方言作为中华民族优秀的一种非物质文化遗产，其不应该随普通话的日益普及而消失。针对目前方言的标注体系有待完善、语料库构建和计算模型研究等方面的不足等问题，本项目拟在以下方面进行深入的研究和探索：（1）建立一套适用于赣方言处理的标注规范，并手工标注完成一定规模的高质量赣方言篇章平行语料库。（2）基于标注的平行语料，研究基于潜在语义对偶空间的赣方言词对齐模型，并利用自动后处理的词映射机制对抽取的词对齐加以修正。（3）通过融合抽取的词对齐、句子分布式表示等多方面的特征，研究递归神经网络下的句子级赣方言识别模型和基于强化学习机制的篇章级赣方言识别模型，并探索赣方言篇章衔接性与连贯性联合学习机制。本项目开展的研究工作对于推进教育部的语言资源保护工程建设具有重要理论意义和应用价值，同时也有利于江西省有声资源库的后续开发和利用。

Dialect, an excellent intangible cultural heritage of China, should not vanish with the increasing popularization of Mandarin. To improve the dialect annotation system and solve the problems of corpus building and computational modelling, we will work on the following aspects: (1) Propose an annotation scheme for the Gan dialect handling and manually build a high-quality annotated discourse-level parallel corpus of the Gan dialect with a certain scale. (2) Study how the latent semantic dual space mechanism and the automatic post-handling of the word mapping can interact with and benefit from each other to generate the word alignment of the Gan dialect based on the annotated corpus. (3) Propose optimal identification models of the Gan dialect via studying the recurrent neural network and deep reinforcement learning for the sentence-level and discourse-level dialect discrimination respectively based on the fusion of the extracted word alignment and distributed sentence representation features. Meanwhile, study the combination learning mechanism for discourse cohesion and discourse coherence in the Gan dialect. This research has important theoretical significance and application value for promoting the Project of Chinese Language Resources Protection proposed by the Ministry of Education. It is also beneficial to the subsequent development and utilization of the audio resource of Jiangxi Province.