Elementary discourse unit segmentation for Vietnamese texts. (12th April 2022)

• Record Type: Journal Article
• Authors: Nguyen, Chinh Trong; Nguyen, Dang Tuan
• Is Part Of: International journal of intelligent information and database systems. Volume 15:Number 3(2022)
• Page Start: 249
• Page End: 266
• Publication Date: 2022-04-12
• Subjects: EDU segmentation -- sequential labelling -- BERT -- transfer learning; Database management -- Computer programs -- Periodicals; Information retrieval -- Computer programs -- Periodicals; Information storage and retrieval systems -- Computer programs -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Intelligent agents (Computer software) -- Periodicals; 006.33
• Languages: English
• ISSNs: 1751-5858

Abstract

Elementary discourse unit (EDU) segmentation is an important problem in discourse analysis of text. In Vietnam, we do not have any tool or model official published to solve this problem yet. Therefore, we would like to propose a solution for this problem. Our approach is to apply a sequential labelling method for identifying the beginning of each EDU in a sentence. For sequential labelling method, we use a deep neural network architecture containing a BERT for generating word feature vectors as transfer learning approach and a feed forward neural network for identifying the tag of every word. For building the model, we have automatically built an EDU segmentation dataset from a Vietnamese constituent treebank NIIVTB and used this dataset to fine-tune PhoBERT pretrained model. The results show that our EDU segmentation model has span-based F1 score of 0.8, which is sufficient to be used in practical tasks.