This paper presents the development and results of two distinct modules designed to detect toxic language in Ukrainian texts, primarily implemented within the "TextAttributor 1.0" expert system. The first module utilizes a rule-based approach, analyzing text through predefined linguistic rules and lexiconbased methods, while the second employs machine learning techniques, specifically leveraging the fastText and LLAMA-3 models, to automatically detect toxic content. The rule-based module outputs a detailed linguistic analysis, mapping toxic vocabulary using a precompiled lexicographic database, while the machine learning module calculates toxicity based on statistical models. The performance of both methods was evaluated by comparing their results on a corpus of Ukrainian texts, with the Pearson to effectively identify toxic content, contributing to ongoing efforts to mitigate the spread of harmful information. This paper contains rude texts that only serve as illustrative examples.
The study of text toxicity represents a relatively new field of inquiry within the broader domain of
sentiment analysis. The tonality of a text is a significant element that influences how it is perceived
and understood by the reader. It also enables the author to achieve their communicative objectives.
For this reason, the task of determining the tone of a text is of great interest today not only to
modern linguists and computer scientists but also to political scientists, managers, marketers,
advertisers, image makers, and other professionals working with a particular brand. This task has
become particularly crucial with the advent of the Internet, as it has provided a new platform for
the analysis of media texts that "transmit, store, and reproduce information that influences public
opinion" [
0000-0001-8932-9301 (N. Darchuk); 0000-0002-2644-3892 (O. Zuban); 0000-0003-2266-7650 (V. Robeiko); 0000-00029684-3840 (Yu. Tsyhvintseva); 0000-0003-1169-6851 (M. Sazhok) leads to an inadequate perception of reality, underscores the acute urgency of the problem of "ecology" and the protection of the Internet space.
The issue of addressing the proliferation of destructive online content is a matter of global
concern and significance. Since 2019, June 18 has been designated by the United Nations as the
International Day against Hate Speech. This year, the Council of Europe held the Week against
Hate Speech on June 17-20 in Strasbourg. In 2022, the Committee of Ministers of the Council of
Europe developed Recommendations on combating hate speech [
The full-scale Russian-Ukrainian war has highlighted the urgency of this issue, as hate speech and other forms of aggressive propaganda represent a crucial element of the Russian propaganda apparatus in its efforts to erode the identity and integrity of the Ukrainian nation. To address this issue, it is essential to develop tools for automated analysis and detection of Ukrainian-language textual content that negatively impacts an individual's psychological state, public consciousness, and infringes upon the rights and legitimate interests of users, society, and the state.
In light of the pressing necessity for such tools, our team has developed IT solutions for the
automated identification of toxicity in Ukrainian-language text. These tools are integrated into
TextAttributor 1.0 [
The objective of the present article is twofold: firstly, to analyze the implementation of two methods the lexicon and rules-based method, and the deep learning-based method in the automatic analysis of the toxicity of Ukrainian-language texts; secondly, to investigate the effectiveness of the TextAttributor 1.0 system using the two methods on the texts analyzed by the system. The object of this study is Ukrainian-language texts. The subject of this study is the criteria and methods used for the automatic identification of toxic Ukrainian-language texts. The study employs a range of methods, including: a rules and lexicon-based sentiment analysis method, a deep learning-based method, and a combination of linguistic methods, namely component analysis, distributional analysis, and taxonomic analysis. Additionally, statistical methods, such as toxicity indexing and the calculation of the Pearson correlation coefficient for sample data, are utilized. Graphical modeling of statistical data is also employed as a method to visualize 520 Ukrainianlanguage texts that were submitted by users of the TextAttributor 1.0 web application.
The latest developments in the field indicate that the task of automatic toxicity detection is
currently solved mainly by applying deep learning methods based on various architectures (CNN,
LSTM, BERT) [
Recently, there have also been publications devoted to the detection of toxic language in
lowresource languages, including Ukrainian. This is primarily due to the release of multilingual large
language models (LLMs) and the development of new methods for creating datasets, including the
development of translated datasets and the generation of synthetic data. To date, there is no
publicly available expert-annotated toxic text dataset. The only existing corpus of this kind is the
Ukrainian tweets corpus [
The lexicon-based method is employed to determine the toxicity of a text. This entails identifying toxic words within the text in accordance with pre-compiled dictionaries of toxic vocabulary. The text is then evaluated on a toxicity scale according to the number of instances of toxic vocabulary identified. This method is employed in conjunction with automatic morphological and syntactic analysis, namely automatic rules-based text analysis. For this reason, the dictionary-based toxicity assessment may also be referred to as the dictionary and lexicon-based method or the rules-based method. The accuracy and completeness of the toxicity determination using this method are contingent upon the scope and quality of the compiled lexicon, as well as the quality of the lemmatization procedure for the text. The scope and semantics of the words in the toxic dictionary are contingent upon the style, genre, and subject matter of the texts for which the analysis system is being constructed.
