Toxic Comment Research Articles

Advanced Social Media Toxic Comments Detection System Using AI

Abstract: In the contemporary digital landscape, social media platforms have evolved into essential components of our daily lives, fostering connections, idea sharing, and meaningful conversations. Nevertheless, the escalating volume of online interactions has ushered in a concerning rise in toxic comments and cyberbullying, casting a shadow over the potential for a healthy onlineenvironment. Toxic comments, spanning hate speech, harassment, and offensive content, not only inflict harm on individuals but also detrimentally affect the overall user experience. This project seeks to address this pressing issue through the development of a cutting-edge Toxic Comment Detection system, leveraging the power of Natural Language Processing (NLP) and Machine Learning (ML) techniques. The primary objective of this endeavor is to create an automated system capable of identifying and flagging toxic comments in realtime across various social media platforms. By employing advanced NLP algorithms and ML models, the system aims to analyze textual content swiftly and accurately, pinpointing instances of toxicity. Once identified, the system will promptly notify moderators, enabling swift intervention and potential removal of the harmful content. By implementing this technological solution, the project aspires to contribute significantly to fostering a safer and more inclusive online environment where users can engage without fear of encountering toxic behavior. Through the fusion of NLP and ML, this endeavor aims to exemplify the transformative potential of technology in mitigating the challenges posed by toxic comments in the digital age.

Enhancing Transparency and Interpretability in Toxic Comment Classification: A Study on the Integration of Explainable Artificial Intelligence (XAI) Techniques || Dr. Pallavi Devendra Tawde, Mr. Jadyn Dias

More than ever, robust, and interpretable toxic comment recognition methods are required to manage the growing frequency of toxic comments on online platforms. The research tries to incorporate techniques in Explainable Artificial Intelligence (XAI) to improve the transparency and comprehensibility of toxic comment classification. Using a comprehensive dataset, we designed a model architecture which includes the latest practices in XAI. Through rigorous experimentation, our study proves the usefulness of such methods as tools that not only increase classification accuracy but also illuminate model decision-making processes. One view is that by adding LIME and Eli5 to toxic comment classification, model performance improves both in terms of accuracy and interpretation for decisions. Our results provide valuable insights into the model's strengths and areas for refinement, contributing to the transparency and interpretability of toxic comment classification. This research contributes to the evolving landscape of interpretable machine learning, offering a pathway to more accountable and trustworthy toxic comment moderation systems. Keywords: Explainable artificial intelligence, Model interpretability, toxic comment classification, LIME, Eli5

Dataset Augmentation for Counteracting Bias in Toxic Comment Classification


 Toxic comments are a prevalent issue on online social media and networking platforms. These comments contain offensive, malicious, hate speech, or other harmful content that negatively impacts audiences and communities. Effectively detecting and categorizing toxic comments is essential for maintaining order in the online environ ment, protecting user safety, and enhancing user experience. This is despite the fact that researchers and companies have developed various models to recognize toxicity in online chats and comments, achieving some success. However, many of the currently used models incorrectly classify non-toxic comments that contain certain identity terms as potentially toxic. This misclassification hinders the ability to accurately identify categorized comments. In this paper, the detection and classification of toxic comments were implemented using Term Frequency-Inverse Document Frequency (TF-IDF) and machine learning techniques. Additionally, two dataset-specific optimizations were proposed to mitigate the impact of bias on text classification by expanding the number of datasets. Comparative analysis of bias evaluation metrics demonstrates that this approach can effectively mitigate bias while maintaining the accuracy of the original model as much as possible.

