to highlight more promising approaches. The most notable difference can be observed with the Emotions attribute, where an improvement of 0.82% of weighted F l was reached. Writing that leverages specific emotions is characterized by an easily recognizable writing style that was properly captured in most experi- ments. 37 5- EVALUATION Table 5.2: A n overview of the best-performing methods for manipulative techniques. The first row refers to the best-performing technique without secondary features for each attribute, whereas the second row uses stylometric features, r. style refers to stylometry with features discarded using feature selection. Attribute Method Type Weighted F l Argumentation X L M R Large X L M R Large text text + style 70.69 70.64 Blaming X L M R Large X L M R Large text text + r. style 74.55 74.92 Demonization X L M R Large X L M R Large text text + style 96.13 96.19 Emotions X L M R Large X L M R Large text text + r. style 81.81 82.63 Fabulation X L M R Large X L M R Large text text + r. style 80.57 80.92 Fear Mongering X L M R Large X L M R Large text text + r. style 91.71 91.94 Labelling Perceptron Random Forest text r. style 83.59 83.22 Relativizing Random Forest Random Forest text r. style 93.01 93.31 The average weighted F l score of the style-based approach is lower than the text-only classification in cases of Argumentation and Labelling. Argumentation is considered a complex and noisy attribute with a relatively low inter-annotator agreement. The current definition of Argumentation allows for anything from simple reasoning to a solid argument to be included in this class. The argument may also contain various logical fallacies. Due to such high variation, the Argumentation may be challenging to grasp using automated machine learning methods. The difference between approaches is 0.05% in favor of the non-stylometric one, which is considered a non-significant difference. The Labelling manipulative technique usually refers to a small segment of the text containing a short, powerful phrase. Stylometric features, on the other hand, summarize the writing style of the entire 38 5- EVALUATION document. The reason behind the 0.36% decrease in the performance metric may be that the stylometric features could not correctly capture this attribute's characteristics. A possible solution could incorporate stylometry on the span of individual paragraphs instead of the entire document. The prominence of perceptron in the Labelling attribute could indicate multiple things. First, insufficient data is available to fine-tune deep transformer models properly. Also, when looking at Figure 5.2, most of the Labelling annotations consist of less than four words. These words usually stand out from the rest of the document as they are not a part of usual writing (words like presstitute, Haveljugend). For each class, the perceptron computes the final score as a linear combination between the TF-IDF representation and their associated weights. For the learning of perceptron, it is sufficient to assign high weights for the words that are typically used as labels, which might explain the high performance of this method. Range Length (#tokens) Figure 5.2: Range length distribution in tokens for the attribute La- belling 39 5- EVALUATION 5.3.2 Document-Level Properties Table 5.3 and Figure 5.3 summarize the results for document-level properties. Overall, the stylometric techniques have outperformed the other methods in three of five attributes. | U s e s s t y l o m e t r y N a i v e B a y e s S u p p o r t V e c t o r M a c h i n e C z e r t R o b e c z e c h ^ U s e s r e d u c e d s t y l o m e t r y P e r c e p t r o n R a n d o m F o r e s t S l a v i c Bert X L M R o b e r t a L a r g e • U s e s t e x t o n l y 92.5 93.0 9 3 . 5 94.0 94.5 9 5 . 0 95.5 9 6 . 0 96.5 / • • • • • * 80.5 81.0 81.5 82.0 8 2 . 5 8 3 . 