The efficacy of the dictionary-based approach is contingent upon its capacity to provide a comprehensive linguistic analysis of text toxicity, culminating in the formulation of an expert opinion. This entails the identification of a list of lexical toxic units that characterize the text in question. In contrast, the considered machine learning method is limited to binary classification of text toxicity based on the features and degree of toxicity. It is worth noting that the lexicon-based method can be used to assess the toxicity of texts of varying lengths, even in the absence of toxic datasets. One disadvantage of the method based on dictionaries and rules is the limited nature of the lexicon. It is inherently incomplete because communication generates new means of expressing toxicity and new toxic discourses that transform neutral vocabulary into the vocabulary of destructive influence. This in turn requires the ongoing addition of new terms to the lexicon.
In the present study, our task was to compile lexical lists of toxic lexical means (including idiomatic expressions) regardless of the topic of toxic texts in Ukrainian-language online media discourse.
A toxic text is typically defined as an offensive comment or publication that exhibits one or
more of the following characteristics: harassment, threat, obscenity, cyberbullying, trolling,
indignation, and identity-based hate text [
Accordingly, the lexical lists of toxic vocabulary were compiled with the notion of this broader
interpretation of the concept of a "toxic text" based on the following data: 1) A textual sample of
approximately two million words [
Lexicographic List No. 1, the "Dictionary of Emotionogens," contains over 5,000 lexemes that verbalize negative facts, emotions, and assessments, causing the recipient to experience anxiety, fear, confusion, shame, guilt, oppression, and control. The dictionary includes words as independent parts of speech with a negative tonality (rated "-2" or "very negative"), namely nouns (e.g., (debt), (ignore), ), adjectives (e.g., (difficult), (nuclear)), adverbs (e.g.,
), verbs (e.g., subvert), (to ignore), (to nullify), scam)) and adjectives formed from verbs (e.g., (knocked),
(torn), (to prohibit), (to risk), (squandered)).
The lexical list, designated No. 2, "Hate Speech Dictionary," contains 3,000 lexemes that verbalize aggressive communication, including harassment, threats, obscenities, cyberbullying, trolling, (to (to (lesbo), (scammer), outrage, and identity-based hate text. The list covers the following lexical groups: negative names for people 1,620 (e.g., (scammer), (fatty), (torturer), ), obscene vocabulary
613 (e.g., (fucker), (to fuck up)) and abusive vocabulary and vulgarisms 787 (e.g., shitstorm), (bastard)). Each item in the list is marked with semantic characteristics according to the developed classifications. For example, in the dictionary "Negative Names for People," the lexemes are grouped into 18 categories that correspond to key semantic features: negative designations of a person based on age ( ), sexual orientation ( (queer), (lesbian), ), gender ( ), appearance ( (beanpole), (dystrophic), (fatty)), antisocial behavior ( (swindler), (swindler), (bandit), (clown), (gangster), (kingpin), (traitor)), social activity / passivity ( - (defector), (indifferent)), nationality ( (defector), (banderite), (negro)), social status ( (bastard),
), profession and financial status ( (bankrupt), ), political affiliation ( (nazi), (leftist)), place of residence ( (pedorussia), (hillbilly)), religion ( (atheist), (sectarian)), intellectual abilities ( (idiot), (hack), ), diseases ( (down)), and comparison to plants or animals ( (amoeba), (sheep), (otter), (lamb), (pig)) etc. Each word of hate speech received has been annotated as belonging to one of the following categories: s sexism, r racism, e ageism, etc.
Lexicographic List No. 3, "Toxic Compounds," encompasses 1,500 stable phrases that idiomatically reflect toxic sentiment. The items in the list are classified by 26 semantic features, with each combination assigned a semantic label of toxicity (T), identity-based hate speech (IBYS), or expressiveness (E), e.g., IBYS 4. Claims of inferiority, moral flaws ( (dirty gypsy), ) IBYS 7. Mention in a derogatory or insulting context, obscene comments ( (lousy (pigs have no word around here), ) IBYS 10.
