Detection of offensive content in the Kazakh language using machine learning and deep learning approaches

Defining destructive speech

Destructive speech encompasses various forms of communication that cause harm or promote negative outcomes, particularly within digital environments. The categories of bullying, racism, nation-based extremism, and religious extremism were selected for this study because they represent the most prevalent and socially significant forms of destructive speech identified in the Kazakh online environment. Prior studies conducted by author’s research group revealed that these categories are particularly common in Kazakh-language hate speech, making them critical targets for classification. Each of which poses unique challenges and risks in the digital context:

• 1.

  Religious extremism refers to the ideology held by certain movements, groups, or individuals within religious denominations and organizations, characterized by a strict adherence to extreme interpretations of religious doctrine (Eraliev, 2022).

• 2.

  Cyberbullying is an act of using the Internet to inflict harm or fear on another person, particularly by sending them distressing or threatening messages, known as online harassment (Cambridge Dictionary, 2024).

• 3.

  National origin discrimination (national extremism) occurs when individuals are treated unfairly due to their country of origin, ethnicity, accent, or perceived ethnic background, regardless of whether these characteristics are accurate (U.S. Equal Employment Opportunity Commission, 2024).

• 4.

  Racism involves discrimination and bias directed at individuals due to their race or ethnicity (Wikipedia Contributors, 2024).

Motivation

The rapid advancement of digital technologies has significantly transformed communication patterns across the globe, offering unprecedented opportunities for interaction and information exchange. However, this digital revolution has also introduced new challenges, particularly in the realm of online destructive speech. In Kazakhstan, where the Kazakh language is a cornerstone of national identity and cultural heritage, addressing the issue of harmful online content becomes crucial for maintaining social harmony and protecting individuals from harm.

Despite the growing presence of the Kazakh language in digital spaces, there is a notable scarcity of effective tools and methodologies for detecting destructive speech in Kazakh. The unique linguistic characteristics of Kazakh, including its agglutinative structure, dialectal diversity, and substantial influence from other languages, present significant hurdles for conventional natural language processing (NLP) techniques. These challenges are compounded by the limited availability of annotated datasets and language-specific resources, which are essential for training accurate and reliable detection models.

The motivation for this article stems from the urgent need to bridge this gap and develop effective solutions for detecting destructive speech in the Kazakh language. By addressing the specific linguistic and contextual challenges, this research aims to contribute to the broader field of NLP and artificial intelligence (AI), providing tools that can enhance the safety and well-being of Kazakh-speaking online communities. This study seeks to promote responsible AI development by ensuring that detection systems are culturally sensitive and aligned with ethical standards.

Through this research, we aspire to advance the understanding of destructive speech detection in less commonly spoken languages and to offer practical solutions that can be applied in Kazakhstan and similar contexts. The ultimate goal is to create a safer digital environment where individuals can engage in meaningful dialogue without fear of harm or discrimination, thus supporting the principles of free expression while mitigating the risks associated with online destructive speech.

Cyberbullying detection using machine learning models

The application of machine learning techniques for detecting extremist content, hate speech, and cyberbullying has seen considerable advancement, particularly in high-resource languages.

A notable contribution in this direction is presented in Bolatbek & Mussiraliyeva (2023), where the authors develop semantic analysis models trained specifically on Kazakh-language data to identify extremist content. Utilizing bigrams and word embedding approaches, the study reports strong performance across multiple machine learning classifiers. By constructing and testing models on a dedicated Kazakh-language corpus, this research not only contributes a practical framework for local language content moderation but also marks a critical step forward in low-resource language adaptation.

Heidari, James & Uzuner (2021) offers a comparative assessment of several machine learning algorithms—including support vector machine, random forests, logistic regression, and neural networks—for the task of bot account detection. While these methods demonstrate high accuracy and speed, their applicability remains constrained by monolingual training data. The absence of cross-linguistic evaluation limits the transferability of the findings to languages with different syntactic and semantic characteristics.

A systematic review presented in Mansur, Omar & Tiun (2023) explores a taxonomy of hate speech detection approaches, ranging from traditional rule-based methods to deep learning architectures such as transformers. Although the review provides valuable insights into algorithmic performance and key challenges such as contextual ambiguity and evolving hate speech patterns, the evaluation is again limited to English datasets. As a result, the review overlooks the performance variance that may arise in multilingual or morphologically complex environments.