TOXIC COMMENT CLASSIFICATION USING BI-LSTM

In the age of internet and social media platforms, the problem of toxic remarks has become increasingly prominent. This research addresses the application of advanced deep learning technique, specifically C Bidirectional Long Short-Term Memory networks (Bi-LSTM), for the classification of toxic comments. Additionally, we employ pre-trained GloVe word embeddings to enhance the performance of the models. The project intends to increase the accuracy and efficiency of toxic comment classification, enabling platforms to automatically recognize and filter out harmful information. By utilizing Bi-LSTM architectures, which excel in capturing spatial and temporal correlations in textual data, we can effectively identify toxic language. The integration of GloVe embeddings significantly strengthens the semantic comprehension of words, contributing to more exact categorization results. Through a comprehensive analysis and evaluation of the proposed models on benchmark datasets such as the Jigsaw Multilingual Toxic Comment Classification dataset, we demonstrate the effectiveness of Bi- LSTM with GloVe embeddings in accurately identifying toxic comments. By reaching high classification accuracy, precision, recall, and F1 scores, the models highlight their potential in minimizing the harmful impact of toxic comments in online contexts. The results of this study have significant implications for online platforms, social media companies, and community moderators seeking automated solutions for content moderation. Key Words: Toxic comment classification, Bidirectional Long Short-Term Memory(BI-LSTM), GloVe embeddings, Deep learning.

A hybrid deep learning approach for Assamese toxic comment detection in social media

The presence of toxic comments on online platforms creates significant barriers to encouraging positive conversation and user involvement. The present research introduces a novel hybrid deep learning methodology that combines Bidirectional long-short-term memory (BiLSTM) and Convolutional Neural Network (CNN) architectures to improve toxic comment identification. The major goal is to improve accuracy and efficiency in detecting Assamese toxic content, particularly on social media sites. Due to insufficient existing datasets, information is manually gathered from a wide range of public domains, allowing for a thorough evaluation of the performance of the hybrid method. We used two alternative activation functions in our experiments: sigmoid and softmax. The sigmoid activation obtained 88.43% accuracy, while the softmax activation outperformed with 90.51% accuracy. We have also made an effort to use our suggested Approach in Bengali and Hindi, two additional Indian languages. Due to their comparable Subject-Object-Verb (SOV) linguistic structure to Assamese, we chose these languages, and the results have been fairly encouraging. The findings from the research are extremely significant since they highlight that the hybrid deep learning approach is a promising option for effectively identifying toxic comments on social media networks.

Automated Spelling Error Detection in Assamese Texts using Deep Learning Approaches

The growing prevalence of regional languages, including Assamese, in user-generated content on social media platforms, has brought the problem of spelling errors into focus. In the realm of natural language processing (NLP) tasks, especially in hate speech and toxic comment detection models, the accuracy of spelling becomes of utmost importance to ensure that misspelled terms do not slip through undetected. This highlights the need for effective approaches to address spelling issues in such language-based applications. This research work suggests a novel approach for identifying word-level spelling errors in Assamese text. The approach employs two deep learning techniques, Long-short-term memory (LSTM) and Bidirectional Long-short-term memory (BiLSTM), to accurately detect misspelled words by analyzing the context of each word inside a sentence. The models are trained using a dataset of 2677 Assamese sentences that include both correctly and incorrectly spelled words. The LSTM and Bi-LSTM models have shown significant improvement over previous spelling detection research in Assamese. LSTM achieved an accuracy of 92.72%, while Bi-LSTM achieved 94.59% accuracy. These results highlight the efficacy of deep learning methods in identifying spelling errors in the Assamese language. By ensuring precise spelling detection and reducing misunderstandings caused by misspelled words, the proposed approach can be implemented to improve the reliability of NLP tasks such as identifying toxic comments and hate speech.