Threats of physical destruction or any violence ( (death to Jews)) ) Bullying ( (degenerate brat)) and malicious jokes aimed at eliciting an emotional response from readers, or trolling ( (supreme traitors), etc.( tonality ( (rotten essence), (atrophied brain), ), etc.
)
The described lexicon-based method for toxicity detection in Ukrainian-language texts was implemented within the TextAttributor 1.0 system as a rule-based module. The general characteristics of the rule-based toxicity detection module are as follows: the input is user-provided text, while the output is: 1) a numerical value of the statistical index of text toxicity; 2) the absolute frequencies of toxic words and phrases classified by semantic classes; 3) the text with toxic words and phrases highlighted.
The linguistic statistical analysis of toxicity is performed for each text using the program code for calling the analyzers in the following sequence of actions: 1. 2. 3. 4.
The text is tokenized into separate sentences and words; Morphological annotation of words; Followed by a contextual analysis that refines the morphological annotation codes; Lemmatization of words;
Identification of toxic vocabulary in the analyzed text: comparison of words in the text with the registry of the toxic lexicon database and calculation of the absolute frequencies of toxic words and phrases;
Calculation of the text toxicity index;
Generation of a linguistic expert report: a statistical map of toxic vocabulary in addition to the text where toxic words and phrases are highlighted.
The rule-based module operates on the basis of a lexicographic database of toxic tabled vocabulary, compiled from three lexicographic lists. The respective table (developed based on MS SQL Server) contains 9,500 rows.
Description of the data structure of the table: [wid] [did] [wrd] record identifier; dictionary identifier (used for grouping phenomena by dictionary); word form or lemma; if the column then this column would be a lemma; [mitka]
sub-category label; [updateDate]
record update date; [updatedBy] the user who updated the record; [sdcatitemID]
category identifier; [wrd2]
next words of the phrase after the first (if phrase, if `sltype=1`; [wrd3]
the third lemma of the phrase, if [sltype] an indicator that this line contains a phrase stored as text or as lemmas.
The text toxicity index (Itox) is calculated using the following formula: = 10 ( + | |(
+ ))⁄ , where: e the number of emotionogen words (Lexicographic List No. 1), m the number of hate speech words (Lexicographic List No. 2), t the number of toxic phrases (Lexicographic List No. 3) in the text analyzed by the system; K
an intensifier of aggressive toxic speech units (K = 2); n number of words in the analyzed text; 10 normalization factor for the statistical parameter.
(1)
The formula for the statistical parameter the text toxicity index takes into account the usage frequency for various classes of vocabulary in the text, differentiated by the semantic features of three lexicons: emotionogenic words (e), hate speech lexemes (m), toxic phraseological compounds (t). To enhance the weight of toxic means of aggressive communication, a coefficient K (on a fivepoint scale [-2, -1, 0, +1, +2] it corresponds to -2) has been introduced into the formula. This coefficient intensifies the weight of hate speech words (Lexicographic List 2) and toxic phraseological compounds (Lexicographic List 3) in determining the level of toxicity. The procedure of multiplying by 10 has been included into the formula to increase the empirical value of the toxicity index for normalizing the linguistic statistical parameters of the stylometric analysis in the TextAttributor 1.0 system.
Let us consider the results of toxic speech analysis using a sample text: « form, rather than my Ukraine"), which consists of 52 sentences and 500 words. The toxicity index of the analyzed text (Itox) has been assessed at 0.7, indicating a slightly higher level of toxicity compared to the upper limit of the average toxicity value for the media style in the Ukrainian language (Fig. 1).
The system generates a statistical map of the linguistic expert report on text toxicity (Fig. 1) which presents a list of word semantic classes identified by the system in the analyzed text, according to the classification markers of the toxic vocabulary database: Semantic category column 1; name of semantic feature column 2; number of words/phrases in the analyzed text according to semantic features column 3. In particular, the statistical map of the text "In 2019, a completely different country began to form, rather than my Ukraine" systematizes the following linguistic data: emotionogens 18; negative names for a person based on intellectual ability 2; vulgarisms 2; negative names for a person by comparison with mythical creatures 1; negative names for a person based on health characteristics 1; negative names for a person (sarcasm, idiomatic expression) 1; negative names for a person based on body parts or physiological processes 1.