Similarly, the study in Ayo et al. (2020) compares various deep learning approaches—including recurrent neural networks (RNNs), transformers, and support vector machines (SVMs)—on hate speech classification tasks.

Recent contributions such as Yadav, Bajaj & Gupta (2021) further investigate the use of neural networks, hybrid models, and contextual embeddings in detecting offensive and hateful speech. These studies highlight technical challenges including imbalanced datasets and ambiguous terms. However, their models are trained exclusively on English-language content, raising concerns about the effectiveness of such systems in other linguistic contexts.

An innovative approach explored in Aljero & Dimililer (2021) employs genetic programming to automatically generate classifiers for hate speech detection. The method is praised for its adaptability and accuracy, but its evaluation also remains confined to English data, leaving questions about its cross-lingual generalization potential.

A critical dimension in the development of hate speech recognizers (HSRs) lies in addressing unintended bias and ethical concerns in automated content moderation. In this context, a recent study introduced KERM-HATE, a syntax-based HSR framework designed to reduce prejudice effects often observed in state-of-the-art models. KERM-HATE leverages syntax heat parse trees as post-hoc explanations to enhance model interpretability and emphasize syntactic patterns over potentially biased semantic cues. The system outperformed established architectures such as BERT, RoBERTa, and XLNet on standard benchmark datasets, demonstrating superior classification performance (Mastromattei et al., 2022).

Racism detection using machine learning models

The study in Vidgen & Yasseri (2020) explores methods for detecting both mild and explicit forms of Islamophobic hate speech on social media. Using natural language processing and machine learning techniques, the authors develop models capable of distinguishing varying levels of hostility. The importance of contextual interpretation is emphasized, with deep learning approaches enhancing detection accuracy.

National origin extremism detection using machine learning models

Previous work by the authors of this article has significantly contributed to the field of extremism detection, with several publications addressing various aspects of identifying extremist content in Kazakh language (Mussiraliyeva et al., 2023, 2021; Zhenisbekovna, Aslanbekkyzy & Bolatkyzy, 2024; Bolatbek et al., 2024).

In the majority of reviewed studies, model development and evaluation rely heavily on English-language datasets. Although multilingualism is often acknowledged at a theoretical level, practical adaptation and cross-linguistic validation of models are rarely implemented. Preprocessing and modeling techniques tailored to the morphological characteristics of languages like Kazakh are still scarce.

Our article seeks to address this gap. Multiple machine learning models were trained and evaluated, resulting in the first comparative analysis of destructive content detection methods specifically tailored to Kazakh. This research highlights the necessity and feasibility of language-specific approaches and represents an important step toward the development of inclusive NLP systems.

Table 1 presents a comparative analysis of recent studies employing large language models and traditional machine learning (ML) methods for hate speech detection. The studies span a range of languages and platforms—including English, Kazakh, Arabic, Urdu, and Hindi—and utilize widely adopted models such as BERT, RoBERTa, GPT-3, LSTM, support vector machines (SVM), decision trees, and genetic programming.

Although high accuracy and F1-scores have been reported, persistent challenges remain in the areas of language adaptation, contextual understanding, fairness, and interpretability—especially for low-resource languages such as Kazakh and Arabic. This review clearly underscores the need for adaptive, interpretable, and equitable systems to address hate speech effectively in a multilingual digital landscape.

Data collection

In this study, authors used their own previously collected and published dataset of Kazakh-language texts, focusing on destructive speech. The dataset was compiled using public Application Programming Interfaces (APIs) from platforms such as YouTube, Telegram, and VK (Bolatbek et al., 2024). To extract relevant destructive speech content, a keyword-based crawling strategy was implemented. The list of keywords and phrases was systematically developed by analyzing a preliminary set of open-source destructive messages in the Kazakh language. This initial analysis helped identify common linguistic patterns, slurs, and terminology associated with cyberbullying, racism, religious extremism, and nationalistic extremism. Examples of selected keywords include “қорлау” (insult), “дискриминация” (discrimination), “радикализм” (radicalism), “ұлтшылдық” (nationalism). Additionally, typical hashtags and phrases frequently appearing in hate speech and extremist discourse were incorporated to enhance data retrieval effectiveness.

After filtering and cleaning over 100,000 raw entries, the final dataset included 10,190 annotated texts across five categories: cyberbullying (2,136 texts), racism (2,205 texts), religious extremism (2,110 texts), national extremism (2,076 texts), and neutral content (1,663 texts). This dataset was originally introduced at Table 2. Figure 2 demonstrates distribution of text length for corpora texts.

