"Words are powerful; they can either build bridges or walls. Let’s choose to build a world without hate."
📜 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐓𝐢𝐭𝐥𝐞: "Enhancing Multilingual Hate Speech Detection: From Language-Specific Insights to Cross-Linguistic Integration"
Excited to announce that our research article has been published in the early access section of IEEE Access. This research extends our previous work on hate speech detection, which has been explored within unilingual contexts published by Springer Nature Group: https://lnkd.in/eXiXZTa6, and bilingual contexts, available in IEEE Xplore: https://lnkd.in/d4u7d8uw.
This new study expands our scope to 13 languages: English, German, Chinese, Arabic, Russian, Turkish, Roman Urdu, Korean, Italian, Spanish, Portuguese, French, and Indonesian expanding the limits of what AI can understand and handle in our diverse global community.
𝐈𝐧𝐧𝐨𝐯𝐚𝐭𝐢𝐯𝐞 𝐀𝐧𝐚𝐥𝐲𝐬𝐢𝐬 𝐌𝐞𝐭𝐡𝐨𝐝𝐬:
- Prompt-based Fine-Tuning customizes the learning process to enhance the interpretation of nuanced language usage.
- Transfer Learning with N-1 Strategy and Incremental Learning to help our models learn efficiently and effectively, adapting to new languages and contexts.
- The Unified multilingual hate speech Detection framework integrates advanced techniques into a single, robust model.
- Benchmark testing involves rigorous evaluation of our multilingual hybrid model against benchmarks like GermEval-2018, GermEval-2021, AMI 2018, HateEval-2019, and the Roman Urdu hate speech dataset highlighting its robustness and adaptability across diverse linguistic contexts.
- Integration of Explainable AI (XAI) with LIME, enhancing transparency and understanding in AI decision-making.
Link to the journal article: https://lnkd.in/d2R4NcrN
This work, part of the SOCYTI project: https://lnkd.in/dJbyH3uU led by Rajendra Akerkar and funded by the Research Council of Norway (NFR), explores the complexities of online hate speech.
Thanks to all the co-authors for their contributions to the work. Sule Yildirim Yayilgan, Ibrahim A. Hameed, Muhammad Mudassar Yamin, Mohib Ullah, Mohamed Abomhara
Norwegian University of Science and Technology (NTNU), NTNU Fakultet for informasjonsteknologi og elektroteknikk, MRPET Research Group