The verbalization of statistics for identified toxic vocabulary in the text, presented in the statistical map according to semantic categories, is visualized in a separate window using bold black font marking specific lexical means (Fig. 2).
By comparing the statistical map data with the text, one can systematize toxic means, for example: emotionogens 18 (nouns: (horror), (victim), (cruelty), (corruption), (propaganda), ; adjectives: (painful), (uncontrollable), (doomed), (terrible); adverb: (unfortunately); verbs:
); vulgarisms 2 ( (finished)); toxic phrase (to kiss ass); negative person descriptors based on various characteristics ( (idiot), ).
The provided text also includes indication of other features: 1) phraseoligisms are highlighted in blue font ( etc.); some words and symbols not recognized by the system's automatic morphological analysis are highlighted in red font ( ). The system does not consider these units when calculating the text toxicity index. Testing the system's performance demonstrates the need to refine the lexicographic database, especially regarding enhance the text preprocessing and lemmatization procedure. We consider this one of the primary prospective tasks.
4.1. Data For our experimental study of toxicity in media texts using machine learning methods, two datasets containing short Internet texts from Ukrainian-language blogs, comments, articles, etc., have been prepared.
Dataset 1. For the initial analysis of the issue, a trial corpus of media texts was formed, consisting of 668 documents featuring hate speech and expert annotations (each text sample was classified by an expert into a specific category neutrality or hate speech). 226 texts were identified as toxic. The obtained dataset was randomly split into a training set (600 texts, 192 of which are toxic, 94,905 word samples, 11,852 stems) and a test set (68 texts, 34 of which are toxic, 10,800 word samples).
Dataset 2. The training set was augmented by incorporating annotated 11,387 text documents, among which 2,155 are toxic. The test set remained unchanged.
The typical pipeline of Machine Learning methods involves preprocessing the text data, extracting features, and then using a classifier to determine if the text is toxic. Feature extraction techniques include:
• Bag-of-Words (BoW) is a simple approach where text is represented as an unordered collection of words, ignoring grammar and word order but preserving frequency.
• Term Frequency-Inverse Document Frequency (TF-IDF) enhances the BoW model by weighing word frequency relative to its importance across documents, thus reducing the weight of commonly used words.
• Pre-trained word embeddings like Word2Vec or GloVe can be used to represent words as dense vectors, capturing semantic similarities.
• N-grams capture short phrases or word sequences to better model context compared to BoW, useful for identifying common toxic word patterns.
• Other linguistic features may include sentiment scores, lexical resources like swear word dictionaries, Part-of-Speech (POS) tags, and syntactic features.
• In experimental research described in this paper we rely on BoW and word embeddings.
In experimental research described in this paper we rely on BoW and word embeddings.
Classical, or primitive, Machine Learning Approaches use linguistic features and classical algorithms like SVM, logistic regression and Naive Bayes. They require significant domain knowledge and are limited in handling complex language patterns.
Naive Bayes: Suitable for text classification due to the assumption of word independence, which often works well despite its simplicity.
Logistic Regression: A widely used linear model for binary classification.
Support Vector Machines (SVM): Works well with high-dimensional data like text, especially when combined with kernel methods to separate non-linear toxic and non-toxic classes.
Random Forest and Decision Trees: Tree-based models are sometimes used to capture nonlinear relationships, but they tend to struggle with sparse data in large text corpora.
In this paper we describe experiments exploiting the Naive Bayes approach.
Deep Learning Approaches leverage neural networks, RNNs, CNNs, LSTMs, and especially transformers like BERT for sophisticated contextual understanding of text. These models can automatically learn features but require substantial computational resources and large datasets.
Hybrid Approaches combine the strengths of both classical and deep learning methods for better performance and robustness.
Challenges across both methods include handling nuanced language, interpretability, data imbalance, and avoiding bias.
In this paper our choice is fastText [
FastText's main advantages are its relative simplicity, speed, and ability to handle large vocabularies and datasets. It is particularly efficient because it: • uses rather shallow neural network, which is much faster than deep models, • embeds subword (character n-grams) information, allowing better generalization, • utilizes hierarchical softmax for computational efficiency during classification, especially for large-scale problems.
The subword presentation is important since the Ukrainian language is highly inflective, and plenty of unseen words as well as words with spelling errors must be covered. The effectiveness of the selected approach is also determined by the technical conditions of system operation and available textual resources for model training.
