A system for detecting and mitigating multilingual hate speech using deep reinforcement learning

DE202024104313U1Active Publication Date: 2024-11-21CHAURASIA SANDEEP DR JAIPUR +3
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
DE202024104313
Authority / Receiving Office
DE · DE
Patent Type
Utility models
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-21
Estimated Expiration
2034-08-31
Patent Text Reader

Abstract

System for the detection and mitigation of multilingual hate speech using deep reinforcement learning, including: a ZSL (Zero-Shot Learning) component that uses finely crafted argumentation prompts to drive large language models for context-aware hate speech detection; a classifier for conversation outcomes, integrated as a reward mechanism to guide the reinforcement learning process by assigning one of the conversation outcomes as the desired label; a training module that calculates the log probabilities of tokens during model training and in the base model; a mechanism that uses the Kullback-Leibler (KL) divergence between the logarithmic probabilities of two outputs as an additional reward to ensure that the generated responses remain consistent with the base language model; a reinforcement learning module that uses proximal policy optimization (PPO) to train the active model until local stability is achieved; an iterative learning process for the continuous adaptation and improvement of system responses in the effective fight against hate speech.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] The present invention relates to a field of natural language processing and, more particularly, to a system for detecting and curbing multilingual hate speech using deep reinforcement learning

[0002] In response to the rise in hate speech (HS), particularly highlighted in the India Hate Lab's 2023 report with 668 incidents, there is an urgent need to combat the spread of online toxicity. While there are efforts to combat HS using detection and mitigation systems, existing machine learning methods that effectively detect and mitigate HS online are significantly limited, especially in languages ​​other than English and, more importantly, in the Indian context. A key limitation of existing methods is that hate speech detection is a highly context-dependent problem, and these methods cannot fully capture the context of hate speech to make accurate predictions. Recently, LLMs have demonstrated cutting-edge performance in various natural language tasks.LLMs have undergone extensive training with vast amounts of natural language data, enabling them to grasp intricate contextual details.

[0003] However, a fundamental problem with using LLMs for hate speech detection is that there are very few studies on the effective invocation of LLMs for context-aware hate speech detection. Deep RL remains underexplored in modern detection and mitigation systems, especially in the context of HS moderation. While efforts have been directed toward automated counterspeech generation systems, RL techniques have been integrated into these systems to a limited extent. Existing work focuses on generating counterspeech with specific linguistic attributes, such as polite, informative, and intent-driven. However, it remains unclear what impact counterspeech might have in an online environment, especially in the Indian context where multilingualism is crucial.

[0004] To solve the problem, the present invention provides a system for detecting and mitigating multilingual hate speech using deep reinforcement learning.

[0005] The multilingual hate speech detection and containment system that uses deep amplification learning to achieve effective hate speech detection and containment across diverse language landscapes.

[0006] The multilingual hate speech detection and containment system uses deep reinforcement learning to detect hate speech in underrepresented languages ​​without explicit training.

[0007] The Deep Reinforcement Learning Multilingual Hate Speech Detection and Mitigation System uses transformer-based models and LLMs to accurately detect hate speech across diverse language landscapes, an innovative approach to hate speech detection.

[0008] The system for detecting and combating multilingual hate speech using deep reinforcement learning integrates deep RL to generate contextual counter-speech.

[0009] The system for detecting and curbing multilingual hate speech using deep reinforcement learning is commercially viable and suitable for industrial production.

[0010] The multilingual hate speech detection and mitigation system uses deep reinforcement learning to process large amounts of online content, which is critical for real-world applications.

[0011] The multilingual hate speech detection and mitigation system uses deep reinforcement learning to handle large volumes of online content, critical for real-world applications.

[0012] The system for detecting and curbing multilingual hate speech using deep reinforcement learning is suitable for use in various regions and markets, especially in the Indian subcontinent.

[0013] The multilingual hate speech detection and mitigation system using deep reinforcement learning can also be used for integration with existing social platforms that are witnessing a constant rise in Indian users from Tier II and Tier III cities.

[0014] In one embodiment, a system for detecting and mitigating multilingual hate speech using deep reinforcement learning is provided. In this system, zero-shot learning (ZSL) prompting is performed to detect hate speech. The system uses a carefully crafted argument prompt that effectively captures the context of the hate speech and fully exploits the knowledge base in large language models (LLMs), significantly outperforming existing techniques. The system includes a conversation outcome classifier that serves as a reward function, similar to reward models acting as human proxies. This function drives the model's reinforcement learning by specifying one of the conversation outcomes as the desired label.

[0015] In one embodiment, the training process includes calculating the logarithmic probabilities of tokens within the model during training and within the base model. Additionally, the Kullback-Leibler (KL) divergence between the logarithmic probabilities of two outputs is used as an additional reward to ensure that the generated responses do not deviate significantly from the base language model. The active model is trained using the Proximal Policy Optimization (PPO) step until it reaches a state of local stability. This iterative learning process allows the system to continuously adapt and improve its responses to effectively combat hate speech. The integration of reinforcement learning thus offers a promising path to improving the efficiency of counterpeech generation in combating online hate.

[0016] The invention is explained below.

