NL2SQL-规则

内容增强的基于BERT的文本到SQL生成 https://arxiv.org/abs/1910.07179

动机

合理：将数据库设计规则融入文本到SQL生成中：

我们使用表格单元格和问题字符串的匹配信息来构造一个向量，其长度与问题长度相同。这个问题向量主要改善了WHERE-VALUE推理结果的性能。因为它注入了答案单元格及其对应表头绑定在一起的知识。如果我们定位了答案单元格，那么我们就定位了包含答案单元格的答案列。
我们使用所有表头和问题字符串的匹配信息来构造一个向量，其长度与表头长度相同。这个表头向量主要改善了WHERE-COLUMN推理结果的性能。

要求

python 3.6

torch 1.1.0

运行

步骤1

数据准备：下载所有原始数据（https://drive.google.com/file/d/1iJvsf38f16el58H4NPINQ7uzal5-V4v4 或 https://download.csdn.net/download/guotong1988/13008037）并将它们放在`data_and_model`目录下。

然后运行data_and_model/output_entity.py

步骤2

训练和评估：train.py

在不使用执行引导解码的BERT-Base-Uncased上的结果

模型	开发集逻辑形式准确率	开发集执行准确率	测试集逻辑形式准确率	测试集执行准确率
SQLova	80.6	86.5	80.0	85.5
我们的方法	84.3	90.3	83.7	89.2

数据

一个数据视图：

{
	"table_id": "1-1000181-1",
	"phase": 1,
	"question": "Tell me what the notes are for South Australia ",
	"question_tok": ["Tell", "me", "what", "the", "notes", "are", "for", "South", "Australia"],
	"sql": {
		"sel": 5,
		"conds": [
			[3, 0, "SOUTH AUSTRALIA"]
		],
		"agg": 0
	},
	"query": {
		"sel": 5,
		"conds": [
			[3, 0, "SOUTH AUSTRALIA"]
		],
		"agg": 0
	},
	"wvi_corenlp": [
		[7, 8]
	],
	"bertindex_knowledge": [0, 0, 0, 0, 4, 0, 0, 1, 3],
	"header_knowledge": [2, 0, 0, 2, 0, 1]
}

对应的表格：

{
	"id": "1-1000181-1",
	"header": ["State/territory", "Text/background colour", "Format", "Current slogan", "Current series", "Notes"],
	"rows": [
		["Australian Capital Territory", "blue/white", "Yaa·nna", "ACT · CELEBRATION OF A CENTURY 2013", "YIL·00A", "Slogan screenprinted on plate"],
		["New South Wales", "black/yellow", "aa·nn·aa", "NEW SOUTH WALES", "BX·99·HI", "No slogan on current series"],
		["New South Wales", "black/white", "aaa·nna", "NSW", "CPX·12A", "Optional white slimline series"],
		["Northern Territory", "ochre/white", "Ca·nn·aa", "NT · OUTBACK AUSTRALIA", "CB·06·ZZ", "New series began in June 2011"],
		["Queensland", "maroon/white", "nnn·aaa", "QUEENSLAND · SUNSHINE STATE", "999·TLG", "Slogan embossed on plate"],
		["South Australia", "black/white", "Snnn·aaa", "SOUTH AUSTRALIA", "S000·AZD", "No slogan on current series"],
		["Victoria", "blue/white", "aaa·nnn", "VICTORIA - THE PLACE TO BE", "ZZZ·562", "Current series will be exhausted this year"]
	]
}