Leveraging deep learning for toxic comment detection in cursive languages

Novel contributions and innovations

The present study presents a significant advancement and a novel approach to cursive language text categorization and characteristic extraction. Our work goes beyond conventional methods and tackles the unique challenges these languages pose. We specifically focus on Urdu, but our approach is applicable to other cursive languages such as Arabic and Persian. Initially, we present an innovative method for extracting features from cursive language strings that employs unsupervised learning techniques. The methodology used in this study involves the implementation of bidirectional encoder representations from transformers (BERT) for several essential tasks, such as tokenization and the identification of lemma words. Given the complexity of cursive scripts and their inherent linguistic nuances, this represents a significant advancement over traditional methods.

Secondly, we pioneered the application of memorization algorithms for text categorization in cursive languages. This approach provides an effective solution for dealing with the high dimensionality and complexity of cursive language data.

Data collection procedure

The dataset employed in this study was systematically collected using a methodical and structured approach to ensure an inclusive and representative sample of both non-toxic and toxic remarks. The initial step involved developing a Google form to facilitate and organize the submission of comments from participants.

This document was widely circulated among the undergraduate population of the university. The intention was to elicit multiple responses from each participant, spanning a spectrum of emotions from impartial to detrimental. This methodology facilitated the collection of a diverse array of remarks indicative of the wide range of human discourse.

Following the collection of form submissions, a meticulous review process was instituted to determine the genuineness and relevance of the data. A group of annotators systematically reviewed each submission, carefully analyzing the comments to assess their accuracy and relevance. A dual validation process was implemented to ensure that the dataset used in the study was not only voluminous but also relevant and of high quality, thus aligning with the research objectives.

After the conclusion of the data collection and validation procedures, the data was further refined through a sequence of preprocessing measures. These procedures included the normalization of textual data, the elimination of extraneous components, and the classification of the remarks. The compiled CSV file of the final dataset is a robust and comprehensive collection of remarks, ready for thorough examination and investigation.

The meticulous and methodical approach to collecting and organizing data not only yielded a comprehensive dataset for our current study but also established a standard for future research in this field. It highlights the importance of rigorous data collection and verification procedures.

Dataset pre-processing

The manual dataset curation necessitates a rigorous preprocessing stage. This involves a series of operations to clean and prepare the data for subsequent statistical analysis, as described in Fig. 3. Given the diverse nature of raw text data, cleaning involves several crucial steps. These steps ensure that irrelevant content does not significantly influence the results of language processing and classification tasks. The following are the steps taken for the preprocessing of the dataset.

• Character normalization: Accented characters, incorrectly typed characters, and dialects in Urdu or similar languages within the comments that do not contribute to the meaning were normalized. This includes normalizing similar characters with different ASCII codes for uniformity.

• Tag and metadata removal: Textual data may contain various HTML tags, hashtags, and other metadata that do not contribute semantically to the comment’s content. These were programmatically removed during the cleaning process and further verified manually.

• Emoji and symbol removal: In modern digital communication, emojis and symbols are prevalent. However, these did not contribute to the textual data processing and were removed from the dataset.

• Punctuation and special character handling: Given that Urdu is written in the UTF-8 encoding scheme, which differs from ANSI, certain punctuation marks and special characters were converted into their encoding numeric value during data collection. For instance, double quotation marks (”) were replaced with their corresponding HTML code.

Upon completion of the initial cleanup phase, a comprehensive list of words was generated from all comments. This list was meticulously reviewed once again to remove any unnecessary residual content. Following this procedure, the UTOXS21 dataset was effectively cleaned of all extraneous elements, producing a high-quality, noise-free resource for subsequent analysis.

Dataset description

To sum up, our dataset UTOXS21 consisted of 17,078 short-length comments, of which 51.6% were nontoxic comments, and the remaining were manually marked as toxic comments. Characteristics of the dataset are shown in Table 2. The dataset consists of 17k Urdu language short text labeled comments (both toxic and nontoxic) on topics such as politics, sectarianism, identity, hate, and obscenity.

