Reproduce “Combination of multiple Deep Learning architectures for Offensive Language Detection in Tweets” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nicolò Frisiani, Alexis Laignelet, Batuhan Güler
Researchers at Imperial College London developed a system for the SemEval-2019 Task 6 competition to detect offensive language in tweets. Their hybrid deep learning model, utilizing custom FastText embeddings, achieved macro-average F1 scores of 0.76 for offensive language identification, 0.68 for categorization, and 0.54 for target identification on the official test set.
Key insights
- A combined deep learning architecture, leveraging both sequential (LSTM) and local feature (CNN) extraction, is effective for multi-task offensive language detection in social media.
- Domain-specific, subword-aware embeddings like FastText, especially when fine-tuned, yield superior performance for noisy social media text compared to generic pre-trained alternatives.
- Transfer learning via a shared network backbone feeding into task-specific output layers proves beneficial for solving related multi-task problems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.