0 83.5 J- • • • • • • • • * 68 69 70 71 a. 0.70 / 0.3 0.68 4 6 8 10 12 14 16 4 6 8 10 12 14 16 Epoch Epoch Figure 5.5: A visualization of training and validation metrics for the stylometric Slavic Bert experiment on the attribute Blaming In order to improve the current deep learning approaches, a data augmentation technique has to be utilized that affects both the stylometric features and the input text. The most obvious approach would be to generate various paraphrases of the input documents, either by altering the entire document or by modifying the fragments of text containing the desired attribute. According to the previous results, in some attributes, the deep transformer models were outperformed by BiGRUs. A promising approach that was not covered in this thesis could combine the word2vec representation of the document and the stylometric features in a similar way that was used in the transformer models. Various recurrent neural network architectures could be used to process the word2vec sequence, and the resulting hidden state could be combined with the stylometric feature vector to generate the final prediction. Besides the training and augmentation, several steps can be taken to improve the stylometric features themselves. The most obvious improvements are to revise the existing problematic features (for example,full morphology) and add new ones into the feature set. Promising candidates for feature extensions would incorporate psycholinguistic wordlists that map specific words of the document into the psychological and emotional background of the writer. Other expansions could utilize specific corpus statistics like hapax legomena frequencies or 46 5. EVALUATION other lexical summary measures besides the word richness, for example, readability index. 47 6 Conclusion and Future Work This thesis designs, implements, and evaluates various manipulative style recognition techniques for propaganda detection in the Czech language. The foundation of the solution is based on stylometric text analysis. A l l proposed solutions were benchmarked using the Propaganda dataset and compared with their counterparts that use text representations other than stylometry. The first chapter presents the current trends in the detection of manipulative techniques. The second half of the chapter has described well-established methods in stylometric text analysis on which this thesis is based. The second chapter describes the Propaganda dataset and its attributes with their comparison to other datasets. The third and fourth chapters describe the solutions used in this thesis. Among these are stylometric features and machine learning algorithms that generate the final predictions. The last chapter performs a detailed evaluation and comparison of the proposed approaches. Overall, 918 models were trained for both style-based and other text representations. The stylometric techniques outperform the nonstylometric ones in most attributes. The most prominent weighted F l increase was measured in the Source attribute by 3.4%. It was a common theme that the utilization of X L M Roberta Large yielded the best result, as it is the most complex model used in this thesis. However, in some attributes like Relativizing, classical machine learning methods like Random Forest outperformed the deep learning approaches. Compared to the previous results, the style-based methods outperform the previous approaches in almost every dataset attribute. In the last segments of this thesis, several problems and their possible solutions were pinpointed as a point of reference for future work. The most promising future directions include using data augmentation to combat the imbalance of the domains and careful modifications and expansions to the available stylometric feature set. 48 Bibliography 1. BOLSOVER, Gillian; HOWARD, Philip. Computational propaganda and political big data: Moving toward a more critical research agenda. Vol. 5. Mary A n n Liebert, Inc. 140 Huguenot Street, 3rd Floor New Rochelle, N Y 10801 USA, 2017. No. 4. 2. MARTINO, Giovanni Da San; CRESCI, Stefano; BARRÓN-CEDEŇO, Alberto; YU, Seunghak; DI PIETRO, Roberto; NAKOV, Preslav. A survey on computational propaganda detection. arXiv preprint arXiv:2007.08024. 2020. 3. HORAK, Ales; BAISA, Vít; HEŘMAN, Ondřej. Technological Approaches to Detecting Online Disinformation and Manipulation. ArXiv. 2021, vol. abs/2108.11669. 4. ESTEVES, Diego; REDDY, Aniketh Janardhan; CHAWLA, Piyush; L E H M A N N , Jens. Belittling the Source: Trustworthiness Indicators to Obfuscate Fake News on the Web. 2018. Available from DOI: 10.48550/ARXIV. 