学习笔记（13）- decaNLP训练WikiSQL

将自然语言转为sql语句，达到对话查询报表的效果。

参考资料

参考1

https://mp.weixin.qq.com/s/i7WAFjQHK1NGVACR8x3v0A

语义解析。SQL查询生成与语义解析相关。基于WikiSQL数据集的模型将自然语言问题转化成结构化的SQL查询，以便用户可以使用自然语言与数据库进行交互。WikiSQL通过逻辑形式精确匹配（lfEM）进行评估，以确保模型不会从错误生成的查询中获得正确的答案。

参考2

http://decanlp.com/

Semantic Parsing
Semantic parsing requires models to translate unstructured information into structured formats so that users can interact with structured information (e.g. a database) in natural language . decaNLP includes the WikiSQL dataset, which maps natural language questions into structured SQL queries.

参考3

https://github.com/salesforce/WikiSQL

安装

创建python虚拟环境
下载源码：
git clone https://github.com/salesforce/WikiSQL
cd WikiSQL
pip install -r requirements.txt
tar xvjf data.tar.bz2

数据

解压之后的数据文件目录：

.jsonl文件每行是一个json文件，

.db是SQLite3数据库格式。

查看db文件，可以从这里下载工具：https://github.com/pawelsalawa/sqlitestudio/releases/tag/3.2.1

问题、查询命令和表ID

文件/Users/huihui/git/WikiSQL/data/dev.jsonl

{
	"phase": 1,
	"table_id": "1-10015132-11",  
	"question": "What position does the player who played for butler cc (ks) play?", 
	"sql": {
		"sel": 3, 
		"conds": [
			[5, 0, "Butler CC (KS)"]
		],
		"agg": 0
	}
}

• phase: 数据集收集的阶段，在2个阶段收集WikiSQL。
• table_id: 该问题所在的表ID。
• question: 工作人员编写的自然语言问题。
• sql: 该问题对应的SQL查询语句。有以下子字段:
  • sel: 列的下标。
  • agg: 聚合操作的下标。agg_ops = ['', 'MAX', 'MIN', 'COUNT', 'SUM', 'AVG']
  • conds: 三元组列表:
    • column_index: 列下标
    • operator_index: 操作符的下标。['=', '>', '<', 'OP']
    • condition: 条件的比较值，float或者string

可以进行max、min、count、sum、avg、大于小于等于、这些查询。

表文件

/Users/huihui/git/WikiSQL/data/dev.tables.jsonl

{
	"header": ["Player", "No.", "Nationality", "Position", "Years in Toronto", "School/Club Team"],
	"page_title": "Toronto Raptors all-time roster",
	"types": ["text", "text", "text", "text", "text", "text"],
	"id": "1-10015132-11",
	"section_title": "L",
	"caption": "L",
	"rows": [
		["Antonio Lang", "21", "United States", "Guard-Forward", "1999-2000", "Duke"],
		["Voshon Lenard", "2", "United States", "Guard", "2002-03", "Minnesota"],
		["Martin Lewis", "32, 44", "United States", "Guard-Forward", "1996-97", "Butler CC (KS)"],
		["Brad Lohaus", "33", "United States", "Forward-Center", "1996", "Iowa"],
		["Art Long", "42", "United States", "Forward-Center", "2002-03", "Cincinnati"],
		["John Long", "25", "United States", "Guard", "1996-97", "Detroit"],
		["Kyle Lowry", "3", "United States", "Guard", "2012-Present", "Villanova"]
	],
	"name": "table_10015132_11"
}

数据库db文件

表中列名用col0、col1等替代，目的是为了节省空间。

测试

测试的样例，可见文件test/example.pred.dev.jsonl

{
	"query": {
		"sel": 3,
		"agg": 0,
		"conds": [
			[5, 0, "butler cc (ks)"]
		]
	},
	"seq": {
		"words": ["symselect", "symagg", "symcol", "position", "symwhere", "symcol", "school/club", "team", "symop", "=", "symcond", "butler", "cc", "-lrb-", "ks", "-rrb-"],
		"after": [" ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", "", "", " "],
		"num": [1, 12, 4, 28, 2, 4, 32, 33, 9, 20, 10, 40, 41, 42, 43, 44],
		"gloss": ["SYMSELECT", "SYMAGG", "SYMCOL", "Position", "SYMWHERE", "SYMCOL", "School/Club", "Team", "SYMOP", "=", "SYMCOND", "butler", "cc", "(", "ks", ")"]
	},
	"error": ""
}

提供了一个测试文件test/example.pred.dev.jsonl.bz2. 使用命令 bunzip2 test/example.pred.dev.jsonl.bz2 -k 进行解压。

提供了一个docker文件，打包了一些依赖文件，可以运行评估脚本。

首先在根目录构建镜像
docker build -t wikisqltest -f test/Dockerfile .
然后运行镜像文件
docker run --rm --name wikisqltest wikisqltest
如果一切运行正常，输入如下：
{
  "ex_accuracy": 0.5380596128725804,
  "lf_accuracy": 0.35375846099038116
}

我用了sudo
xuehp@haomeiya002:~/git/WikiSQL$ sudo docker build -t wikisqltest -f test/Dockerfile .
xuehp@haomeiya002:~/git/WikiSQL$ sudo docker run --rm --name wikisqltest wikisqltest