tootouch/sparta

Semantic Parsing And Relational Table Aware Model that generates SQL from question written in Korean language

SPARTA (Semantic Parsing And Relational Table Aware)

This is a term project in Unstructured Text Analysis class. We implement the deep learning model for converting Korean language to SQL query.

Team Members

• Hoonsang Yoon
• Jaehyuk Heo
• Jungwoo Choi
• Jeongseob Kim

Information

• Korea University DSBA Lab
• Advisor: Pilsung Kang

Demo

Check about Demo in here.

Video

Text2SQL Result Video

Dataset

• WikiSQL

tar xvjf data/data.tar.bz2

Korean WikiSQL dataset

unzip data/ko_token.zip
unzip data/ko_token_not_h.zip
unzip data/ko_from_table.zip
unzip data/ko_from_table_not_h.zip

Translation

We translated English question into Korean question in four ways as follows.

Download dataset

No	Method	Data Name	Description
1	Where+Select	ko_token	Keep where values in label and column used in select clause among the words in English question
2	Where	ko_token_not_h	Keep header of table among the words in English question
3	Table+Header	ko_from_table	Keep values and header in table among the words in English question
4	Table	ko_from_table_not_h	Keep values in table among the words in English question

Method 1 (Where+Select)

Method 2 (Where)

Method 3 (Table+Header)

Method 4 (Table)

Run translation

1. Create a question dataframe to translate English to Korean.

bash run_translate.sh value

2. Translate English to Korean by using Google Tanslator (click here!) and copy a text file in ko_data directory such as 'ko_train_question.txt'

3. Insert Korean question

bash run_translate.sh token

SPARTA Model

We use pretrained multilingual BERT as encoder.

Sub Task

1. SQLova [ paper | github ]
2. HydraNet [ paper | github ]

Seq2Seq

1. BRIDGE(TabularSemanticParsing)[ paper | github ]

Evaluation

• Logical Form Accuracy
• Execution Accuracy

Experiments

Model	Task	Test Logical Form Accuracy(%)	Test Execution Accuracy(%)
SQLova	Subtask	65.8	74.3
HydraNet	Subtask	40.4	40.7
Bridge	Generation	54.6	62.1

Download Trained Models

Method	SQlova	Bridge
Where+Select	Download	-
Where	Download	-
Table+Header	Download	-
Table	Download	-

Presentation

Proposal

Interim Findings

Final

Reference

• [1] Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning.
• [2] Hwang, W., Yim, J., Park, S., & Seo, M. (2019). A comprehensive exploration on wikisql with table-aware word contextualization. KR2ML Workship at NeurIPS 2019
• [3] Lyu, Q., Chakrabarti, K., Hathi, S., Kundu, S., Zhang, J., & Chen, Z. (2020). Hybrid ranking network for text-to-sql. arXiv preprint arXiv:2008.04759.
• [4] Xi Victoria Lin, Richard Socher and Caiming Xiong. Bridging Textual and Tabular Data for Cross-Domain Text-to-SQL Semantic Parsing. Findings of EMNLP 2020.