The annotation was performed manually by five trained native Kazakh speakers. All annotators held academic degrees in computer linguistics, computer science. Before beginning the annotation, a detailed set of guidelines was provided. These guidelines clearly defined each category (cyberbullying, racism, religious extremism, national extremism, neutral content), with examples for each class to ensure consistency. The annotation process followed these steps:

• 1.

  Annotators independently labeled each text according to the provided category definitions.

• 2.

  Regular cross-check sessions were held to discuss ambiguous cases and resolve disagreements.

• 3.

  Annotated data was reviewed collectively to ensure consistency across annotators.

To measure annotation reliability, Cohen’s kappa coefficient was calculated between annotators. The resulting agreement score was 0.82, indicating a substantial level of agreement and confirming the consistency of the annotations.

The dataset collection and annotation processes strictly adhered to ethical standards. Only publicly available posts and comments were used, and no personal user data or sensitive private information was collected. Annotators were instructed to avoid any personally identifiable information. Given the use of publicly accessible data, and the focus on textual content rather than individuals, formal ethical approval was not required according to the applicable research guidelines. Nevertheless, all procedures were conducted in alignment with responsible research practices to ensure respect for user privacy.

Data preprocessing

The collected text data underwent several preprocessing steps to prepare it for analysis:

• Text Cleaning. Removal of irrelevant information, such as HTML tags, special characters, and excessive whitespace.

• Tokenization. Splitting text into words or subwords using a tokenization algorithm tailored for the Kazakh language.

• Normalization. Converting text to a consistent format, including lowercasing and handling of diacritics.

• Stopword Removal. Eliminating common words that do not contribute to the meaning of the text.

• Stemming. Stemming was applied to reduce morphological variations in the Kazakh language, facilitating better matching of word forms while avoiding the complexity and potential data sparsity issues introduced by full lemmatization. More advanced text representation techniques, such as word embeddings (e.g., Word2Vec, FastText), typically require substantially larger datasets to achieve optimal performance. Given the size of the current dataset and the focus on interpretability and computational efficiency, stemming was deemed the most appropriate choice for the traditional machine learning baseline models.

• The dataset was split into training (70%), validation (15%), and testing (15%) sets using stratified random sampling to ensure class balance across splits.

These preprocessing steps were selected to enhance the quality of textual features, reduce noise in the data, and ensure that linguistic patterns specific to the Kazakh language could be effectively captured. They provide a balanced trade-off between preserving semantic meaning and maintaining computational efficiency necessary for building robust machine learning models.

Feature extraction and machine learning models

Feature selection is a crucial step in developing effective models for detecting destructive speech. In this study, we employed two primary techniques for feature extraction: TF-IDF and n-grams. This choice was based on extensive prior research demonstrating that TF-IDF is effective in capturing important lexical features for text classification tasks, especially in low-resource settings where deep learning models may underperform due to limited data availability Aljwari et al. (2022).

TF-IDF is a statistical measure used to evaluate the importance of a word in a document relative to a collection of documents (corpus). It combines two factors: term frequency (TF), which measures how often a word appears in a document, and inverse document frequency (IDF), which assesses how unique or rare a word is across the entire corpus.

Term frequency (TF): measures how often a word appears in a document.

(1) $$TF(w,d)=\frac{\mathrm{N}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\mathrm{s}\mathrm{w}\mathrm{o}\mathrm{r}\mathrm{d}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{o}\mathrm{c}\mathrm{u}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}d}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{w}\mathrm{o}\mathrm{r}\mathrm{d}\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{o}\mathrm{c}\mathrm{u}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}d}$$where $w$ is the word, and $d$ is the document.

Inverse document frequency (IDF): measures how important a word is across the entire corpus.

(2) $$IDF(w,D)=log\left(\frac{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{d}\mathrm{o}\mathrm{c}\mathrm{u}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{p}\mathrm{u}\mathrm{s}D}{\mathrm{N}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{d}\mathrm{o}\mathrm{c}\mathrm{u}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{a}\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{w}\mathrm{o}\mathrm{r}\mathrm{d}w}\right)$$where $w$ is the word, and D is the corpus of documents.

TF-IDF: Combines TF and IDF to give a balanced measure of a word’s importance in a document relative to the corpus.

