The Tibetan language is widely distributed and exhibits significant dialectal differences, with systematic variations in phonetics, vocabulary, and syntax. Its complex honorific system and flexible syntactic structures result in strong regional characteristics in stylistic expression. These characteristics can easily introduce stylistic noise and semantic ambiguity in machine translation and style conversion tasks, affecting the model's ability to capture core semantics accurately. To address these issues, this paper proposes a neural network fusion strategy model that integrates stylistic feature modeling with semantic alignment. The model introduces Word2Vec pre-trained word vectors at the representation layer to enhance semantic prior knowledge. At the structural layer, it employs a Bi-LSTM encoder and an LSTM decoder, combined with an attention mechanism, to achieve dynamic alignment of contextual information and the collaborative modeling of word vector representations, context encoding, and the attention mechanism. Experimental results based on parallel corpora of the three major dialects—Ü-Tsang, Kham, and Amdo—show that the model achieves an average improvement of 9.61 percentage points in the BLEU score compared to baseline methods and an average reduction of 6.61 in perplexity (PPL), significantly enhancing the accuracy of dialect conversion and the naturalness of stylistic expression.
在人类复杂的语言交际系统中,语言表达的形式往往随着语境、社会关系、交际目的以及地域文化等差异而呈现出高度变异性。这种变异性集中体现在“语体”的区别上,涵盖了从正式的法律、学术文本到非正式的日常口语、社交媒体用语等多种语体形态。语体转换作为自然语言处理(Natural Language Processing,NLP)领域中一项极具挑战性的关键任务,其核心学术价值在于构建一种从源语体向目标语体的高效映射机制,从而解决语言在不同交际场景下产生的异质性问题。实现语体转换的意义不仅在于文本风格的迁移,更在于通过深层的语义保持与句法重构,消除语言表达中的非规范因素,为自然语言理解(Natural Language Understanding,NLU)和自然语言生成(Natural Language Generation,NLG)等上游及下游任务提供标准化的数据支撑,从而在语音识别、机器翻译以及人机交互等多个维度发挥基础性作用。
ShaalanK, Abo BakrH M, ZiedanI. Transferring egyptian colloquial dialect into modern standard arabic[C]. Proceedings of the International Conference on Recent Advances in Natural Language Processing. Borovets:RANLP,2007:525-529
[2]
MilonH R, SabbirS N U, InanA, et al. A comprehensive dialect conversion approach from chittagonian to standard bangla[C]. Proceedings of the 2020 IEEE Region 10 Symposium. Dhaka: IEEE, 2020:214-217.
[3]
格桑居冕,格桑央京.藏语方言概论[M].北京:民族出版社,2002.
[4]
华侃,索南才让.语言学概论[M].北京:中央民族大学出版社,1992.
[5]
吉太加.安多方言特点及成因研究[J].中国藏学,2003(2):5-37.
[6]
周毛草.玛曲藏语研究[M].北京:民族出版社,2003.
[7]
周毛草.古藏语作格助词在现代方言中的表现[J].民族语文,2011(2):34-39.
[8]
东主才让.藏语方言学[M].北京:民族出版社,2020.
[9]
拉毛吉.藏语方言语体转换方法研究[D].拉萨:西藏大学,2025.
[10]
JinD, JinZ, HuZ, et al. Deep learning for text style transfer: A survey[J]. Computational Linguistics, 2022,48(1):155-205.
[11]
SproatR, JaitlyN. An RNN model of text normalization[C]. Proceedings of Interspeech 2017. Stockholm: ISCA, 2017:754-758.