Toxic Comment Classification based on Personality Traits Using NLP

Concerns about the frequency of harmful remarks have been raised by the growth of online communication platforms, which makes it difficult to create inclusive and safe digital spaces. This study explores the creation of a strong framework that uses machine learning algorithms and natural language processing (NLP) methods to categorise harmful comments. In order to improve the accuracy and comprehensiveness of categorization, the study investigates the integration of personality trait analysis in addition to identifying hazardous language. A wide range of online comments comprised the dataset that was gathered and put through extensive preparation methods such as text cleaning, lemmatization, and feature extraction. To facilitate the training and assessment of machine learning models, textual data was converted into numerical representations by utilising TF-IDF vectorization and word embeddings. Furthermore, personality traits were extracted from comments using sentiment analysis and language clues, which linked linguistic patterns with behavioural inclinations. The study resulted in the development and assessment of complex categorization models that combined features from textual content and inferred personality traits. The findings show encouraging associations between specific personality qualities and the use of toxic language, providing opportunities to identify subtle differences in toxic comment contexts. In order to provide insights into developing more sophisticated and successful methods of reducing toxicity in online discourse, this study outlines the methodology, major findings, and consequences of incorporating personality traits analysis into the classification of toxic comments.

How do Machine Learning Algorithms Effectively Classify Toxic Comments? An Empirical Analysis

Toxic comments on social media platforms, news portals, and online forums are impolite, insulting, or unreasonable that usually make other users leave a conversation. Due to the significant number of comments, it is impractical to moderate them manually. Therefore, online service providers use the automatic detection of toxicity using Machine Learning (ML) algorithms. However, the model's toxicity identification performance relies on the best combination of classifier and feature extraction techniques. In this empirical study, we set up a comparison environment for toxic comment classification using 15 frequently used supervised ML classifiers with the four most prominent feature extraction schemes. We considered the publicly available Jigsaw dataset on toxic comments written by human users. We tested, analyzed and compared with every pair of investigated classifiers and finally reported a conclusion. We used the accuracy and area under the ROC curve as the evaluation metrics. We revealed that Logistic Regression and AdaBoost are the best toxic comment classifiers. The average accuracy of Logistic Regression and AdaBoost is 0.895 and 0.893, respectively, where both achieved the same area under the ROC curve score (i.e., 0.828). Therefore, the primary takeaway of this study is that the Logistic Regression and Adaboost leveraging BoW, TF-IDF, or Hashing features can perform sufficiently for toxic comment classification.

XRBi-GAC: A hybrid deep learning framework for multilingual toxicity detection

Social media platforms allow people across the globe to share their thoughts and opinions and conveniently communicate with each other. Apart from various advantages of social media, it is also misused by a set of users for hate-mongering with toxic and offensive comments. The majority of the earlier proposed toxicity detection methods are primarily focused on the English language, but there is a lack of research on low-resource languages and multilingual text data. We propose an XRBi-GAC framework comprising XLM-RoBERTa, Bi-GRU with self-attention and capsule networks for multilingual toxic text detection. A loss function is also presented, which fuses the binary cross-entropy loss and focal loss to address the class imbalance problem. We evaluated the proposed framework on two datasets, namely, the Jigsaw Multilingual Toxic Comment dataset and HASOC 2019 dataset and achieved F1-score of 0.865 and 0.829, respectively. The results of the experiments show that the proposed framework has outperformed the state-of-the-art multilingual models XLM-RoBERTa and mBERT on both datasets, which shows the versatility and robustness of the proposed XRBi-GAC framework.

An Intense Study of Machine Learning Research Approach to Identify Toxic Comments

A large number of online public domain comments are usually constructive, but a significant proportion is toxic. The comments include several errors that allow the machine-learning algorithm to train the data set by processing dataset with numerous variety of tasks, in the method of conversion of raw comments previously feeding it to Classification models using a ML method. In this study, we have proposed classification of toxic comments using a ML approach on a multilinguistic toxic comment dataset. The logistic regression method is applied to classify processed dataset, which will distinguish toxic comments from non-toxic comments. The multi-headed model comprises toxicity (obscene, insult, severe toxic, threat, & identity-hate) or Nontoxicity Estimation. We have implemented four models (LSTM, GRU RNN, and BiLSTM) and detected the toxic comments. In Python 3, all models have a simple structure that can adapt to the resolution of other tasks. The classification problem resolution findings are presented with the aid of the proposed models. It has been concluded that all models solve the challenge effectively, but the BiLSTM is the most effective to ensure the best practicable accuracy.