These selected comments conveyed either some toxicity or were nontoxic. To extract valuable information from the collected text, a month-long annotation process was performed by native Urdu experts. Two experts from the Urdu-speaking community reviewed the labels of each comment to ensure accuracy. Some comments were even hard to recognize by humans, so a third opinion was used to finalize the label.

The dataset distribution, with approximately 51.6% classified as nontoxic and the remaining comments identified as toxic, as shown in Fig. 4. The comments cover various subjects, such as politics, sectarianism, identity, hate, and obscenity. Native Urdu-speaking experts thoroughly reviewed and labeled each comment, which took a month to complete. To ensure accuracy, two experts independently reviewed each comment’s label, with a third opinion considered for any ambiguous comments. The following are the major distribution details of the dataset instances.

Dataset characteristics:

• Number of instances: 17,078 (8,259 of which were classified as toxic)

• Number of attributes: String attribute (vector of words) and the class

• Class distribution: 48.4% toxic and 51.6% nontoxic

Specific linguistic considerations for Urdu

In contrast to short texts, analyzing grammar is easier in longer texts. Longer texts provide better meaning. Short texts often use informal vocabulary with spelling and grammatical errors. This is because proper grammar analysis methods work poorly in short texts. Many words in Urdu can have different meanings. In this situation, it becomes important to know the context of the solutions. Some of these examples are given in Fig. 5. Urdu is written using a derivation of the Persian script based on the Arabic alphabet. The following are some of the critical challenges in the Urdu language. There are no small and capital characters in the Urdu alphabet. Therefore, it is difficult to recognize the start of sentences and nouns. This makes it difficult to tag parts of speech. Urdu writing is bidirectional, with alphabets written from right to left and numbers written from left to right. It is a cursive language, and the shape of the character is context-dependent. This means that the shape of a character changes because of its neighboring character.

Employing N-gram tokenization in cursive languages

The act of segmenting sentences into discrete units of language, specifically in the context of cursive scripts such as Urdu, poses distinctive obstacles that are not commonly observed in other linguistic systems. The Urdu language frequently constructs words consisting of two or three letters, which may or may not be delimited by interstitial gaps. Conventional tokenization techniques, which typically utilize spaces or other forms of punctuation for word segmentation, are insufficient for the precise disintegration of Urdu sentences into their constituent words.

To address this challenge, we have implemented a tokenization method based on n-gram creation, as shown in Fig. 6. The methodology employed involves the segmentation of the input text into consecutive groupings of ‘n’ elements. The nature of the element can vary depending on the level of scrutiny and may encompass characters, syllables, or words. The utilization of n-grams allows us to capture the local structure of the language and overcome the space insertion and omission problem inherent in cursive languages.

In detail, we first generate potential tokens by creating n-grams of varying lengths. Character n-grams ranging in length from 2 to 10 grams are created. The most extended sequence is placed at the start of the character gram list, while the shortest sequence is at the end. The term frequency of the character n-grams is also counted and ordered from the most frequent to the least frequent sequence. Extremely rare tokens are eliminated from consideration. The thresholds were determined by analyzing the n-gram distribution of the dataset, the diversity of the token, and the variation in the length of the token. Unigrams, bigrams, and trigrams were considered with frequency thresholds set at [7, 6, 5, 4, 3] to select the most relevant tokens. This selection process involved filtering n-grams by frequency and evaluating the performance of our tokenizer against other tokenization techniques. The optimal frequency threshold, which balances precision and recall, was identified through this iterative process. In addition, n-gram probabilities and confidence scores were integrated to prioritize tokens that were more likely to occur in the text, further refining tokenization. This approach allowed us to handle the complexity of the Urdu language effectively, enhancing the tokenizer’s performance in processing Urdu text for various NLP tasks.

This approach is especially advantageous when dealing with unstructured text that does not adhere to standard grammar and language rules, such as offensive sentences. In these cases, the ability of the n-gram approach to recognize words based on local structure, rather than relying on spaces or punctuation, makes it highly effective.

