Text2Sql研究（二）- 大模型测试对比结果

已于 2024-02-22 09:19:53 修改
阅读量1.9k 收藏 16
点赞数 21
分类专栏： Text2SQL 文章标签： 数据库 sql 大数据
于 2024-02-22 09:18:42 首次发布
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/IT_Fly/article/details/136217076
版权

一. 前言

基于之前的文章Text2SQL研究-Chat2DB体验与剖析-CSDN博客，我们了解了Text2Sql基本原理和实现要点，下一步我们就对市面各个商业大模型做了SQL生成测试对比，我们挑选了9条语句，

序号问题
12023年度，累计加油金额是多少元？
22023年，每个季度的加油金额各是多少元？
32023年，每个月的加油金额各是多少元？
42023年，每个月的加油金额总体的变化趋势是怎样的？
5请列出2023年每个月加油金额相比上个月的环比变化百分比
6请列出2023年加油金额最高的十台车，并显示每台车的年度加油费用
72023年加油金额最高的十台车，他们的使用性质各是什么
82023年，加油金额最高的月份是哪个月？这个月哪十辆车加油费用最高
9请计算2023年车辆的平均油耗并列出油耗较低的前十辆车

并针对千帆Qianfan-Chinese-Llama-2-13B、通义千问-Max、讯飞星火-3.5、智普GLM-4做了测试结果分别如下：

二. 测试结果

  1. 千帆Qianfan-Chinese-Llama-2-13B

    千帆Qianfan-Chinese-Llama-2-13B
    语句生成的SQL语句准确性
    1SELECT  SUM(ROUND(fuel_consumption * cost, 2)) AS total_refueling_cost,  SUM(ROUND(mileage * 100)) AS total_driven_distance FROM (SELECT     SUM(fuel_consumption * cost) AS fuel_consumption_cost,   SUM(mileage) AS mileage, organization_id,  vehicle_id,   license_plate,  year  FROM (SELECT  id AS refueling_order_id,  order_number,   license_plate,  organization_id,。。。错误
    2SELECT YEAR(refueling_order.refueling_time) as year,QUARTER(refueling_order.refueling_time) as quarter,SUM(refueling_order.cost) as total_amount FROM  refueling_order GROUP BY  YEAR(refueling_order.refueling_time),  QUARTER(refueling_order.refueling_time)ORDER BY  YEAR(refueling_order.refueling_time),  quarter正确
    3SELECT  DATE_FORMAT(refueling_order.refueling_time, '%Y-%m') AS year,  SUM(refueling_order.fuel_consumption * refueling_order.cost) AS total_amountFROM  refueling_orderGROUP BY  DATE_FORMAT(refueling_order.refueling_time, '%Y-%m')正确
    4SELECT YEAR(refueling_order.refueling_time) as ‘Year’, MONTH(refueling_order.refueling_time) as ‘Month’,SUM(refueling_order.cost) as ‘Total_Refueling_Amount’FROM refueling_order WHERE YEAR(refueling_order.refueling_time) = 2023 GROUP BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time) ORDER BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time)错误
    5SELECT 
      YEAR(refueling_order.refueling_time) AS year, 
      MONTH(refueling_order.refueling_time) AS month, 
      SUM(refueling_order.cost) AS total_amount, 
      SUM(refueling_order.cost - LAG(refueling_order.cost) OVER (PARTITION BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time)) AS prev_month_amount) AS prev_month_total_amount,
      (SUM(refueling_order.cost) - SUM(refueling_order.cost - LAG(refueling_order.cost) OVER (PARTITION BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time))) / SUM(refueling_order.cost - SUM(refueling_order.cost - LAG(refueling_order.cost) OVER (PARTITION BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time)))) * 100 AS percentage_change
    FROM refueling_order GROUP BY  YEAR(refueling_order.refueling_time), 
     MONTH(refueling_order.refueling_time)
    ORDER BY 
      YEAR(refueling_order.refueling_time),     		错误
    6SELECT 
      YEAR(refueling_order.refueling_time) AS year, 
      MONTH(refueling_order.refueling_time) AS month, 
      SUM(refueling_order.cost) AS total_amount, 
      SUM(refueling_order.cost - LAG(refueling_order.cost) OVER (PARTITION BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time)) AS prev_month_amount) AS prev_month_total_amount,
      (SUM(refueling_order.cost) - SUM(refueling_order.cost - LAG(refueling_order.cost) OVER (PARTITION BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time))) / SUM(refueling_order.cost - SUM(refueling_order.cost - LAG(refueling_order.cost) OVER (PARTITION BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time)))) * 100 AS percentage_change
    FROM 
      refueling_order
    GROUP BY 
      YEAR(refueling_order.refueling_time), 
      MONTH(refueling_order.refueling_time)
    ORDER BY 
      YEAR(refueling_order.refueling_time), 
      MONTH(refueling_order.refueling_time)    		正确
    7SELECT 
      YEAR(refueling_order.refueling_time) AS year, 
      MONTH(refueling_order.refueling_time) AS month, 
      SUM(refueling_order.cost) AS total_amount, 
      SUM(refueling_order.cost - LAG(refueling_order.cost) OVER (PARTITION BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time)) AS prev_month_amount) AS prev_month_total_amount,
      (SUM(refueling_order.cost) - SUM(refueling_order.cost - LAG(refueling_order.cost) OVER (PARTITION BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time))) / SUM(refueling_order.cost - SUM(refueling_order.cost - LAG(refueling_order.cost) OVER (PARTITION BY YEAR(refueling_order.refueling_time), MONTH(refueling_order.refueling_time)))) * 100 AS percentage_change
    FROM 
      refueling_order
    GROUP BY 
      YEAR(refueling_order.refueling_time), 
      MONTH(refueling_order.refueling_time)
    ORDER BY 
      YEAR(refueling_order.refueling_time), 
      MONTH(refueling_order.refueling_time)    		错误
    8SELECT YEAR(refueling_order.refueling_time) AS year,
    MONTH(refueling_order.refueling_time) AS month,
    SUM(refueling_order.cost) AS total_cost FROM
    refueling_order GROUP BY YEAR(refueling_order.refueling_time),
    MONTH(refueling_order.refueling_time)
    ORDER BY year DESC,month DESC LIMIT 10;    		错误
    9SELECT 
        r.license_plate,
        r.organization_name,
        r.vehicle_id,
        r.refueling_time,
        r.fuel_consumption,
        r.cost,
        r.year,
        AVG(r.mileage) AS avg_fuel_consumption
    FROM 
        refueling_order r
    JOIN 
        driven_distance d ON r.vehicle_id = d.vehicle_id
    WHERE 
        YEAR(r.refueling_time) = 2023
    GROUP BY 
        r.vehicle_id
    ORDER BY 
        avg_fuel_consumption ASC
    LIMIT 10;    		错误
    【结果评价】6错3对，回答非常啰嗦，准确性不高