1809.00494. 5. ABBASI, Mohammad-Ali; LIU, Huan. Measuring User Credibility in Social Media. In: GREENBERG, Ariel M . ; KENNEDY, William G.; BOS, Nathan D. (eds.). Social Computing, BehavioralCultural Modeling and Prediction. Berlin, Heidelberg: Springer Berlin Heidelberg, 2013, pp. 441-448. ISBN 978-3-642-37210-0. 6. VOSOUGHI, Soroush; ROY, Deb; A R A L , Sinan. The spread of true and false news online, science. 2018, vol. 359, no. 6380, pp. 1146- 1151. 7. SIERING, Michael; KOCH, Jascha-Alexander; DEOKAR, Amit V. Detecting fraudulent behavior on crowdfunding platforms: The role of linguistic and content-based cues in static and dynamic contexts. Journal of Management Information Systems. 2016, vol. 33, no. 2, pp. 421-455. 8. HOCHREITER, Sepp; SCHMIDHUBER, Jürgen. Long short-term memory. Neural computation. 1997, vol. 9, no. 8, pp. 1735-1780. 49 BIBLIOGRAPHY 9. CHO, Kyunghyun; MERRIENBOER, Bart van; BAHDANAU, Dzmitry; BENGIO, Yoshua. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches. arXiv, 2014. Available from DOI: 10.48550/ARXIV.1409.1259. 10. VASWANI, Ashish; SHAZEER, Noam; PARMAR, Niki; USZKOREIT, Jakob; JONES, Llion; GOMEZ, Aidan N ; KAISER, Lukasz; POLOSUKHIN, Illia. Attention Is All You Need. arXiv, 2017. Available from DOI: 10.48550/ARXIV. 1706.03762. 11. RASHKIN, Hannah; CHOI, Eunsol; JANG, Jin Yea; VOLKOVA, Svitlana; CHOI, Yejin. 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. Copenhagen, Denmark: Association for Computational Linguistics, 2017, pp. 2931-2937. Available from DOI: 10.18653/ vl/D17-1317. 12. MINTZ, Mike; BILLS, Steven; SNOW, Rion; JURAFSKY, Dan. Distant supervision for relation extraction without labeled data. In: Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP. 2009, pp. 1003-1011. 13. How to Detect Propaganda. Publications of the Institutefor Propaganda Analysis. 1937, pp. 210-218. 14. DA SAN MARTINO, Giovanni; YU, Seunghak; BARRON-CEDENO, Alberto; PETROV, Rostislav; NAKOV, Preslav. Fine-grained analysis of propaganda in news article. 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, pp. 5636-5646. 15. BARRON-CEDENO, Alberto; MARTINO, Giovanni Da San; JARADAT, Israa; NAKOV, Preslav. [N.d.]. Available from DOI: 10 . 48550/ARXIV.1912.06810. 16. BAISA, Vit; H E R M A N , Ondfej; HORAK, Ales. Benchmark Dataset for Propaganda Detection in Czech Newspaper Texts. In: Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2019). Varna, Bulgaria: I N C O M A 50 BIBLIOGRAPHY Ltd., 2019, pp. 77-83. Available from DOI: 10 . 26615/978-954- 452-056-4_010. 17. NEAL, Tempestt; SUNDARARAJAN, Kalaivani; FATIMA, Aneez; Y A N , Yiming; XIANG, Yingfei; WOODARD, Damon. Surveying stylometry techniques and applications. ACM Computing Surveys (CSuR). 2017, vol. 50, no. 6, pp. 1-36. 18. CHASKI, Carole E. Author Identification In The Forensic Setting. In: The Oxford Handbook of Language and Law. Oxford University Press, 2012. ISBN 9780199572120. Available from DOI: 10 .1093/ oxfordhb/9780199572120.013.0036. 19. DAELEMANS, Walter. Explanation in Computational Stylometry. In: 2013, pp. 451-462. ISBN 978-3-642-37255-1. Available from DOI: 10.1007/978-3-642-37256-8_37. 20. LAGUTINA, Ksenia; LAGUTINA, Nadezhda; BOYCHUK, Elena; VORONTSOVA, Inna; SHLIAKHTINA, Elena; BELYAEVA, Olga; PARAMONOV, Ilya; DEMIDOV, P G . A Survey on Stylometric Text Features. In: 2019 25th Conference of Open Innovations Association (FRUCT). 2019, pp. 184-195. Available from DOI: 10.23919/ FRUCT48121.2019.8981504. 21. PEARL, Lisa; STEYVERS, Mark. Detecting authorship deception: a supervised machine learning approach using author writeprints. Literary and Linguistic Computing. 2012, vol. 27, no. 2, pp. 183-196. ISSN 0268-1145. Available from DOI: 10.1093/llc/f qs003. 22. STAMATATOS, Efstathios; PARDO, Francisco Manuel Rangel; TSCHUGGNALL, Michael; STEIN, Benno; KESTEMONT, Mike; ROSSO, Paolo; POTTHAST, Martin. Overview of P A N 2018 Author Identification, Author Profiling, and Author Obfuscation. In: CLEF. 2018. 23. CRISTANI, Marco; ROFFO, Giorgio; SEGALIN, Cristina; B A Z Z A N I , Loris; VINCIARELLI, Alessandro; M U R I N O , Vittorio. Conversationally-inspired stylometric features for authorship attribution in instant messaging. In: Proceedings of the 20th ACM international conference on Multimedia. 