Text toxicity detection using Large Language Models (LLMs), such as GPT-3 or GPT-4, is an emerging area in Natural Language Processing (NLP) that leverages advanced capabilities of LLMs for identifying harmful, abusive, or toxic language in text. Here are the methods and strategies for using LLM prompting to detect toxicity: • • • • • • •
Direct Toxicity Classification Prompting Chain-of-Thought Prompting (Step-by-Step Reasoning) Few-Shot Learning (Providing Examples) Zero-Shot Prompting with Contextual Framing Prompt Tuning for Toxicity Detection Debiasing and Calibration Prompting
Toxicity Detection as Part of a Larger System (Hybrid Models)
In this work we consider one of the most straightforward methods involves directly prompting
the LLM to classify text as "toxic" or "non-toxic" exploiting Llama v. 3 [
Classify the following sentence as 'toxic' or 'non-toxic': "You are an idiot and nobody likes you."
The LLM is provided with text snippets, and based on its training and understanding of language, it identifies the toxic content. Implementation of this approach is quick and simple, no fine-tuning required. Another advantage is that LLMs can identify nuanced and context-dependent toxicity that may escape simpler classifiers. On the other hand, (a) LLMs may not always be consistent or reliable without additional constraints or examples, (b) false positives or negatives could occur due to ambiguity or complexity in language, (c) huge amounts of computation resources are required, including a powerful GPU with 24GB of RAM.
Evaluating models for toxicity detection requires a careful choice of metrics because of the class imbalance and Precision, Recall, and F1-Score meets these requirements: precision helps in minimizing false positives (important to avoid over-censoring), while recall ensures detection of most toxic content and F1 generalizes these two metrics.
• Classical techniques: The following results were obtained from the testing of Dataset 1 and 2: F1 71 %; precision 60 %; recall 85 %.
• Deep Learning: The best result according to the generalized metric (F1 = 79.4%) was obtained for the following hyperparameter values: the dimension of the space of the vector representation of words 56, the initial learning rate 0.15, the number of epochs 500, the lexical context bigrams, the length of subwords from 2 to 5, the decision-making threshold 0.4. At the same time, for a sensitivity of 85%, an accuracy of about 70% is achieved. • LLM prompting: The following results were obtained on the test sample: F1 precision 74.3 %; recall 76.5 %.
The experimental system for determining the toxicity of media texts based on ML is implemented in a "client-server" architecture using a REST interface. A basic web user interface has been developed, allowing users to input text and receive a response from the system on whether the text is toxic (__label__target) with a confidence score ranging from 0 to 1. Figure 3 shows an example of using the basic web interface to assess the toxicity of a given media text. In turn, using the REST API, TextAttributor acts as a client, sending the user-entered text to the server, receiving a response, and visualizing the result along with other parameters calculated for the given text.
Let us consider the experimental comparison of the toxicity indices of a single text, determined by the two methods considered.
The experimental study aims to compare the results of the system's performance using two methods rules-based and machine learning to determine the effectiveness of these methods in identifying toxic texts. To achieve this goal, a sample of 520 texts was compiled and analyzed using the TextAttributor 1.0 system. The sample included texts of varying lengths (from 33 words to 46,000 words), encompassing different themes, styles, and genres. Each text was evaluated for toxicity using two indices: ToxR is the toxicity index based on the lexicon and rule-based method, while ToxML is the toxicity index as determined by the machine learning method.
Using the two methods in the TextAttributor 1.0 system, toxicity grading occurs on two scales: ToxR is graded on a scale from 0 to 5.78, while ToxML, based on machine learning, ranges from 0 to 1.00. ToxML is limited to a range of 0 1.00, with a decision threshold for text toxicity set at 0.4 (0.35). (neutral text) 0 < lowdetermined experimentally based on the best F1 score for the trained model. In contrast, ToxR has no upper limit or decision-making threshold value for the system. Establishing toxicity using this method involves the identification of lexical markers of text toxicity, their statistical evaluation the text toxicity index while the decision on toxicity is left to the linguist.
During the linguistic analysis of the toxicity of texts examined by the system, the average ToxR value for the media style of the Ukrainian language was empirically measured at 0.4 (0.35). This value will be considered as the decision threshold on a text's toxicity, despite the fact that this numerical value results from statistical analysis of entirely different features than those used in the calculation of ToxML. Empirical numerical values of toxicity indices, determined by two methods for each text in the sample, were ranked in descending order of numerical values: from highest to lowest. Ranking the toxicity indices, calculated by the two methods, allows for the distribution of texts according to toxicity levels: • •
ToxML: toxic texts with an index of 0.4+ constitute 52.7% of the sample (276 texts).
