Titulo Estágio
Context-Aware Detection of Harmful and Malicious Content in text using Large Language Models
Local do Estágio
Rua Dom João Castro n.12, 3030-384 Coimbra, Portugal
Enquadramento
The rapid growth of user-generated content across social media, forums, and messaging platforms has brought significant challenges in moderating online environments. Harmful content such as hate speech, racism, fraudulent schemes, and toxic behavior can propagate quickly, damaging individual well-being and societal cohesion.
Traditional keyword-based or shallow machine learning filters often miss subtle, context-dependent cases of harm — such as sarcasm, implicit bias, or coded language. Recent advancements in Large Language Models (LLMs) and contextual language representations offer a powerful way to analyze language with semantic depth, capturing intent, sentiment, and hidden meaning more effectively.
This internship proposes to explore and prototype a system that uses state-of-the-art NLP techniques to classify text into categories such as racist, fraudulent, hateful, or safe, with a strong emphasis on contextual understanding.
Objetivo
The main objective of this master's internship is to develop and evaluate a text classification system capable of detecting harmful or malicious intent in natural language, based on deep semantic context. Specifically, the internship will:
● Investigate current techniques for context-aware classification, using LLMs such as BERT, RoBERTa, or GPT models;
● Build or fine-tune models to distinguish between racist, hateful, fraudulent, and safe content;
● Analyze how semantic nuances, tone, and intent can influence classification decisions;
● Evaluate the model’s accuracy and robustness, particularly in ambiguous or subtle cases;
● Explore explainability methods (e.g., SHAP, attention visualization) to justify model decisions.
Plano de Trabalhos - Semestre 1
1. Review the literature on hate speech detection, fraud detection, and context-aware Natural Language Processing (NLP);
2. Preprocess the dataset, including text cleaning, tokenization, and handling of noise or imbalanced classes;
3. Experiment with Large Language Models (LLMs) for initial classification using zero-shot or few-shot learning approaches;
4. Begin fine-tuning the models using annotated datasets, exploring different preprocessing strategies and training techniques
5. Write the intermediate version of the thesis, covering the state of the art, problem definition, and proposed methodology.
Plano de Trabalhos - Semestre 2
1. Implement contextual inference improvements, such as using conversation history or topic memory (if applicable to the detection task);
2. Evaluate the models using metrics such as precision, recall, F1-score, and assess robustness against adversarial or ambiguous language (e.g., irony, sarcasm, evasion;
3. Integrate explainability techniques (e.g., SHAP, attention visualization) to ensure transparency and interpretability of model predictions;
4. Develop a simple prototype or interface (optional but valuable) to visualize results and interact with the model in a decision support context;
5. Write the final version of the thesis, including results analysis, discussion, conclusions, and future work.
Condições
The trainee will have all the necessary conditions to carry out the planned tasks, being integrated into the research and development teams within European research projects in which OneSource is involved.
Orientador
Luís Miguel Batista Rosa
luis.rosa@onesource.pt 📩