Furthermore, the tokens generated through this method serve as the basis for subsequent analysis. They are used to compute the distances between different sentences, a key factor in assessing document similarity. By accurately identifying the constituent words in each sentence, we can effectively measure the semantic distance between them, enabling a more accurate document comparison. N-gram tokenization provides a robust and effective method for handling the unique challenges posed by cursive languages. It enables accurate tokenization even in the absence of consistent use of spaces, facilitating more precise language analysis and document comparison.

Comments similarity with vector space models

Textual data can be considered as a sequence of tokens, where each token can represent a character, a word, or any other linguistic unit. In many text processing tasks, a text document is represented as a Bag-of-Words (BoW) model, where each distinct word corresponds to a unique feature in the feature space. The BoW model represents each document as a vector embedded in a high-dimensional space, with each dimension corresponding to a unique word in the document.

For example, consider the Roman Urdu sentences “yeh mausam khubsurat hai”, “mausam bahut khubsurat hai aaj”, and “aaj ka mausam bahut acha hai”. In these sentences, each word is vectorized independently, and the BoW representation of each sentence is obtained by summing the vector representations of its constituent words. In real-world problems, the dimensionality of the feature space can be on the order of hundreds of thousands, given the rich vocabulary of natural languages.

Figure 7 provides a graphical representation of the vectorized words and their sum in the BoW model. It shows how each word is represented in a separate vector with padded zeros and how these vectors are summed to yield the BoW representation of the sentence.

The similarity of the comments can be measured as the distance or angle between the vector representations of the comments. One common measure of document similarity is cosine similarity, which is calculated as the cosine of the angle between two vectors. Given two vectors a → and b →, the cosine similarity between them is computed as shown in Eq. (1): (1)cosine similarity a → , b → = a → ⋅ b → | a → | 2 | b → | 2

where a → ⋅ b → denotes the dot product of a → and b →, and | a → | and | b → | are the magnitudes of a → and b →, respectively.

In our example, Fig. 7 shows the cosine similarities between the BoW representations of the sentences. For instance, the cosine similarity between “yeh mausam khubsurat hai” and “mausam bahut khubsurat hai aaj” is 0.85, indicating a high degree of similarity between these sentences.

The TF given in the Eq. (2) is a ratio of the occurrences of a term in a d and the number of instances of the most recurring word within the comment d. (2)t f t , d = f d t m a x ω ∈ t f d w .

The term idf represents the transposed count of the comments terms; see Eq. (3). The lower frequency of terms relative to documents increases the factor. This would severely prejudice the ranking of the praise of infrequent terms, although the tf factor has a high value. It is exceptional that the term t is off with a high tf value in document d but rarely used anywhere else. (3)i d f t , D = l n D d ∈ D : t ∈ d .

The Inverse Document Frequency (IDF) is a metric that gauges the significance of a given term ‘t’ in a corpus of documents ‘D’, as expressed in Eq. (3). The expression denotes the logarithm with base ‘e’ of the ratio between the total number of documents and the number of documents that contain the specific term ‘t’. The purpose of this design is to assign greater importance to terms that have a lower frequency across all documents, as they are presumed to possess greater informational value. (4)t f i d f t , d , D = t f t , d . i d f t , D . The TF-IDF score, as stipulated in Eq. (4), is computed by multiplying the term frequency (TF) with the inverse document frequency (IDF). The approach aims to achieve a harmonious equilibrium between the intrinsic significance of a given term within a specific document (i.e., its frequency) and its extrinsic significance across all documents (i.e., its rarity). The aforementioned metric assigns a significant weight to terms that exhibit a high frequency within a given comment, yet are comparatively infrequent across all comments, thereby suggesting that such words may hold particular significance for the comment in question. (5)t f i d f ′ t , d , D = i d f t , D D + t f i d f t , d , D .

