A Template-Based Approach for Generating Vietnamese References from Flat MR Dataset in Restaurant Domain

Abstract

In recent years, researchers in natural language generation (NLG) focus on corpus-based systems on specific or across domains. The training data should consist of meaning representations (MRs) paired with Natural Language (NL) references. In the first content of the article, we introduce a Vietnamese Flat MR dataset which is the first Vietnamese dataset for training end-to-end, data-driven NLG systems in restaurant domain. We establish a method of generating references on this dataset. The core of the method are two important stages: (i) sentence planning which determine semantic template of the output text; (ii) surface realization which selecting appropriate Vietnamese phrases to replace the corresponding predicates (slot-value) of the Flat MR in the semantic template. The evaluation results show that the dataset and proposed generating method have contributed well to the development of the NLG research direction.