2012, pp. 1121-1124. 51 BIBLIOGRAPHY 24. DE BRUYNE, Luna; VERHOEVEN, Ben; DAELEMANS, Walter. Stylometric Text Analysis for Dutch-speaking Adolescents with Autism Spectrum Disorder. 2018, vol. 8, pp. 3-23. Available also from: https: //clinjournal. org/clinj/article/view/77. 25. FENG, Song; BANERJEE, Ritwik; CHOI, Yejin. Syntactic stylometry for deception detection. In: Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2012, pp. 171-175. 26. PENNEBAKER, James; BOOTH, Roger; BOYD, Ryan; FRANCIS, Martha. Linguistic Inquiry and Word Count: LIWC2015. 2015. 27. SIRTS, Kairit; PIGUET, Olivier; JOHNSON, Mark. Idea density for predicting Alzheimer's disease from transcribed speech. arXiv preprint arXiv:1706.04473. 2017. 28. HOUVARDAS, John; STAMATATOS, Efstathios. N-gram feature selection for authorship identification. In: International conference on artificial intelligence: Methodology, systems, and applications. Springer, 2006, pp. 77-86. 29. LUYCKX, Kim; D A E L E M A N S , Walter. Authorship Attribution and Verification with Many Authors and Limited Data. In: Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008). Manchester, U K : Coling 2008 Organizing Committee, 2008, pp. 513-520. Available also from: https : //aclanthology.org/C08-1065. 30. HOOVER, David L. Multivariate Analysis and the Study of Style Variation. Literary and Linguistic Computing. 2003, vol. 18, no. 4, pp. 341-360. ISSN 0268-1145. Available from DOI: 10.1093/llc/ 18.4.341. 31. PANDIAN, A. Anbarasa; RAGAVI, R ; R A M A L I N G A M , Vennila. Feature Extraction and Feature Selection process in Authorship Identification for Tamil Language. In: 2020. 32. JUOLA, Patrick et al. Authorship attribution. Foundations and Trends® in Information Retrieval. 2008, vol. 1, no. 3, pp. 233-334. 33. SHAMS, Rushdi; MERCER, Robert E. Supervised classification of spam emails with natural language stylometry. Neural Computing and Applications. 2016, vol. 27, no. 8, pp. 2315-2331. 52 BIBLIOGRAPHY 34. ANCHIÉTA, Rafael T; NETO, Francisco Assis Ricarte; SOUSA, Rogério Figueiredo de; MOURA, Raimundo Santos. Using stylometric features for sentiment classification. In: International Conference on Intelligent Text Processing and Computational Linguistics. Springer, 2015, pp. 189-200. 35. MARKOV, Ilia; DAELEMANS, Walter. Improving Cross-Domain Hate Speech Detection by Reducing the False Positive Rate. In: Proceedings of the Fourth Workshop on NLP for Internet Freedom: Censorship, Disinformation, and Propaganda. Online: Association for Computational Linguistics, 2021, pp. 17-22. Available from DOI: 10.18653/vl/2021 .nlp4if -1.3. 36. BAISA, Vít; HEŘMAN, Ondřej; H O R A K , Ales. Manipulative Propaganda Techniques. In: RASLAN. 2017. 37. MICHELFEIT, Jan; POMIKÁLEK, Jan; S U C H O M E L , Vit. Text Tokenisation Using unitok. In: HORÁK, Aleš; RYCHLÝ, Pavel (eds.). RASLAN 2014. Brno, Czech Republic: Tribun E U , 2014, pp. 71-75. ISBN 2336-4289. 38. SMERK, Pavel. Fast morphological analysis of czech. Proceedings of recent advances in slavonie natural language processing, RASLAN. 2009, vol. 2007, pp. 13-16. 39. ŠMERK, Pavel. K počítačové morfologické analýze češtiny [online]. 2010 [cit. 2022-11-09]. Available also from: https : //theses . cz/ id/28r7vj/. Disertační práce. Masarykova univerzita, Fakulta informatikyBrno. SUPERVISOR: prof. PhDr. Karel Pala, CSc. 40. RYGL, Jan. Automatic adaptation of author's stylometric features to document types. In: International Conference on Text, Speech, and Dialogue. Springer, 2014, pp. 53-61. 41. SIMPSON, Edward H . Measurement of diversity. Nature. 1949, vol. 163, no. 4148, pp. 688-688. 42. YULE, C Udny. The statistical study of literary vocabulary. Cambridge University Press, 2014. 53 BIBLIOGRAPHY 43. PEDREGOSA, F.; VAROQUAUX, G.; GRAMFORT, A.; MICHEL, V.; THIRION, B.; GRISEL, O.; BLONDEL, M.; PRETTENHOFER, P.; WEISS, R.; D U B O U R G , V.; VANDERPLAS, J.; PASSOS, A.; COURNAPEAU, D.; BRUCHER, M.; PERROT, M.; DUCHESNAY, E. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research. 2011, vol. 12, pp. 2825-2830. 