A Pilot Study of Text-to-SQL Semantic Parsing for Vietnamese

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4079–4085 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 4079 A Pilot Study of Text-to-SQL Semantic Parsing for Vietnamese Anh Tuan Nguyen1,∗, Mai Hoang Dao2 and Dat Quoc Nguyen2 1NVIDIA, USA; 2VinAI Research, Vietnam tuananhn@nvidia.com, {v.maidh3, v.datnq9}@vinai.io Abstract Semantic parsing is an important NLP task. However, Vietnamese is a low-resource lan- guage in this research area. In this paper, we present the first public large-scale Text-to- SQL semantic parsing dataset for Vietnamese. We extend and evaluate two strong seman- tic parsing baselines EditSQL (Zhang et al., 2019) and IRNet (Guo et al., 2019) on our dataset. We compare the two baselines with key configurations and find that: automatic Vietnamese word segmentation improves the parsing results of both baselines; the normal- ized pointwise mutual information (NPMI) score (Bouma, 2009) is useful for schema link- ing; latent syntactic features extracted from a neural dependency parser for Vietnamese also improve the results; and the monolingual language model PhoBERT for Vietnamese (Nguyen and Nguyen, 2020) helps produce higher performances than the recent best mul- tilingual language model XLM-R (Conneau et al., 2020). 1 Introduction Semantic parsing is the task of converting natural language sentences into meaning representations such as logical forms or standard SQL database queries (Mooney, 2007), which serves as an im- portant component in many NLP systems such as Question answering and Task-oriented dialogue (Androutsopoulos et al., 1995; Moldovan et al., 2003; Guo et al., 2018). The significant avail- ability of the world’s knowledge stored in rela- tional databases leads to the creation of large-scale Text-to-SQL datasets, such as WikiSQL (Zhong et al., 2017) and Spider (Yu et al., 2018), which help boost the development of various state-of- the-art sequence-to-sequence (seq2seq) semantic parsers (Bogin et al., 2019; Zhang et al., 2019; Guo ∗Work done during internship at VinAI Research. et al., 2019). Compared to WikiSQL, the Spider dataset presents challenges not only in handling complex questions but also in generalizing to un- seen databases during evaluation. Most SQL semantic parsing benchmarks, such as WikiSQL and Spider, are exclusively for En- glish. Thus the development of semantic parsers has largely been limited to the English language. As SQL is a database interface and universal se- mantic representation, it is worth investigating the Text-to-SQL semantic parsing task for languages other than English. Especially, the difference in lin- guistic characteristics could add difficulties in ap- plying seq2seq semantic parsing models to the non- English languages (Min et al., 2019). For example, about 85% of word types in Vietnamese are com- posed of at least two syllables (Thang et al., 2008). Unlike English, in addition to marking word bound- aries, white space is also used to separate syllables that constitute words in Vietnamese written texts. For example, an 8-syllable written text “Có bao nhiêu quốc gia ở châu Âu” (How many countries in Europe) forms 5 words “Có bao_nhiêuHow many quốc_giacountry ởin châu_ÂuEurope”. Thus it is inter- esting to study the influence of word …

Content truncated to 500 words