The modified version of the TF-IDF score is represented by Eq. (5). The TF-IDF score is increased by the IDF quotient and the aggregate count of documents. The aforementioned equation underscores the significance of uncommon vocabulary in all written materials, thereby minimizing the likelihood of inaccuracies and augmenting the distinguishing capability of less frequent expressions. The aforementioned computations constitute a pivotal cornerstone in the categorization of noxious remarks, furnishing a quantitative framework for machine learning algorithms to detect and mark harmful material.

Both the TF and IDF factors influence favorably for high relevance. One uses global information and the other uses local information. So, the product of these terms sees that Eqs. (4) and (5) have the better possibility of reflecting co-relevance. To further reduce the possibility of error.

Techniques for normalizing variable-length comments

The deep learning framework requires vector representations of the data. In this case, the input length is variable, and this is required to pad the data such that each vector has the same length. These same length vectors allow you to process the data according to your algorithms. In this section, a list of the sequences of each document is created.

Train test split

The dataset is divided into training and test datasets using a fine-tuned train-test split ratio. Some data is required for the validation of the trained model to test the variance between the predicted and observed values. This split ratio is obtained after conducting batch training and testing with train data sizes ranging from 55% to 95%. The best accuracy is achieved when 70% of the dataset is reserved for the training set and the remaining 30% for the test dataset. A sequential model is trained using the training data. The test data are further quantitatively evaluated for their accuracy.

Feature selection

To measure the differences between the expected frequencies and the observed frequencies for two events, the Chi-square test is used in our proposed experimentation. It is a statistical technique. In feature selection, the two events are classified as the occurrence of the term and the occurrence of the class. In Urdu sentiment analysis, many studies have explored the consequences of applying the Chi-square feature selection method, such as the value of each term with respect to the value of the class being measured. In our experiment, the number of features was reduced from 6,243 to the selected number of features used for different models.

Word2Vec embedding

Word2Vec is a neural network that processes a corpus. The text documents are input into Word2Vec, producing a vector for each word. This network is a two-layer neural network. Word2Vec produces a feature vector that is a numerical form for each word. The purpose of this feature vector is to group together word vectors that have similarities in vector space. With the help of word embedding, similarities can be identified between words using the vector of each word. Word2Vec vectors represent the feature space in the context of individual words with other words. By providing data and calculating the embedding of words, the Word2Vec model can calculate the occurrence of a particular word in context with other words. The output of this two-layered neural network is a vocabulary set with a vector attached to it. This vocabulary vector can then be passed to the deep neural network to detect relationships between the queried words.

Here is an example window with the complexity of computing P(w_t+2|w_t).

For each position t = 1...T, context is predicted for each word within the fixed size window m given center word w_j is shown in Eq. (6). (6)L i k e l i h o o d = L θ = ∏ t = 1 T ∏ − m ≤ j ≤ m t P w t + j | w t ; θ . Let θ be all the variables to optimize and let the objective function J(θ) is the average negative log.

It is a long, large corpus that contains T words. This corpus includes many documents, all of which are concatenated to produce a long list of words. In Eq. (7), the first ∏ iterates through all words, and then the second product chooses a fixed size window, m, to predict a word in the context of the words around it. This means the word can be predicted by a center word. All values are multiplied by 2. This is how the model predicts the likelihood. The likelihood depends on the parameters of the model, which are represented by θ. In this model, the only parameter is the vector representation given to words. The model has no other parameters to work with. A word is represented by a vector in the vector space, and that representation is its meaning. Other words are predicted based on these vectors. To minimize the cost, we need to define an objective function. The objective function is the same as given in Eq. (7), except that it has a minus sign, indicating that it is a minimization problem. The one over T means that we are averaging it. The important thing is that we include a log in Eq. (7) because it turns out that everything is in the log. We now add the log of all probabilities. (7)J θ = − 1 T log L θ = − 1 T ∑ t = 1 T ∑ − m ≤ j ≤ m t log P P t + j | w t ; θ .