44. C H A W L A , Nitesh V; BOWYER, Kevin W; H A L L , Lawrence O; KEGELMEYER, W Philip. SMOTE: synthetic minority over-sampling technique. Journal of artificial intelligence research. 2002, vol. 16, pp. 321-357. 45. SIDO, Jakub; PRAŽÁK, Ondřej; PŘIBÁŇ, Pavel; PASEK, Jan; SEJÁK, Michal; KONOPÍK, Miloslav. Czert - Czech BERT-like Model for Language Representation. arXiv preprint arXiv.2103.13031. 2021. Available from arXiv: 2103.13031 [es . CL]. 46. ARKHIPOV, Mikhail; TROFIMOVA, Maria; KURATOV, Yuri; SOROKIN, Alexey. Tuning Multilingual Transformers for Language-Specific Named Entity Recognition. In: Proceedings of the 7th Workshop on Balto-Slavic Natural Language Processing. Florence, Italy: Association for Computational Linguistics, 2019, pp. 89-93. Available from DOI: 10.18653/vl/W19-3712. 47. STRAKA, Milan; NÁ PLAVA, Jakub; STRAKOVÁ, Jana; SAMUEL, David. RobeCzech: Czech RoBERTa, a Monolingual Contextualized Language Representation Model. In: Text, Speech, and Dialogue. Springer International Publishing, 2021, pp. 197-209. Available from DOI: 10.1007/978-3-030-83527-9_17. 48. CONNEAU, Alexis; KHANDELWAL, Kartikay; GOYAL, Naman; CHAUDHARY, Vishrav; WENZEK, Guillaume; G U Z M A N , Francisco; GRAVE, Edouard; OTT, Myle; ZETTLEMOYER, Luke; STOYANOV, Veselin. Unsupervised Cross-lingual Representation Learning at Scale. CoRR. 2019, vol. abs/1911.02116. Available from arXiv: 1911.02116. 54 A The Complete Table of Results The following appendix features full, unfiltered version of results for all the considered approaches. Table A . l : The complete table of results Method Atrribute Weighted F l 0 Multinomial Naive Bayes (text only) genre 85.9 1 Random Forest genre 94.92 2 Random Forest (text only) genre 92.45 3 Random Forest (reduced stylometry) genre 95.28 4 Support Vector Machines (text only) genre 94.48 5 Support Vector Machines genre 94.34 6 Perceptron (text only) genre 67.3 7 Perceptron genre 93.8 8 Czert (text only) genre 95.77 9 Czert (with stylometry) genre 95.51 10 Slavic Bert (text only) genre 95.69 11 Slavic Bert (with stylometry) genre 96 12 Slavic Bert (reduced stylometry) genre 95.8 13 Robeczech (text only) genre 96.15 14 Robeczech (with stylometry) genre 96.11 15 X L M Roberta Large (text only) genre 96.46 16 X L M Roberta Large (with stylometry) genre 96.64 17 X L M Roberta Large (reduced stylometry) genre 96.73 18 Multinomial Naive Bayes (text only) topic 37.05 19 Random Forest topic 59.01 20 Random Forest (text only) topic 59.85 21 Random Forest (reduced stylometry) topic 60.23 22 Support Vector Machines (text only) topic 60.5 23 Support Vector Machines topic 57.8 24 Perceptron (text only) topic 62.29 25 Perceptron topic 47.8 26 Czert (text only) topic 69.98 27 Czert (with stylometry) topic 69.85 28 Slavic Bert (text only) topic 68.47 29 Slavic Bert (with stylometry) topic 69.72 30 Slavic Bert (reduced stylometry) topic 69.58 31 Robeczech (text only) topic 67.3 32 Robeczech (with stylometry) topic 68.22 33 X L M Roberta Large (text only) topic 71.93 - Continued on next page - 55 A. T H E COMPLETE TABLE OF RESULTS - Continued from previous page Method Atrribute Weighted F l 34 X L M Roberta Large (with stylometry) topic 70.72 35 X L M Roberta Large (reduced stylometry) topic 71.12 36 Multinomial Naive Bayes (text only) scope 81.8 37 Random Forest scope 84.56 38 Random Forest (text only) scope 81.27 39 Random Forest (reduced stylometry) scope 85.63 40 Support Vector Machines (text only) scope 86.77 41 Support Vector Machines scope 80.88 42 Perceptron (text only) scope 86.1 43 Perceptron scope 80.46 44 Czert (text only) scope 89.28 45 Czert (with stylometry) scope 89.81 46 Slavic Bert (text only) scope 89.85 47 Slavic Bert (with stylometry) scope 89.51 48 Slavic Bert (reduced stylometry) scope 88.9 49 Robeczech (text only) scope 89.09 50 Robeczech (with stylometry) scope 89.62 51 X L M Roberta Large (text only) scope 89.36 52 X L M Roberta Large (with stylometry) scope 90.12 53 X L M Roberta Large (reduced stylometry) scope 90.04 54 Multinomial Naive Bayes (text only) location 40.74 55 Random Forest location 67.7 56 Random Forest (text only) location 62.51 57 Random Forest (reduced stylometry) location 68.26 58 Support Vector Machines (text