Utilizing subjectivity level to mitigate identity term bias in toxic comments classification

Toxic comment classification models are often found biased towards identity terms, i.e., terms characterizing a specific group of people such as “Muslim” and “black”. Such bias is commonly reflected in false positive predictions, i.e., non-toxic comments with identity terms. In this work, we propose a novel approach to debias the model in toxic comment classification, leveraging the notion of subjectivity level of a comment and the presence of identity terms. We hypothesize that toxic comments containing identity terms are more likely to be expressions of subjective feelings or opinions. Therefore, the subjectivity level of a comment containing identity terms can be helpful for classifying toxic comments and mitigating the identity term bias. To implement this idea, we propose a model based on BERT and study two different methods of measuring the subjectivity level. The first method uses a lexicon-based tool. The second method is based on the idea of calculating the embedding similarity between a comment and a relevant Wikipedia text of the identity term in the comment. We thoroughly evaluate our method on an extensive collection of four datasets collected from different social media platforms. Our results show that: (1) our models that incorporate both features of subjectivity and identity terms consistently outperform strong SOTA baselines, with our best performing model achieving an improvement in F1 of 4.75% over a Twitter dataset; (2) our idea of measuring subjectivity based on the similarity to the relevant Wikipedia text is very effective on toxic comment classification as our model using this has achieved the best performance on 3 out of 4 datasets while obtaining comparative performance on the remaining dataset. We further test our method on RoBERTa to evaluate the generality of our method and the results show the biggest improvement in F1 of up to 1.29% (on a dataset from a white supremacist online forum).

Toxic Comment Classification Based on Bidirectional Gated Recurrent Unit and Convolutional Neural Network

For English toxic comment classification, this paper presents the model that combines Bi-GRU and CNN optimized by global average pooling (BG-GCNN) based on the bidirectional gated recurrent unit (Bi-GRU) and global pooling optimized convolution neural network (CNN) . The model treats each type of toxic comment as a binary classification. First, Bi-GRU is used to extract the time-series features of the comment and then the dimensionality is reduced through global pooling optimized convolution neural network. Finally, the classification result is output by Sigmoid function. Comparative experiments show the BG-GCNN model has a better classification effect than Text-CNN, LSTM, Bi-GRU, and other models. The Macro-F1 value of the toxic comment dataset on the Kaggle competition platform is 0.62. The F1 values of the three toxic label classification results (toxic, obscene, and insult label) are 0.81, 0.84, and 0.74, respectively, which are the highest values in the comparative experiment.

Toxic Comment Classification on Social Media Using Support Vector Machine and Chi Square Feature Selection

The use of social media in society continues to increase over time and the ease of access and familiarity of social media then make it easier for an irresponsible user to do unethical things such as spreading hatred, defamation, radicalism, pornography so on. Although there are regulations that govern all the activities on social media. However, the regulations are still not working effectively. In this study, we conducted a classification of toxic comments containing unethical matters using the SVM method with TF-IDF as the feature extraction and Chi Square as the feature selection. The best performance result based on the experiment that has been carried out is by using the SVM model with a linear kernel, without implementing Chi Square, and using stemming and stopwords removal with the F1 − Score equal to 76.57%.

Cyberbullying Sentiment Analysis with Word2Vec and One-Against-All Support Vector Machine

