The TITAN Platform represents a sophisticated integration of technologies and tools designed to enhance the detection of disinformation, improve critical thinking, and foster collaborative discussions. Below, we discuss the architectural components of the first major release of the TITAN platform, which was utilised during the Living Labs and initial pilot phases to gather user feedback. Each component is essential for achieving TITAN's overarching goal of facilitating informed and productive interactions in the fight against disinformation.
User Content
The User Content component serves as the foundation for the platform's functionality. This module handles all user-provided textual data, including articles sourced from the web or freeform text. It is an integral part of TITAN’s processing pipeline, feeding the Socratic Chatbot, the Critical Thinking Assessment, and the Collaboration Environment.
The Socratic Chatbot, for instance, utilises the user content to engage in meaningful discussions while assessing both the accuracy of the information and the potential presence of disinformation signals. The User Content module communicates seamlessly with other system components to store, analyze, and manage textual data, ensuring a smooth flow of information for disinformation detection and user interaction.
Critical Thinking Assessment
The Critical Thinking Assessment component focuses on evaluating users' ability to analyse and interpret information critically. This module implements questionnaires that help gauge the users' skills in critical thinking.
These questionnaires are vital for determining how well users can recognise logical inconsistencies, evaluate evidence, and assess the validity of claims. By directly engaging users in critical thinking exercises, this component enhances their ability to engage in more nuanced, informed discussions.
Disinformation Signal Detection
The Disinformation Signal Detection component is a powerful tool designed to automatically identify disinformation markers such as hate speech, offensive language, clickbait, and logical fallacies in user-generated content. There are three main sub-components:
Hate Speech, Offensive Language & Clickbait Binary Classification Models
Using state-of-the-art Transformer-based models, TITAN employs pre-trained models such as XLM-RoBERTa and RoBERTa to detect hate speech, offensive language, and clickbait in English texts. The models are fine-tuned using a variety of datasets, achieving high accuracy in classifying texts as either containing these disinformation signals or not.
These models are integrated into the TITAN API, enabling users to submit texts for real-time disinformation detection. The models' performance is continually improved using metrics such as Macro-F1 scores.
Multi-Label Hate Speech Detection Model
Given the complexity of hate speech detection, TITAN developed a multi-label hate speech detection model. This model can identify hate speech across multiple categories—such as racism, sexism, and religious intolerance—by fine-tuning a pre-trained distilroberta-base model on a curated dataset. This model allows for more nuanced detection of hate speech beyond binary classification.
Logical Fallacy Detection with Large Language Models
For identifying logical fallacies, the TITAN platform fine-tunes a Llama2-13b model. By training the model on specialized datasets, TITAN ensures that it can accurately detect various fallacies, such as slippery slopes and bandwagon arguments. The model's output is structured in JSON format to facilitate efficient extraction and categorisation of logical inconsistencies.
Socratic Chatbot
The Socratic Chatbot is a conversational agent that guides users through structured dialogues, based on Socratic questioning techniques. The chatbot helps users dissect disinformation by asking probing questions, encouraging critical analysis of the content they encounter.
This chatbot was initially built using a rule-based system and trained on meticulously crafted Socratic dialogues. As users interact with the chatbot, the system tailors its responses based on the specific disinformation signals detected in the provided text, ensuring a customised and impactful conversational flow.
The TITAN Platform's core components are evolving all the time to work together to detect disinformation, foster critical thinking, and encourage collaborative discussions. By leveraging advanced language models, engaging in Socratic dialogue, and integrating user feedback, the platform continuously refines its ability to address the growing challenges of misinformation in the digital age. Each component—from content analysis to critical thinking assessments—plays a crucial role in ensuring that users can engage in well-informed, productive conversations online.
Read more in our Architecture Report (D1.4)
Comments