only) location 74.16 59 Support Vector Machines location 65.76 60 Perceptron (text only) location 72.22 61 Perceptron location 50.67 62 Czert (text only) location 81.66 63 Czert (with stylometry) location 82.05 64 Slavic Bert (text only) location 81.96 65 Slavic Bert (with stylometry) location 82.08 66 Slavic Bert (reduced stylometry) location 82.08 67 Robeczech (text only) location 80.41 68 Robeczech (with stylometry) location 80.77 69 X L M Roberta Large (text only) location 82.95 70 X L M Roberta Large (with stylometry) location 82.82 71 X L M Roberta Large (reduced stylometry) location 83.77 72 Multinomial Naive Bayes (text only) argumentation 50.18 73 Random Forest argumentation 66.58 74 Random Forest (text only) argumentation 67.25 75 Random Forest (reduced stylometry) argumentation 67.61 - Continued on next page - 56 A . T H E COMPLETE TABLE OF RESULTS - Continued from previous page Method Atrribute Weighted F l 76 Support Vector Machines (text only) argumentation 68.84 77 Support Vector Machines argumentation 70.28 78 Perceptron (text only) argumentation 68.29 79 Perceptron argumentation 61.97 80 Czert (text only) argumentation 69.82 81 Czert (with stylometry) argumentation 69.39 82 Slavic Bert (text only) argumentation 68.85 83 Slavic Bert (with stylometry) argumentation 69.87 84 Slavic Bert (reduced stylometry) argumentation 70.07 85 Robeczech (text only) argumentation 69.46 86 Robeczech (with stylometry) argumentation 70.54 87 X L M Roberta Large (text only) argumentation 70.69 88 X L M Roberta Large (with stylometry) argumentation 70.64 89 X L M Roberta Large (reduced stylometry) argumentation 69.89 90 Multinomial Naive Bayes (text only) emotions 77.83 91 Random Forest emotions 80.54 92 Random Forest (text only) emotions 79.55 93 Random Forest (reduced stylometry) emotions 80.87 94 Support Vector Machines (text only) emotions 77.92 95 Support Vector Machines emotions 79.96 96 Perceptron (text only) emotions 82 97 Perceptron emotions 77.23 98 Czert (text only) emotions 81.42 99 Czert (with stylometry) emotions 79.72 100 Slavic Bert (text only) emotions 78.09 101 Slavic Bert (with stylometry) emotions 78.31 102 Slavic Bert (reduced stylometry) emotions 78.06 103 Robeczech (text only) emotions 78.15 104 Robeczech (with stylometry) emotions 78.74 105 X L M Roberta Large (text only) emotions 81.81 106 X L M Roberta Large (with stylometry) emotions 81.75 107 X L M Roberta Large (reduced stylometry) emotions 82 108 Multinomial Naive Bayes (text only) overall sentiment 74.58 109 Random Forest overall sentiment 79.62 110 Random Forest (text only) overall sentiment 80.17 111 Random Forest (reduced stylometry) overall sentiment 80.04 112 Support Vector Machines (text only) overall sentiment 79.77 113 Support Vector Machines overall sentiment 79.96 114 Perceptron (text only) overall sentiment 81.33 115 Perceptron overall sentiment 78.07 116 Czert (text only) overall sentiment 81.98 117 Czert (with stylometry) overall sentiment 81.73 - Continued on next page - 57 A. T H E COMPLETE TABLE OF RESULTS - Continued from previous page Method Atrribute Weighted F l 118 Slavic Bert (text only) overall sentiment 81.31 119 Slavic Bert (with stylometry) overall sentiment 79.71 120 Slavic Bert (reduced stylometry) overall sentiment 79.91 121 Robeczech (text only) overall sentiment 80.41 122 Robeczech (with stylometry) overall sentiment 80.35 123 X L M Roberta Large (text only) overall sentiment 83.14 124 X L M Roberta Large (with stylometry) overall sentiment 82.56 125 X L M Roberta Large (reduced stylometry) overall sentiment 82.9 126 Multinomial Naive Bayes (text only) Russia 56.26 127 Random Forest Russia 81.27 128 Random Forest (text only) Russia 78.43 129 Random Forest (reduced stylometry) Russia 81.61 130 Support Vector Machines (text only) Russia 78.41 131 Support Vector Machines Russia 79.34 132 Perceptron (text only) Russia 78.14 133 Perceptron Russia 75.75 134 Czert (text only) Russia 81.77 135 Czert (with stylometry) Russia 81.46 136 Slavic Bert (text only) Russia 81.12 137 Slavic Bert (with stylometry) Russia 81.37 138 Slavic Bert (reduced stylometry) Russia 81.84 139 Robeczech (text only) Russia 80.9 140 Robeczech (with stylometry) Russia 