[0017] The system includes zero-shot learning, a conversation outcome classifier, a training module, a mechanism using Kullback-Leibler, a reinforcement learning module, and an iterative learning process. The zero-shot learning (ZSL) prompting component drives large language models for context-dependent hate speech detection. The integrated conversation outcome classifier drives the reinforcement learning process. The training module calculates the logarithmic probabilities of tokens during model training. The mechanism uses the Kullback-Leibler (KL) divergence between the logarithmic probabilities of two outputs to ensure that the generated responses remain consistent with the base language model. The reinforcement learning module uses proximal policy optimization (PPO) to train the active model until local stability is achieved.The iterative learning process for continuously adapting and improving system responses to effectively combat hate speech.

[0018] After detection, the system uses deep reinforcement learning (RL) to generate appropriate counteractions that mitigate the impact of the identified hate speech. This approach enables the system to respond to hate speech dynamically and in real time, thus promoting a safer and more inclusive online environment by effectively combating harmful content across language barriers.

[0019] The transformative impact of HS detection and mitigation in various domains is highlighted. First, it demonstrates how zero-shot learning (ZSL) with the help of instruction-tuning models such as LLMs can serve as a benchmark for cross-platform HS detection in textual content. Second, it precisely identifies specific HSs using transformer-based and instruction-tuning models. Finally, it combats them by generating counter-speech while preserving context using deep reinforcement learning (RL). Counter-speech generation utilizes human proxy-based reward models with pre-trained baseline models, facilitating reinforcement learning through human feedback (RLHF). In the Indian context, it lays the groundwork by training a system on multilingual HS datasets, including major Indian languages.This work aims to stimulate further research and promote collaborative efforts toward more sophisticated, responsible, and human-centered approaches to hate speech moderation in the digital age. The system is designed to detect and counter hate speech, with a focus on underrepresented languages ​​and cultures, such as those in India. It combines several innovative approaches to achieve its goals, including zero-shot learning (ZSL) prompting, transformer-based models, and large language models (LLMs), as well as deep reinforcement learning (RL) for counterpeech generation. ZSL prompting is used to guide LLMs to detect hate speech without explicit training on hate speech datasets. This is crucial for languages ​​with limited prior research. The system uses a coherent prompt that effectively captures the context of hate speech and outperforms existing techniques.Transformer-based models and LLMs are used to precisely identify specific instances of hate speech. These models are known for their ability to capture complex linguistic patterns and can be fine-tuned to detect hate speech with high accuracy and generate appropriate counter-speech. For example, the system uses models such as BERT (Bidirectional Encoder Representations from Transformers) for this task along with the LLM, FLAN-T5. The system uses Deep RL to generate counter-speech that combats hate speech while maintaining the context of the conversation. This is achieved by training the model to generate responses that lead to a desired conversation outcome, similar to human feedback. The RL algorithm used is Proximal Policy Optimization (PPO), which iteratively improves the model's responses based on the rewards it receives.The overall system combines zero-shot hate speech detection, transformer-based identification, and deep reinforcement learning for counterspeech generation. The method of using argumentation prompts to leverage large language models for context-aware, no-shot hate speech detection. The integration of a conversation outcome classifier as a reward function for reinforcement learning for counterspeech generation. The use of KL divergence between the model output and the baseline model output as an additional reward. The iterative reinforcement learning process uses proximal policy optimization (PPO) to continuously adapt counterspeech responses. The application of this system specifically to multilingual hate speech, with a focus on Indic languages.

Claims

[1] A system for detecting and mitigating multilingual hate speech using deep reinforcement learning, comprising: a zero-shot learning (ZSL) component that uses finely crafted reasoning prompts to drive large language models for contextual hate speech detection; a conversation outcome classifier integrated as a reward mechanism to guide the reinforcement learning process by assigning one of the conversation outcomes as the desired label; a training module that calculates the log-probabilities of tokens during model training and in the base model; a mechanism that uses the Kullback-Leibler (KL) divergence between the log-probabilities of two outputs as an additional reward to ensure that the generated answers remain consistent with the base language model; a reinforcement learning module that uses proximal policy optimization (PPO) to train the active model until local stability is achieved; an iterative learning process to continuously adapt and improve system responses to effectively combat hate speech. [2] The system of claim 1, wherein the zero-shot learning (ZSL) prompting component uses reasoning prompts tailored to capture complex contextual cues indicative of hate speech. [3] The system of claim 1, wherein the argumentation prompts are dynamically adjusted based on feedback from previous iterations to improve the accuracy of hate speech detection. [4] The system of claim 1, wherein the conversation outcome classifier undergoes training with a plurality of conversation scenarios, acting as a human proxy to determine the desired conversation outcomes. [5] The system of claim 1, wherein the training module evaluates the performance improvements by comparing the log-likelihoods of tokens during model training with those of a pre-trained baseline model. [6] The system of claim 1, wherein the mechanism using Kullback-Leibler (KL) divergence includes a regularization term to penalize deviations from the base model and thus ensure consistency of the generated responses. [7] The system of claim 1, wherein the reinforcement learning module optimizes the training process by using proximal policy optimization (PPO) with adaptive learning rates to more effectively achieve stable model performance. [8] The system of claim 1, wherein the iterative learning process includes periodic evaluations of the generated counter-speech against a benchmark dataset to continuously refine and improve the effectiveness of the model in mitigating hate speech. [9] The system of claim 1, wherein the ZSL prompting component and the reinforcement learning module are tailored to handle multiple languages, so that the system is able to effectively detect and mitigate hate speech in different linguistic contexts.