Understanding and Improving Limitations of Multilingual AI Text Detection
Abstract: With the advances in multilingual large language models (LLMs), recent research has embarked on investigating diverse approaches towards multilingual AI-generated text (AI text) detection, including the fine-tuning of monolingual detectors. In this paper, we pinpoint the limitations in the evaluation procedures of current multilingual AI text detection.
Our extensive analysis uncovers significant inadequacies in all of the available multilingual datasets, including $\textbf{(i)}$ a primary focus on a limited set of languages, $\textbf{(ii)}$ imbalanced data distribution between human and AI-generated samples, and $\textbf{(iii)}$ a lack of diverse yet rich data collection sources.
Amidst these challenges, we propose new methods to $\textbf{(a)}$ improve cross-lingual transfer, $\textbf{(b)}$ exploit novel fine-tuning strategies, $\textbf{(c)}$ analyze the complexities of using neural machine translation (NMT) with monolingual detectors, and $\textbf{(d)}$ a detailed analysis on adversarial robustness. Our results facilitate the engineering of a more resilient model for multilingual text detection, demonstrating superior performance and adaptability across a spectrum of languages.
Paper Type: Long
Research Area: Multilingualism and Cross-Lingual NLP
Research Area Keywords: multilingualism, cross-lingual transfer, multilingual representations, multilingual benchmarks, multilingual evaluation
Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis, Surveys
Languages Studied: Arabic, Catalan, Chinese, Czech, Dutch, English, German, Portuguese, Russian, Spanish, Ukrainian, Urdu, Bulgarian, Indonesian
Submission Number: 997
Loading