80.64 141 X L M Roberta Large (text only) Russia 82.88 142 X L M Roberta Large (with stylometry) Russia 83.42 143 X L M Roberta Large (reduced stylometry) Russia 83.37 144 Multinomial Naive Bayes (text only) opinion 80.52 145 Random Forest opinion 87.78 146 Random Forest (text only) opinion 87.25 147 Random Forest (reduced stylometry) opinion 88.51 148 Support Vector Machines (text only) opinion 86.81 149 Support Vector Machines opinion 85.32 150 Perceptron (text only) opinion 87.72 151 Perceptron opinion 85.56 152 Czert (text only) opinion 88.27 153 Czert (with stylometry) opinion 87.77 154 Slavic Bert (text only) opinion 88.36 155 Slavic Bert (with stylometry) opinion 88.29 156 Slavic Bert (reduced stylometry) opinion 88.71 157 Robeczech (text only) opinion 88.9 158 Robeczech (with stylometry) opinion 88.2 159 X L M Roberta Large (text only) opinion 87.61 - Continued on next page - 58 A . T H E COMPLETE TABLE OF RESULTS - Continued from previous page Method Atrribute Weighted F l 160 X L M Roberta Large (with stylometry) opinion 87.64 161 X L M Roberta Large (reduced stylometry) opinion 88.35 162 Multinomial Naive Bayes (text only) expert 48.82 163 Random Forest expert 68.39 164 Random Forest (text only) expert 68.2 165 Random Forest (reduced stylometry) expert 69.19 166 Support Vector Machines (text only) expert 72.54 167 Support Vector Machines expert 70.13 168 Perceptron (text only) expert 70.58 169 Perceptron expert 64.64 170 Czert (text only) expert 72.98 171 Czert (with stylometry) expert 74.28 172 Slavic Bert (text only) expert 74.54 173 Slavic Bert (with stylometry) expert 74.81 174 Slavic Bert (reduced stylometry) expert 74.67 175 Robeczech (text only) expert 74.09 176 Robeczech (with stylometry) expert 73.39 177 X L M Roberta Large (text only) expert 76.1 178 X L M Roberta Large (with stylometry) expert 77.12 179 X L M Roberta Large (reduced stylometry) expert 77.28 180 Multinomial Naive Bayes (text only) source 44.3 181 Random Forest source 52.87 182 Random Forest (text only) source 51.59 183 Random Forest (reduced stylometry) source 54.26 184 Support Vector Machines (text only) source 54.73 185 Support Vector Machines source 51.58 186 Perceptron (text only) source 52.94 187 Perceptron source 51.59 188 Czert (text only) source 52.92 189 Czert (with stylometry) source 53.10 190 Slavic Bert (text only) source 53.12 191 Slavic Bert (with stylometry) source 54.58 192 Slavic Bert (reduced stylometry) source 56.43 193 Robeczech (text only) source 51.84 194 Robeczech (with stylometry) source 55.93 195 X L M Roberta Large (text only) source 52.06 196 X L M Roberta Large (with stylometry) source 54.25 197 X L M Roberta Large (reduced stylometry) source 55.46 198 Multinomial Naive Bayes (text only) fear-monj gering 88.9 199 Random Forest fear-monj gering 90.47 200 Random Forest (text only) fear-monj gering 90.31 201 Random Forest (reduced stylometry) fear-mon^gering 90.84 - Continued on next page - 59 A . T H E COMPLETE TABLE OF RESULTS - Continued from previous page Method Atrribute Weighted F l 202 Support Vector Machines (text only) fear-mongering 90.5 203 Support Vector Machines fear-mongering 90.6 204 Perceptron (text only) fear-mongering 90.95 205 Perceptron fear-mongering 61.6 206 Czert (text only) fear-mongering 90.48 207 Czert (with stylometry) fear-mongering 90.83 208 Slavic Bert (text only) fear-mongering 89.12 209 Slavic Bert (with stylometry) fear-mongering 91.05 210 Slavic Bert (reduced stylometry) fear-mongering 91.22 211 Robeczech (text only) fear-mongering 91.80 212 Robeczech (with stylometry) fear-mongering 91.42 213 X L M Roberta Large (text only) fear-mongering 91.71 214 X L M Roberta Large (with stylometry) fear-mongering 91.5 215 X L M Roberta Large (reduced stylometry) fear-mongering 91.94 216 Multinomial Naive Bayes (text only) blaming 60.68 217 Random Forest blaming 72.74 218 Random Forest (text only) blaming 70.91 219 Random Forest (reduced stylometry) blaming 72.74 220 Support Vector Machines (text only) blaming 73.87 221 Support Vector Machines blaming 63.32 222 Perceptron (text only) blaming 72.03 223 Perceptron blaming 63.5 224 Czert (text only) blaming 73.52 225 Czert (with stylometry) blaming 73.61 