(3) $$TF\text{-}IDF(w,d,D)=TF(w,d)\times IDF(w,D).$$We applied TF-IDF to transform the text data into a numerical format that reflects the significance of each term within the context of destructive speech detection. This method helps to highlight terms that are more relevant to the classification task while reducing the impact of common words that occur frequently across documents.

N-grams are contiguous sequences of $n$ words (or tokens) extracted from the text. They capture contextual relationships and patterns within the text that single words alone may not convey. For instance, bigrams (2-grams) and trigrams (3-grams) help to understand phrases or combinations of words that frequently occur together. We generated $n$-grams from the preprocessed text data, including both bigrams and trigrams, to capture more nuanced patterns in destructive speech. By incorporating these multi-word features, the model can better recognize context-specific expressions of harmful content.

(4) $$n\text{-}gram=w_i,w_{(i+1)},w_{(i+2)},\dots ,w_{(i+n-1)}$$where $i$ ranges from $1$ to $T-n+1.$

Feature Selection Process. The preprocessed text data was transformed into numerical feature vectors using TF-IDF. Simultaneously, $n$-grams were extracted and included as additional features. The final feature vectors for each document were constructed by combining TF-IDF scores with $n$-gram frequencies. This comprehensive feature set ensures that both individual term importance and contextual patterns are considered in the classification models.

The integration of TF-IDF and n-grams into the feature selection process enhances the model’s ability to accurately identify and classify destructive speech in Kazakh. By capturing both term significance and contextual patterns, these features contribute to the overall effectiveness of the detection system.

While our study emphasizes the importance of language-specific modeling for low-resource languages like Kazakh, we chose to use mBERT over KazBERT due to several critical factors relating to data coverage, generalization capability, and implementation stability. KazBERT, although developed specifically for Kazakh, was trained on a relatively small and localized dataset comprising only 5,904 Kazakh Wikipedia texts, along with limited English and Russian corpora (Amandyk, 2023). This narrow scope reduces the robustness and generalizability of the resulting language representations, especially for tasks that require diverse contextual understanding beyond formal Wikipedia-style language. Moreover, the model’s trilingual design (Kazakh, Russian, English) introduces a potential risk of overfitting to the styles of those domains, making it less suitable for broad applications.

In contrast, mBERT was pretrained on the Wikipedias of 104 languages and includes over 200,000 Kazakh Wikipedia entries at the time of training. While it may not specialize in Kazakh, its broad multilingual context supports better cross-lingual transfer and generalization. Studies such as Wu & Dredze (2020) suggest that multilingual models benefit low-resource languages through shared representations across typologically similar languages. Furthermore, mBERT offers a stable, well-integrated pipeline within the Hugging Face ecosystem, whereas KazBERT’s documentation and support are still limited. Ultimately, the choice of mBERT was driven by practical considerations regarding model maturity, infrastructure support, and the broader linguistic diversity captured during pretraining, which compensates, to an extent, for its Kazakh-specific limitations. Future work could include a controlled empirical comparison between mBERT and KazBERT on downstream Kazakh NLP tasks.

In this study, we adopt two complementary approaches to destructive speech detection: (1) classical machine learning using feature-based methods and (2) a deep learning architecture combining BERT embeddings with an LSTM layer. This dual strategy enables a comprehensive evaluation of both traditional and modern techniques for Kazakh-language content classification (Table 3).

The machine learning models selected for this study—Naive Bayes, logistic regression, random forest and others—are widely recognized as strong baseline algorithms in text classification tasks, particularly in the domains of hate speech detection, cyberbullying identification, and toxic content moderation (Schmidt & Wiegand, 2017; Fortuna & Nunes, 2018; Badjatiya et al., 2017). Naive Bayes is known for its effectiveness in handling high-dimensional sparse feature spaces generated by TF-IDF representations. Logistic regression has consistently demonstrated competitive performance in binary and multi-class text classification scenarios, offering a balance between simplicity and accuracy. random forest provides robustness through ensemble learning, capturing complex feature interactions without overfitting easily.

For the deep learning approach, the combination of BERT embeddings with an LSTM layer was motivated by the need to model both contextual semantics and sequential dependencies, particularly important for the morphologically rich Kazakh language. BERT-based models have achieved state-of-the-art results in various NLP tasks (Devlin et al., 2019), while LSTM networks are effective in capturing long-range dependencies in sequential data (Hochreiter & Schmidhuber, 1997).