ToxR: toxic texts with an index of 0.4+ constitute 58.8% of the sample (306 texts).
The remaining texts in the sample exhibit either a low degree of toxicity (0 < 0.4) or are devoid of toxic characteristics (neutral text = 0):
• ToxML: low-toxicity texts with an index below 0.4 constitute 37.7% of the sample (196 texts); neutral texts with a toxicity level of 0 constitute 9.6% (50 texts).
• ToxR: low-toxicity texts with an index below 0.4 constitute 37.5 % of the sample (195 texts); neutral texts with a toxicity level of 0 constitute 3.7 % (19 texts).
The percentage ratio of toxic and low-toxicity texts, as determined by the two methods, demonstrates close results; however, the numerical toxicity indicators determined by the two methods for a single text can be drastically different. Ranking texts by descending numerical values of ToxML while preserving the numerical values of ToxR for these texts indicates that: 1. 276 texts are toxic by ToxML; however, 84 of these texts exhibit low toxicity (ToxR < 0.4) or are non-toxic (ToxR = 0) according to the ToxR method, meaning only 192 texts are assessed as toxic by both methods.
2. 244 texts are low-toxicity or neutral (ToxML < 0.4), but 107 of these texts exhibit high the system as low-toxicity/non-toxic by both methods.
Thus, it can be stated that 329 out of 520 texts were classified as toxic/low-toxicity/emotionally neutral by both methods. Thus, we are presented with a question: Is there a statistical relationship between the two variables empirical numerical values of the toxicity indices determined by the two methods?
Comparing the two sets of numerical data ToxML and ToxR toxicity indices we have formulated a task to determine whether there is a statistical dependency in the changes of ToxML and ToxR toxicity levels on a sample of 520 texts, i.e., to ascertain the degree of correlation between the two data sets: the ranked list of ToxML and corresponding numerical values of ToxR. The correlation between the two data sets was measured using three methods: linear Pearson correlation (0.2068); Spearman rank correlation (0.2513); Kendall rank correlation (0.1753).
Let us consider the value of Pearson's linear correlation coefficient, which is most frequently used in humanities and social studies to analyze correlation rather than causal dependency. The correlation coefficient r = 0.2068 is low and indicates a weak positive correlation. However, with an error p = 0.001, considering the degrees of freedom f = 600, the critical correlation coefficient value is r = 0.134. The empirical value of Pearson's coefficient, r = 0.2068, is higher than the critical value and demonstrates high significance for a sample size of 520 texts. This implies that, although the degree of correlation is weak, the indicator has high reliability.
Let us consider a scatter plot for Pearson's coefficient of variation (0.2068), which indicates a weak measure of joint variability between two random variables two toxicity indices, ToxML and ToxR, determined by two methods for a single text (Fig. 4). For the convenience of graphical modeling of correlation, the empirical values of ToxR were converted to a scale from 0 to 1 using the min- norm = (Xi - Xmin) / (Xmax Xmin).
X-axis: numerical values of toxicity levels determined by the machine learning method (ToxML independent variables); Y-axis, the numerical values of the toxicity index determined by the rulebased method (ToxR dependent variables). Each point on the diagram represents a text, with its coordinates being two toxicity indices. The red line is a regression line modeling the relationship between the two variables. Given a weak positive Pearson coefficient, the regression line has a low slope. This indicates a vague trend of increasing Y values with the increase in X values. Considering the decision threshold of 0.4 for determining text toxicity/low toxicity for ToxML and 0.07 (after min-max normalization) for ToxR, four zones were identified on the graph: 1) Texts toxic by ToxR, but low-toxicity or neutral according to ToxML dark gray zone; 2) Texts toxic according to both methods light gray zone; 3) Texts considered low-toxicity or neutral according to both methods light gray zone; 4) Texts toxic by ToxML, but low-toxicity or neutral by ToxR dark gray zone. The correlation of variables (ToxR and ToxML) for texts in zone 1 and zone 4 consequently affects the coefficient of variation, as an inversely proportional correlation is observed in these zones. Low X values correspond to high Y values (Zone 1), whereas high X values correspond to low Y values (Zone 4). We consider the tests of these areas to be contentious with respect to toxicity indicators.