Everything is predicated on having this probability function method to predict the probabilities of the words in the context of the given centered word. This model can have only a vector representation of the words and that is the only parameter of the words. To make it simple, every word has two vector representations for each word in the neighborhood context as shown in Fig. 9. So, there is one vector of words where it is centered, and another vector for each word when it is in context. This probability of words can be worked for words and context words, a center word purely in terms of these vectors, as shown in Eq. (8). (8)P O | C = exp μ o T ν c ∑ i = 1 w ɛ v n exp μ w T ν c . The dot product compares the similarity of μ and c.

u^Tμ = u.v = ∑u_iv_i The larger dot product means the greater similarity. The fractional part of Eq. (8) is a normalization factor across the entire vocabulary to produce a probability distribution.

Mapping a vector of words to other context words is an exam of softmax, i.e., ℝⁿ → ℝⁿ.

Softmax distribution, so the two parts of expanding and normalizing give you a softmax distribution.

The softmax function will start any numbers in a probability distribution for two reasons. So, it refers to softmax because it works like softmax. If you have numbers, you could just say what the max of these numbers is. It is called hard max if you map the original numbers to the max and everything else is set to zero; it is considered softmax because of the exponential if you just ignore the problem of negative numbers for a moment. When you use an exponent, the bigger numbers will become bigger for a moment and get within the x refer to Eq. (9). The softmax primarily places the mass with the maxes or with a couple of maxes, so it is the max part, and the soft part is that this is a difficult decision that still saves a little probability mass everywhere else. (9)s o f t m a x x i = exp x i ∑ j = 1 n exp x j = p i

So now we have a loss function with the probability model on the inside.

The word vector representation may be a good predictor.

Neural network embedding is used as word embedding. One word is described by another, even if those are entirely different words. Vocabulary vectors enable computers to read sentences by performing mathematical operations on them. These are numbers calculated based on the neighborhood, in other words. It does this in different ways based on two different methods. In one method, the target word is predicted as a Continuous Bag of Words (CBOW) or to predict a context word.

Impact of embedding and vocabulary size on model performance

The text is then tokenized using a tokenizer available in Keras. The tokenizer is used to convert text into vectors, the sequence of binaries. By default, this class’s tokenize function removes punctuation marks and converts sentences into a sequence of words encoded in UTF-8. These are then converted into a list that can be indexed.

Any sequence that is less than the maximum length is then padded with the remaining words. After that, all unique words are presented by an integer. The vocabulary size is defined as the max in this model.

Sequential modeling approach for text classification

Keras has two models available. One is sequential, and the other is functional. Data here are in the form of sequence, which are comments. CNNs are developed to use sequential data efficiently, so sequence processing is used for this problem. The sequential model allows the creation of layer-by-layer processes, which suits most problems. For this problem, this is suitable. The model is a linear stack of neural layers which can be easily described.

Hyperparameters optimization

A hyperparameter is usually used to control the learning process of the deep learning model. Changing the preset values of a deep learning model to the best values provides the best accuracy and efficiency of the deep learning model. The values of the hyperparameter depend on the attributes of the dataset. The parameters are recorded and represented in graphs over different iterations to observe the trend in the behavior of the learning system. During training, the connection between the layer and the next layer can be reduced to a subset of neurons for weight updation, known as dropout. This dropout helps to avoid overfitting the model. Dropout layers are generally used where a large number of neurons is passed to the next layers. This technique allows setting to 0 and then excluding the activation of a certain percentage of the neurons of the preceding layer. The probability that the neuron’s activation is set to 0 is indicated by the dropout ratio parameter within the layer via a number between 0 and 1. We used values ranging from 0.1 to 0.2 and found that the 0.1 dropout layer was the best.

Utilizing k-fold cross-validation for model validation

K-fold cross-validation scrutinizes the machine learning method when a dataset is medium-sized. This cross-validation method has only one parameter, k, which is the number of data samples to split repeatedly. Due to this, repeated training and testing usually prevent any bias and learning ability of the ML model on unseen data. Firstly, the model is trained without cross-validation. Finally, a 10-fold validation parameter is used after attempting a range of 5-10 folds to ensure an unbiased ML model.