Depression and social anxiety are the two main negative impacts of cyberbullying. Unfortunately, a survey conducted by UNICEF on 3rd September 2019 showed that 1 in 3 young people in 30 countries had been victims of cyberbullying. Sentiment analysis research will be conducted to detect a comment that contains cyberbullying. Dataset of cyberbullying is obtained from the Kaggle website, named, Toxic Comment Classification Challenge. The pre-processing process consists of 4 stages, namely comment generalization (convert text into lowercase and remove punctuation), tokenization, stop words removal, and lemmatization. Word Embedding will be used to conduct sentiment analysis by implementing Word2Vec. After that, One-Against-All (OAA) method with the Support Vector Machine (SVM) model will be used to make predictions in the form of multi labelling. The SVM model will go through a hyperparameter tuning process using Randomized Search CV. Then, evaluation will be carried out using Micro Averaged F1 Score to assess the prediction accuracy and Hamming Loss to assess the numbers of pairs of sample and label that are incorrectly classified. Implementation result of Word2Vec and OAA SVM model provide the best result for the data undergoing the process of pre-processing using comment generalization, tokenization, stop words removal, and lemmatization which is stored into 100 features in Word2Vec model. Micro Averaged F1 and Hamming Loss percentage that is produced by the tuned model is 83.40% and 15.13% respectively.
 
 Index Terms— Sentiment Analysis; Word Embedding; Word2Vec; One-Against-All; Support Vector Machine; Toxic Comment Classification Challenge; Multi Labelling

Social Media Toxicity Classification Using Deep Learning: Real-World Application UK Brexit

Social media has become an essential facet of modern society, wherein people share their opinions on a wide variety of topics. Social media is quickly becoming indispensable for a majority of people, and many cases of social media addiction have been documented. Social media platforms such as Twitter have demonstrated over the years the value they provide, such as connecting people from all over the world with different backgrounds. However, they have also shown harmful side effects that can have serious consequences. One such harmful side effect of social media is the immense toxicity that can be found in various discussions. The word toxic has become synonymous with online hate speech, internet trolling, and sometimes outrage culture. In this study, we build an efficient model to detect and classify toxicity in social media from user-generated content using the Bidirectional Encoder Representations from Transformers (BERT). The BERT pre-trained model and three of its variants has been fine-tuned on a well-known labeled toxic comment dataset, Kaggle public dataset (Toxic Comment Classification Challenge). Moreover, we test the proposed models with two datasets collected from Twitter from two different periods to detect toxicity in user-generated content (tweets) using hashtages belonging to the UK Brexit. The results showed that the proposed model can efficiently classify and analyze toxic tweets.

A Study of Multilingual Toxic Text Detection Approaches under Imbalanced Sample Distribution

Multilingual characteristics, lack of annotated data, and imbalanced sample distribution are the three main challenges for toxic comment analysis in a multilingual setting. This paper proposes a multilingual toxic text classifier which adopts a novel fusion strategy that combines different loss functions and multiple pre-training models. Specifically, the proposed learning pipeline starts with a series of pre-processing steps, including translation, word segmentation, purification, text digitization, and vectorization, to convert word tokens to a vectorized form suitable for the downstream tasks. Two models, multilingual bidirectional encoder representation from transformers (MBERT) and XLM-RoBERTa (XLM-R), are employed for pre-training through Masking Language Modeling (MLM) and Translation Language Modeling (TLM), which incorporate semantic and contextual information into the models. We train six base models and fuse them to obtain three fusion models using the F1 scores as the weights. The models are evaluated on the Jigsaw Multilingual Toxic Comment dataset. Experimental results show that the best fusion model outperforms the two state-of-the-art models, MBERT and XLM-R, in F1 score by 5.05% and 0.76%, respectively, verifying the effectiveness and robustness of the proposed fusion strategy.

Identification and Classification of Toxic Comment Using Machine Learning Methods

The increase in penetration of usage of internet services has increased exponentially in the past 4 months due to the ongoing pandemic, this has empowered an enormous number of dynamic new and old clients utilizing the web for different administrationsranging from academic, entertainment, industrial, monitoring and the emergence of a new trend in the corporate-life i.e work-from-home. Due to this sudden emergence of the crowd using the web, there has been an ascent in the number of mischievous persons too. Now it is the primary task of every online platform provider to keep the conversations constructive and inclusive. The best example can be referred to, can be twitter, a web-based media stage where people share their views. This platform has already drawn a lot of flak because of the spread of hate speech, insults, threat, defamatory acts which becomes a challenge for many such online providers in regulating them. Thus, there is active research being conducted in the field of Toxic comment classification. Here wecollatenon-identicalmachinelearning and other trivial techniques on the dataset and propose a model that outflanks all others and compares them one-on-one. We have undertaken the Kaggle dataset for the above reason which has been broadly used and one of the prime resources for scholars working in deducing the challenge of toxic comment classification. The results would help up to create an online interface where we would be able to identify the toxicity level in the given phrase or sentence and classify them into their order of toxicity.

