Structure of raw Binhvq News Corpus.

This paper introduces a solution to the problem of detecting whether a sequence of text is Vietnamese based on its orthography and contextual features. For those unfamiliar with the language, it is known that understanding the meaning of certain texts can be challenging, since Vietnamese is a complex language that uses Latin characters with diacritics, and many of its words rely heavily on accent marks for semantic distinction. In this paper, we provide insight into how these characteristics influence Transformer-based natural language processing models and propose an approach to address this issue. Transformer-based models are selected due to their superior performance compared to earlier architectures such as RNNs and LSTMs, as well as their widespread application in state-of-the-art NLP systems (GPT, BERT, T5). We examine the specific challenges posed by Vietnamese orthography and word formation, and propose a solution that enhances the model’s ability to distinguish Vietnamese text. Our approach is evaluated on a benchmark dataset, demonstrating high accuracy and robustness in Vietnamese text detection, outperforming conventional methods. The results confirm that Transformer-based models can effectively learn orthographic and contextual patterns in Vietnamese, contributing to improved language identification and multilingual NLP processing.

本文针对基于正字法（orthography）与上下文特征的越南语文本序列检测问题，提出了一种解决方案。对于不熟悉该语言的读者而言，理解部分文本的语义往往颇具挑战——越南语作为一门使用带变音符号拉丁字母的复杂语言，大量词汇需依托重音标记来实现语义区分。本文深入剖析了上述特征对基于Transformer的自然语言处理模型的影响，并提出了针对性的解决方法。选择基于Transformer的模型作为研究对象，是因为相较于循环神经网络（Recurrent Neural Network，RNN）与长短期记忆网络（Long Short-Term Memory，LSTM）等早期架构，其性能更优，且已被广泛应用于当前最优的自然语言处理系统（如GPT、BERT、T5）。我们针对越南语正字法与词汇构成带来的特定挑战展开研究，并提出了一种可提升模型越南语文本识别能力的解决方案。我们在基准数据集上对所提方法进行了评估，结果显示其在越南语文本检测任务中具备高精度与强鲁棒性，优于传统方法。实验结果证实，基于Transformer的模型可有效学习越南语的正字法与上下文模式，有助于提升语言识别与多语言自然语言处理的整体性能。