The integration of an LSTM layer following BERT embeddings in our model is motivated by prior research demonstrating the effectiveness of hybrid transformer-RNN architectures for capturing sequential dependencies in morphologically complex languages. While BERT models capture contextualized representations through self-attention, they are limited in modeling fine-grained morphological variations and long-distance sequential dependencies, which are critical in agglutinative languages such as Kazakh.

The article of Gethsia, Juliet & Anitha (2024) presents a hybrid model that combines BERT for contextual embeddings with Bidirectional LSTM (BiLSTM) for sequential modeling to enhance emotion recognition in text. This approach captures both the deep contextual meaning of words and their order in a sentence. The model outperforms standalone BERT and BiLSTM architectures on benchmark datasets. It shows strong potential for applications like sentiment analysis and emotional tone detection in social media and customer feedback.

The study explores the use of BERT-based embeddings combined with hybrid neural network architectures for Indonesian sentiment analysis. It highlights that BERT significantly improves performance by capturing contextual nuances in the language. Among the tested models, the BERT-based LSTM-CNN achieved the highest accuracy. The research demonstrates that combining BERT with hybrid deep learning models is highly effective for sentiment classification in low-resource languages like Indonesian (Murfi et al., 2024).

In the article “Using BERT and LSTM to Do Text Classification”, Lu (2022) addresses the challenge of accurately interpreting multiword expressions (MWEs) in natural language processing. Recognizing that pre-trained models like BERT may struggle with the non-compositional meanings of MWEs, Lu proposes enhancing BERT with an additional LSTM layer to better capture sequential dependencies. This hybrid BERT-LSTM model demonstrated improved performance in text classification tasks, outperforming both standalone BERT and other hybrid configurations such as BERT with TextCNN. The study underscores the effectiveness of integrating LSTM with BERT to enhance the understanding of complex linguistic structures in text classification?.

In our context, the LSTM layer serves to further refine the contextual sequence representations generated by BERT, particularly for Kazakh’s rich case system, verb morphology, and word order variability.

To evaluate the impact of different textual representations on model performance, we systematically experimented with multiple n-gram configurations across both machine learning and deep learning models. Specifically, we used three n-gram ranges: $(i)$ $N=1,2,$ which includes both unigrams and bigrams to capture single-word tokens and adjacent word pairs; $(ii)$ $N=2,2,$ which uses only bigrams to emphasize local word pair dependencies; and $(iii)$ $N=2,3,$ which combines bigrams and trigrams to represent richer multi-word contextual patterns. Each configuration was tested under two conditions: with stemming and without stemming. Stemming reduces words to their morphological roots, helping to normalize inflectional variations common in Kazakh and thereby reducing feature sparsity. In contrast, the no-stemming condition preserves the original surface forms of words, allowing the models to capture exact lexical usage patterns. These combined configurations were designed to assess the sensitivity of various classifiers to different levels of lexical and contextual abstraction in Kazakh texts. As shown in Tables 4–9, the $N=1,2$ with stemming setting consistently yielded the best overall results across multiple evaluation metrics, indicating that the combination of lexical normalization and local context features is particularly effective for destructive speech classification in morphologically rich languages such as Kazakh.

Deep learning pipeline: BERT+LSTM hybrid model

For the deep learning pipeline, we designed a hybrid architecture combining pre-trained BERT embeddings with a sequential modeling component based on an LSTM layer. The motivation for this design stems from the observation that while BERT embeddings capture rich contextual information at the token level, they may not fully model the sequential structure of texts, especially for languages like Kazakh, which exhibit complex morphological patterns.

In the proposed architecture, the input texts are first transformed into contextual embeddings using a pre-trained BERT model. These embeddings are then passed through a unidirectional LSTM layer to model sequential dependencies between words. The output of the LSTM layer is subsequently fed into dense layers with ReLU activations, followed by a softmax output layer for multi-class classification. These layers were chosen to effectively capture both the contextual and sequential characteristics of the Kazakh language. BERT embeddings provide deep contextual understanding, LSTM models sequential dependencies that transformers alone may overlook, and the dense layer maps the learned representations to the target classes for accurate classification. The model was trained using the Adam optimizer with a learning rate of $1e-4$, and early stopping was applied based on validation loss to prevent overfitting.

LSTM layers

The contextualized embeddings from BERT are then fed into a long short-term memory (LSTM) network (Fig. 3). LSTM is a type of recurrent neural network (RNN) that is well-suited for sequence data, such as text, as it can capture long-term dependencies within the sequence.