A Toxic Comment Classification Model Based on Ensemble

Accurate classification of toxic comments in different languages is an important task in today’s international social networking platforms. In order to improve the classification of toxic comments in different languages, this paper combines the advantages of high accuracy of monolingual model and strong generalization ability of multi-language model, and adopts the ensemble of multilingual model and monolingual model to classify toxic comments. For the monolingual model, the monolingual pre-training model is fine-tuned with labeled task data; for the multilingual model, before fine-tuning the model with labeled task data, a further pre-training is applied to train model on unlabeled data, which aims to make full use of the large amount of unlabeled data and reduce the dependency of amount of labeled comment data while improving the classification effect. Comparative experiments on Conversasion AI’s multilingual toxic comment dataset show that the model in this paper has improved results on different evaluation metrics compared to the XLM-RoBERTa multilingual fine-tuning model, illustrating the validity of the model.

Roman Urdu toxic comment classification

With the increasing popularity of user-generated content on social media, the number of toxic texts is also on the rise. Such texts cause adverse effects on users and society at large, therefore, the identification of toxic comments is a growing need of the day. While toxic comment classification has been studied for resource-rich languages like English, no work has been done for Roman Urdu despite being a widely used language on social media in South Asia. This paper addresses the challenge of Roman Urdu toxic comment detection by developing a first-ever large labeled corpus of toxic and non-toxic comments. The developed corpus, called RUT (Roman Urdu Toxic), contains over 72 thousand comments collected from popular social media platforms and has been labeled manually with a strong inter-annotator agreement. With this dataset, we train several classification models to detect Roman Urdu toxic comments, including classical machine learning models with the bag-of-words representation and some recent deep models based on word embeddings. Despite the success of the latter in classifying toxic comments in English, the absence of pre-trained word embeddings for Roman Urdu prompted to generate different word embeddings using Glove, Word2Vec and FastText techniques, and compare them with task-specific word embeddings learned inside the classification task. Finally, we propose an ensemble approach, reaching our best F1-score of 86.35%, setting the first-ever benchmark for toxic comment classification in Roman Urdu.

AN IMPROVED MULTI-LABELED LSTM TOXIC COMMENT CLASSIFICATION

The origin of text classification was far back to the early '60s. Text classification classified text into different predefined classifications. One of the techniques used for text classification long short term memory, which is an artificial recurrent neural network architecture. Today, all around the world people are expressing themselves with their opinions and also discuss among others via the media. In such a setup, it is quite observable that discussions may arise due to differences in opinion. These discussions might take a dirty side and which may further result in combats over the social media platforms and may lead to offensive language termed as toxic comments. To identify online hate speech, a large number of scientific studies have been devoted to using Natural Language Processing in combination with Machine Learning and Deep Learning methods. Among the challenges of toxic comment, classifiers are the Out-of-vocabulary words problem, which is the occurrence of words that are not present in the training data. Long-Range Dependencies are also a challenge to toxic comment classification. Which is a situation whereby the toxicity of comments often depends on expressions made in the early parts of the comment. This is especially problematic for longer comments. Another challenge is the low accuracy of comment classification techniques. Epoch was used in improving the accuracy of long short term memory. Epoch tends to improve the accuracy of the classifier since it positively affects the speed and quality of the learning process. We have an improvement of 0.4068 in precision, 0.2871 in a recall, 0.2293 in F1, and 0.4291 inaccuracy.