Description:Field of Invention
This invention provides a system to detect and classify hate speech in text. The invention takes a multi-layered approach that incorporates syntactic analysis, lexical features, and the contextual semantics of violent language patterns to analyze hate speech. This is a machine learning-based fusion model, with multiple techniques in natural language processing, which allows for increased accuracy and reliability in detecting hate speech occurring over various online platforms.
Objectives of the Invention
The purpose of this invention is to develop a reliable hate speech detection system. It will utilize numerous models of machine learning to better improve the accuracy of detecting and identifying hate speech in a variety of online forums, and use these to limit the damaging effects of harmful content.
Background of the Invention
DE 10 2019 201 988 A1 and GB 2 572 320 A describe a hate speech detection system for online media. This machine learning system identifies potentially controversial content through a multi-step process involving metadata, a representative sample of natural language processing, and a hierarchical scoring system. The Machine Learning (ML) model generates a confidence score based on data tagged with annotations, content features, and question weighting (emotion, person or entity, similarity with harmful content). The product has a feedback loop with a reviewer to mitigate bias. This system includes stance detection, sentiment analysis, and word embedding to identify not only explicit hate but also implicit hate speech.
US 10 861439 B2 is a model of the utility of determining harmful text created by automated language generation systems (The systems may include text summarizers, chat-bots, smart speakers). In short, the model helps automated online language generation software utilize machine learning, trained on n-grams that have been labelled, in pipelines of software to identify and control the possible objectionable material. The model's classification thresholds shift to different audiences, and cultural norms interacting with feedback loops to increase the safety of generated online language.
US 10,936,817 B2 describes a neural network architecture having two encoders; an encoder for background context information and an encoder for the input text; that can be connected using a cross text interaction module. A model to detect subtle hate speech relies on a multilayer perceptron (MLP) classifier. The model utilizes LSTM encoders along with attention mechanisms, which were employed to detect implicit offensive content and improve upon the detection of hate speech without hate related words included.
US 2020/0250267 A1 focuses on detecting mild or nuanced hate speech. The system identifies objectionable language by using LSTM encoders on background information and input text, incorporating both intra and inter attention mechanisms; a representation was formed that is then classified using a multi-layer perceptron (MLP); this improves the ability of the system to detect mild and even subtle offensive language.
US 2021/0019339 A1 establishes a more comprehensive content analysis method for bias detection, hate speech identification, author credibility and explain ability. The system uses machine learning to conduct analysis of online content. The bias detection system synthesis and compares metadata and annotations, in addition to content with entity recognition, in order to detect stance and author credibility with external data to assess credibility. A user interface is provided to enhance and display the data; and finally offers the ability to estimate advertising cost in order to improve user engagement and ensure brand safety.
Summary of the Invention
This system examines internet-based content in a multi-dimensional manner. The system takes metadata, annotations and content features and ties in entity recognizing features to determine the content's stance and author validity. An interface completes the system allowing to modify/ refine data and view the output of the content stance and author validity analysis results. The system will further assist by predicting advertisement costs related to content quality and engagement helping not only with brand protection and user engagement plans.

Brief Description of Drawings
In the figure which are illustrate exemplary embodiments of the invention.
Figure 1, Architecture for NLP System for hate speech detection
Figure 2, Work Flow of the Proposed Invention