226 Slavic Bert (text only) blaming 72.84 227 Slavic Bert (with stylometry) blaming 73.39 228 Slavic Bert (reduced stylometry) blaming 73.55 229 Robeczech (text only) blaming 72.85 230 Robeczech (with stylometry) blaming 73.43 231 X L M Roberta Large (text only) blaming 74.55 232 X L M Roberta Large (with stylometry) blaming 73.41 233 X L M Roberta Large (reduced stylometry) blaming 74.92 234 Multinomial Naive Bayes (text only) labelling 76.7 235 Random Forest labelling 82.49 236 Random Forest (text only) labelling 81.77 237 Random Forest (reduced stylometry) labelling 83.22 238 Support Vector Machines (text only) labelling 81.18 239 Support Vector Machines labelling 81.6 240 Perceptron (text only) labelling 83.59 241 Perceptron labelling 67.67 242 Czert (text only) labelling 82.72 243 Czert (with stylometry) labelling 81.41 - Continued on next page - 60 A. T H E COMPLETE TABLE OF RESULTS - Continued from previous page Method Atrribute Weighted F l 244 Slavic Bert (text only) labelling 80.85 245 Slavic Bert (with stylometry) labelling 80.88 246 Slavic Bert (reduced stylometry) labelling 81.62 247 Robeczech (text only) labelling 82.28 248 Robeczech (with stylometry) labelling 82.68 249 X L M Roberta Large (text only) labelling 83.37 250 X L M Roberta Large (with stylometry) labelling 83 251 X L M Roberta Large (reduced stylometry) labelling 83.06 252 Multinomial Naive Bayes (text only) demonization 95.7 253 Random Forest demonization 95.89 254 Random Forest (text only) demonization 95.8 255 Random Forest (reduced stylometry) demonization 96.14 256 Support Vector Machines (text only) demonization 95.7 257 Support Vector Machines demonization 95.8 258 Perceptron (text only) demonization 96.11 259 Perceptron demonization 88.94 260 Czert (text only) demonization 95.61 261 Czert (with stylometry) demonization 95.65 262 Slavic Bert (text only) demonization 95.67 263 Slavic Bert (with stylometry) demonization 95.67 264 Slavic Bert (reduced stylometry) demonization 95.67 265 Robeczech (text only) demonization 95.67 266 Robeczech (with stylometry) demonization 95.62 267 X L M Roberta Large (text only) demonization 96.13 268 X L M Roberta Large (with stylometry) demonization 96.19 269 X L M Roberta Large (reduced stylometry) demonization 96.10 270 Multinomial Naive Bayes (text only) relativization 92.3 271 Random Forest relativization 93.27 272 Random Forest (text only) relativization 93.01 273 Random Forest (reduced stylometry) relativization 93.31 274 Support Vector Machines (text only) relativization 89.85 275 Support Vector Machines relativization 93.15 276 Perceptron (text only) relativization 92.72 277 Perceptron relativization 91.86 278 Czert (text only) relativization 92.88 279 Czert (with stylometry) relativization 92.81 280 Slavic Bert (text only) relativization 92.8 281 Slavic Bert (with stylometry) relativization 92.21 282 Slavic Bert (reduced stylometry) relativization 92.56 283 Robeczech (text only) relativization 92.55 284 Robeczech (with stylometry) relativization 92.61 285 X L M Roberta Large (text only) relativization 92.75 - Continued on next page - 61 A . T H E COMPLETE TABLE OF RESULTS - Continued from previous page Method Atrribute Weighted F l 286 X L M Roberta Large (with stylometry) relativization 92.84 287 X L M Roberta Large (reduced stylometry) relativization 92.67 288 Multinomial Naive Bayes (text only) fabulation 74.8 289 Random Forest fabulation 78.37 290 Random Forest (text only) fabulation 78 291 Random Forest (reduced stylometry) fabulation 78.57 292 Support Vector Machines (text only) fabulation 78.17 293 Support Vector Machines fabulation 75.2 294 Perceptron (text only) fabulation 80.15 295 Perceptron fabulation 76.74 296 Czert (text only) fabulation 78.91 297 Czert (with stylometry) fabulation 80.81 298 Slavic Bert (text only) fabulation 79.38 299 Slavic Bert (with stylometry) fabulation 79.48 300 Slavic Bert (reduced stylometry) fabulation 79.14 301 Robeczech (text only) fabulation 79.86 302 Robeczech (with stylometry) fabulation 78.54 303 X L M Roberta Large (text only) fabulation 80.57 304 X L M Roberta Large (with stylometry) fabulation 80.11 305 X L M Roberta Large (reduced stylometry) fabulation 80.92 62