Recall measures and their significance

The recall is a measure that tells what proportion of comments actually were offensive and were diagnosed by the algorithm as toxic. The total positives (TP and FN) and comments were diagnosed as neutral but actually were offensive.

The training process was repeated several times to avoid training overfitting during the parameter tuning process. The number of epochs ranged from 10 epochs to 200. The lowest accuracy was achieved when the number of epochs was 20, and this was 59 1%. However, the highest accuracy was 89. 75%, which was achieved when the number of training epochs was 200 and this is shown in Fig. 10. The accuracy and loss were 1.33, which is relatively low. The predictive accuracy is also shown in the confusion matrix in Fig. 11.

Comprehensive discussion and interpretation of results

Four iterations were performed to find the best configuration of the essential features. The figure shows the top 500 features used for all model’s evaluation matrices in the first run Table 4. In addition, SVM and DNN outperformed all other models in this configuration. CNN recall values struck due to some feature size because this is suitable for two-dimensional data.

The top 1,000 features are selected to input learning models in the second iteration. The accuracy of all models is increased at this configuration, as shown in Table 5. However, the performance of GRU remains unchanged. For example, SVM from machine learning and DNN from deep learning outperform in this configuration. The worst performing algorithms were naive Bayes and CNN for this arrangement.

In the third iteration, 1,500 features are selected. The performance of CNN further decreased in this run, as shown in Table 6. SVM and DNN are also ahead of their rival techniques. The F1 scores of naive Bayes, SVM, DNN, CNN, GRU, LSTM, and BiLSTM are 83.5946, 84.7187, 81.8936, 82.8595, 84.5572, 83.4887, and 84.0556 respectively. The GRU algorithm downgraded its performance. All other parameters remain the same for this run.

Fourth evaluation: 2,000 important features are used and selected by the CH2 technique. BI-LSTM and LSTM reduced their performance after increasing the number of features. As a result, DNN performs better on all configurations, as shown in the Table 7.

Toxic comments are short sentences that do not contain any formal structure. Some comments may even consist of one word. LSTM, BI-LSTM, and GRU lag behind because these techniques use context. The dataset used here contains grammar and context-free structures, so for text classification containing short-length comments, DNN is the best algorithm, achieving a 83.9699% precision and an 81.8936% F1 score. The second high-performing model is GRU with a precision of 85.2020% and an F1 score of 84.5572%.

The results of the toxic comments classification are presented in this section. The features are extracted and optimized during experiments, including stop-words, without less common stop-words, and without stop-words. The machine learning models naive Bayes, SVM and the deep learning models DNN, CNN, GRU, LSTM, and BLSTM are trained and tested using the count vector and the TFIDF vector.

The general comparison in terms of the accuracy performance of the aforementioned techniques is shown in Fig. 12. The worst-performing algorithms in this experimental design were CNN and GRU, which showed the least F1 scores of 82.8595% and 84.5572%, respectively. Figure 12 shows that the accuracy of the aforementioned machine learning models is relatively better when features are extracted after stop-words removal. The F1 score of DNN without removing stop-words is 81.8936%, without stop-words is 81.8936%, and removing less common stop-words is 81.8936%. Naive Bayes performed second best with an accuracy score of 83.6484% with stop-words, 83.6484% without stop-words, and 83.6484% without less common stop-words.

Additionally, BERT achieved an accuracy of 85.4508% and an F1 score of 85.4069%, indicating strong performance in the classification task as shown in Table 8. DeBERTa, while slightly lower, still performed well with an accuracy of 79.2740% and an F1 score of 79.1963%. GPT, on the other hand, showed a lower performance with an accuracy of 77.5468% and an F1 score of 77.5415%, suggesting it may not be as well-suited for this specific task compared to the other models as reflected in Table 9.

This technique is not only applicable to Urdu, Persian, and Arabic but can also be extended to other low-resource languages such as Pashto, Kurdish, and Somali. These languages share similar challenges in terms of resource availability, making our methodology particularly valuable for their text categorization and feature extraction needs.