The results of automatic toxicity analysis also demonstrate similar statistical data by the two methods in determining the toxicity/low-toxicity/neutrality of texts, which, according to the scatter plot, are located in zones 2 and 3. To determine the degree of correlation between the statistical data obtained using the two methods for these texts, Pearson's correlation coefficient was calculated for a sample of 329 texts, of which 192 are toxic, 133 are low-toxicity, and 4 are neutral texts. The correlation coefficient r = 0.5344 is relatively high and indicates a moderate positive correlation. Moreover, the average degree of correlation has high reliability, as with p = 0.001 and considering the degrees of freedom f = 350, the critical correlation coefficient value is r = 0.175. The empirical value of Pearson's coefficient, r = 0.5344, is higher than the critical value and demonstrates high significance for a sample size of 329 texts.
A scatter plot was also constructed for Pearson's coefficient of variation (0.5344) (Fig. 5).
The regression line, which models the average positive correlation coefficient on the diagram, has a significantly greater slope compared to the diagram in Fig. 4, which indicates a moderate measure of joint variability between the two random variables ToxR and ToxML. This indicates a clear trend of increasing Y values with the increase in X values.
In Figure 5, four zones were identified using the same principle: 1) Texts toxic by ToxR, but lowtoxicity or neutral according to ToxML - dark gray zone; 2) Texts toxic according to both methods light gray zone; 3) Texts considered low-toxicity or neutral according to both methods - light gray zone; 4) Texts toxic by ToxML, but low-toxicity or neutral by ToxR - dark gray zone. The correlation of variables (ToxR and ToxML) shows that texts rarely fall into zones 1 and 4; instead, the points in zone 2 clearly represent low-toxicity and neutral texts, while zone 4 contains points representing texts with high toxicity. With an average positive correlation coefficient (Fig. 5), the dispersion of points is significantly lower than with a weak positive coefficient (Fig. 4), yet in both cases, "outliers" from the model's range of inference are observed. These are points of texts that deviate from the general trend of variable correlation and are typically characterized by high levels of toxicity according to one or both methods.
The diagram also clearly shows the formation of three clusters: 1) Cluster of points in the range of 0 0.3 (on the X-axis) represents neutral texts and texts with low toxicity; 2) Cluster of points in the range of 0.8 1.0 (on the X-axis) represents texts with high toxicity; 3) Dispersion of points in the range of 0.4 0.8 represents texts of medium toxicity. In the range of 0.3 0.4, there are 5 points texts whose degree of toxicity is difficult to interpret because they fall within the decision threshold range. However, 2 points are closer to cluster 1 (neutral texts and low-toxicity texts), while 3 points are closer to cluster 3 (medium toxicity texts).
promising outcomes, particularly in the automatic identification of toxic content in Ukrainian texts. The system combines a rule-based and a machine learning approach, both of which provide valuable, yet distinct, insights into text toxicity. The rule-based module excels in providing a detailed linguistic analysis of toxic vocabulary based on a lexicographic database, making it suitable for in-depth expert analysis. On the other hand, the machine learning module provides a scalable solution for handling large volumes of text, offering an efficient and automated method of classifying toxic content.
However, a moderate correlation between the two methods, as demonstrated by the Pearson correlation coefficient, reveals some discrepancies in toxicity assessment. These differences underline the need for further refinement in both modules to improve the overall system accuracy and reliability.
Key tasks for further improvement include enhancing the machine learning model to analyze full-length texts rather than truncated segments and expanding its training dataset to increase accuracy for longer texts. Additionally, testing more advanced models such as those from the BERT family could further improve classification performance. On the rule-based side, recalibration of the toxicity index formula is required to address limitations in how text size and single occurrences of aggressive language impact the final score.
One of the most promising directions for future work is the integration of rule-based and machine learning methods into a hybrid model. Such a model would leverage the strengths of both approaches, applying rule-based analysis for in-depth, context-sensitive interpretation and machine learning for efficient large-scale classification. This hybrid approach could also enhance the system's ability to detect more subtle forms of toxicity, such as covert hate speech or contextdependent insults.
Further development of the user interface to display toxicity results more intuitively would -technical users. Visual tools such as heatmaps of toxic word distributions, graphs showing the progression of toxicity throughout a text, and interactive dashboards could make the system more accessible for use in various domains.
183. http://iti.fit.univ.kiev.ua/wp