LSTM helps to further refine the features learned by BERT by modeling the sequential nature of the text. In this architecture:

• LSTM layers $N\times 128\times 256$: the input to the LSTM is the sequence of embeddings from BERT, and the LSTM transforms this into an output of size $N\times 128\times 256$. Here, 128 refers to the hidden layer size, and 256 refers to the output dimension of the LSTM.

• The LSTM layer processes the data in time steps, sequentially processing the embeddings from BERT and allowing the model to capture time-dependencies (or sequential information) that may exist in the text.

The LSTM layer further enhances the model’s ability to capture dependencies in text sequences, which can help in tasks where understanding the order of words or sentences is crucial, such as in distinguishing between neutral and extremist content.

An LSTM layer was used after the BERT embeddings to enhance the model’s ability to capture sequential dependencies within Kazakh texts. Although transformer models like BERT are highly effective in modeling token-level contextual relationships, they may not fully preserve longer sequential patterns, especially in languages with rich morphology and flexible syntax like Kazakh. LSTM networks are well-suited for learning long-term dependencies and sequential structures, making them an appropriate choice for improving classification accuracy in this setting. Alternative models such as GRU or transformer encoders were considered; however, LSTM was selected due to its proven effectiveness in capturing complex sequence patterns with relatively lower computational overhead.

Dense layer

After passing through the LSTM layer, the output is forwarded to a dense (neural) layer. The role of this layer is to perform the final classification task.

• $N\times 256$ (class count): the dense layer has an output dimension of 256, which corresponds to the number of possible output classes. In the case of Kazakh text classification, the class count could represent categories such as neutral, propaganda, recruitment, radicalization, etc.

• This layer performs the classification using the features extracted by both the BERT and LSTM layers.

The architecture presented in Fig. 3 illustrates the operational workflow of our BERT+LSTM model. In this model, contextual embeddings generated by the pretrained BERT layer are passed into the LSTM network, which captures sequential dependencies within the Kazakh text. The output of the LSTM is then forwarded to a final classification (dense) layer. This architecture was specifically applied in the experiments for destructive content classification shown in Fig. 4. In contrast, the results presented in Tables 4–9, described in “Experimental Results”, were obtained using traditional machine learning (ML) and standalone deep learning (DL) models, and therefore are not based on the BERT+LSTM architecture shown in Fig. 3.

Evaluation metrics

To assess the performance of our model for detecting destructive speech in Kazakh, we utilized several key evaluation metrics. These metrics provide a comprehensive view of the model’s effectiveness in classification tasks. The primary metrics used are:

accuracy measures the proportion of correctly classified instances out of the total number of instances. It provides an overall sense of how well the model is performing.

(7) $$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}.$$

• TP (True Positives)—the number of correctly predicted positive instances;

• TN (True Negatives)—number of correctly predicted negative instances;

• FP (False Positives)—the number of falsely predicted positive instances;

• FN (False Negatives)—number of wrongly predicted negative instances.

Precision evaluates the proportion of true positive predictions relative to the total number of positive predictions made by the model. It indicates how many of the identified positive instances are actually positive.

(8) $$Precision=\frac{TP}{TP+FP}.$$Recall (or Sensitivity) measures the proportion of true positive predictions relative to the total number of actual positive instances. It indicates how well the model identifies all relevant instances.

(9) $$Recall=\frac{TP}{TP+FN}$$

The $F1$ Score is the harmonic mean of precision and recall, providing a single metric that balances both aspects. It is particularly useful when dealing with imbalanced datasets where both false positives and false negatives are critical.

(10) $$F1\text{-}score=2\times \frac{Precision\times Recall}{Recall+Precision}.$$

Area under the receiver operating characteristic curve (AUC-ROC) measures the ability of the model to distinguish between classes. The ROC curve plots the true positive rate against the false positive rate at various thresholds, and AUC represents the area under this curve. A higher AUC indicates better model performance. An AUC value closer to 1 signifies excellent performance, while a value closer to 0.5 indicates performance equivalent to random guessing.

The confusion matrix provides a detailed breakdown of the model’s predictions, showing the counts of true positives, true negatives, false positives, and false negatives. It helps in understanding the types of errors the model is making.

These metrics collectively offer insights into the model’s performance in detecting various types of destructive speech, including its accuracy, precision, recall, and ability to handle imbalanced classes. Evaluating the model using these metrics ensures a comprehensive assessment of its effectiveness and helps in fine-tuning and optimizing its performance.