Detailed Description of the Invention
This product relates to a system to identify and classify hate speech in digital text, focusing on social media. The system classifies text as hate speech, offensive speech, or neutral speech. The system is also designed to have extensive performance to directly identify and classify informal language, common spelling variations, and the inherent noise found in user-generated content. In pursuit of being as broadly applicable and accurate as possible, the system utilizes a dataset comprised of publicly available hate speech corpora and a unique data collection of social media.
A strong preprocessing pipeline is vital to obtaining a good machine learning model for text data, which acts to normalize the text, lowering each case, excluding punctuation, URLs, symbols, and stop words (e.g. "uhm," "you know," "etc.). Tokenization converts raw text to singular words, and lemmatization reduces words to their most basic form. The resulting normalization, reduces the distribution of different words guaranteeing only meaningful data is collected for model training and instead of casual not offensive noise. Thus ultimately, we can expect to obtain accuracy from normalizing input and reducing the noise.
The invention builds upon two complementary approaches to feature extraction and classification. The first feature extraction utilizes the TF-IDF (Term Frequency - Inverse Document Frequency) method that provides weights to words weather their occurrence given the entire dataset. Using these features we finally fit to a Logistic Regression classifier and regularized using L1 and L2 regularization to avoid overfitting. We will use the second feature extraction option, the XG-Boost classifiers, while including semantic feature representation via word embedding’s that were created through pre-training, Word2Vec, and Glove. Word embedding’s place words with similar contexts or meanings close to each other by converting them to numbers and then into vectors to carry their word context meaning on a new dimension. This would enable the model to capture hidden patterns in the data, that translate well across representation of semantic meanings of the text.
The system consists of two complementary feature extraction and classification pipelines. In the first approach, we use Term Frequency–Inverse Document Frequency (TF-IDF) to measure the importance of words in the complete dataset. The TF-IDF features are provided to the Logistic Regression classifier (with L1 and L2 regularization to avoid overfitting), which is a good approach to identify the text with frequent discriminatory terms or hate speech. The second approach is semantics-based feature extraction using word embedding’s in the form of pre-trained models like Word2Vec and Glove. Word embedding’s include contextual information about the words and provide an input to an XG-Boost classifier, which is good for efficiently learning non-linear patterns. The models together look at both explicit and subtle forms of offensive content.
In order to verify the performance of the system, the dataset is split into training and testing sets in an 80:20 ratio. For the hyper-parameter tuning with grid search and cross-validation techniques, to find settings for both of classifiers, Logistic Regression is tuned with each of the penalty coefficients and XG-Boost is tuned with the learning rate, the maximum depth of the trees, and the number of estimators. This guarantees that both models would be fitted with optimal settings, while optimizing performance, and keeping overfitting minimal!
The performance evaluation was performed using standard classification metrics including accuracy, precision, recall and F1-score. The XG-Boost model achieved good results with 90.72% overall accuracy, and in identifying hate and offensive speech specifically 94% precision and 95% recall. The model did perform well in classifying neutral speech specifically too, but the high number of false positives was a limitation. We subsequently developed an ensemble model that included shockingly auditory confound outs combining the advantages of both linear and non-linear quality. An ensemble hybrid model significantly improved the accuracy of our classification and importantly highlighted the potential weaknesses of one single model type by improving balance across all three categories.
The fusion approach will improve performance and generalizability by combining the output of both classifiers. This ensemble model combines the predictions from the two classifiers using soft voting and stacking. This approach will represent both lexical patterns and semantic context. Hyper-parameter tuning using grid search and cross-validation would be conducted for the optimal model accuracy. For Logistic Regression, the regularization strength (C) would be tuned whereas, for XG-Boost, those hyper-parameters are tree depth, learning rate and the number of estimators. Hyper-parameter tuning would find the balance between the complexity of learning and generalizability. The ensemble model would reduce the over-reliance on a single model's limitations, and better understand and generalize beyond the variety forms of online discourse.
XGBoost method was tested against Twitter data, it provided a good accuracy result being labeled as 90.72% accuracy overall with a high precision-recall balance for its intended task (hate and offensive content classification). Logistic Regression still performed acceptable accuracy during the test, especially for tweets that used uniquely defined keywords, but overall the ensemble method provided better accuracy (overall balance across all three categories). Overall balance between the previous two (the standalone models) was more reliable. As a final note, although neutral content led to multiple mis-labels as they were inevitably mis-labeled due to the classification methods, confusion matrix analysis provided additional tuning angles. The fusion concept provides further improvement accuracy / robustness over conventional hate speech classification methods and confusion matrix review.
Further investigations using confusion matrices showed that it is possible to obtain even better distinction between offensive and neutral content indicated in future advancements involving the adoption of some of the advanced language models such as BERT or RoBERTa which are better at understanding context associated with toxic comments, and reduce classification errors. This invention provides a scalable, efficient, and reliable approach to hate speech detection for social media platforms, moderation tools, and governing bodies seeking to maintain safe and respectful digital spaces. , Claims:The scope of the invention is defined by the following claims:

Claims:
A multi-class text classification system using machine learning for hate speech detection, comprising:
a. A classification mechanism configured to classify input text into one of three labels: hate speech, offensive speech or neutral speech, where said classification mechanism employs a dual modality of classification using a Parallel Logistic Regression classifier and an XG-Boost classifier,
b. Wherein the Logistic Regression classifier applies a linear inductive bias and the XGBoost classifier provides slightly better decision making that is non-linear, with multiple text representation technologies including Term Frequency-Inverse Document Frequency (TF-IDF), Word2Vec embedding’s, and Glove embedding’s, by using said combined models and text representations, the system creates robustness and represents the computing ability to ascertain meaning in the input text for accurate classification into the defined category.
2. According to claim 1, the TF-IDF with Logistic Regression and either Word2Vec or GloVe embedding’s with XGBoost is implemented. It performs preprocessing by to turning all text to lower case, removing special characters, URLs, and stopword, then tokenizing and lemmatizing. These preprocessing steps are successful in assisting the system in detecting hate speech more accurately.
3. According to claim 1, the system involves hyper-parameter tuning through grid search and cross-validation with an easy tuning process. Logistic Regression is tuned with L1 and L2 regularization, and XGBoost tuning is performed on learning rate, tree depth, and estimators. These processes reduce overfitting and improve classification performance.