  2. SQLCoder-34B-Beta

    SQLCoder-34B-Beta
    语句生成的SQL语句准确性
    1SELECT SUM(r.cost) AS total_cost FROM refueling_order r WHERE EXTRACT(YEAR，FROM r.refueling_time) = 2023;正确
    2SELECT EXTRACT(QUARTER FROM r.refueling_time) AS QUARTER,SUM(r.cost) AS total_cost
    FROM refueling_order r WHERE EXTRACT(YEAR FROM r.refueling_time) = 2023 GROUP BY QUARTER ORDER BY QUARTER NULLS LAST;    		错误
    3SELECT date_trunc('month', r.refueling_time) AS MONTH,
           SUM(r.cost) AS total_cost
    FROM refueling_order r
    WHERE date_part('year', r.refueling_time) = 2023
    GROUP BY MONTH
    ORDER BY MONTH;    		错误
    4SELECT date_trunc('month', r.refueling_time) AS MONTH,
        SUM(r.cost) AS total_cost FROM refueling_order r
    WHERE date_part('year', r.refueling_time) = 2023
    GROUP BY MONTH ORDER BY MONTH;    		错误
    5SELECT date_trunc('month', r.refueling_time) AS MONTH,
           SUM(r.cost) AS total_cost,
           lag(SUM(r.cost)) OVER (
                                  ORDER BY date_trunc('month', r.refueling_time)) AS previous_month_total_cost
    FROM refueling_order r
    WHERE date_part('year', r.refueling_time) = 2023
    GROUP BY MONTH
    ORDER BY MONTH;    		错误
    6SELECT r.license_plate,SUM(r.cost) AS total_cost
    FROM refueling_order r
    WHERE EXTRACT(YEAR FROM r.refueling_time) = 2023
    GROUP BY r.license_plate ORDER BY total_cost DESC NULLS LAST LIMIT 10;    		错误
    7SELECT v.usage_type,SUM(r.cost) AS total_cost FROM refueling_order r JOIN vehicle v ON r.vehicle_id = v.id
    WHERE EXTRACT(YEAR FROM r.refueling_time) = 2023 GROUP BY v.usage_type ORDER BY total_cost DESC LIMIT 10;    		错误
    8SELECT date_trunc('month', o.refueling_time) AS MONTH,
           COUNT(*) AS number_of_orders,SUM(o.cost) AS total_cost
    FROM refueling_order o
    WHERE date_part('year', o.refueling_time) = 2023
    GROUP BY MONTH
    ORDER BY total_cost DESC
    LIMIT 1;    		错误
    9SELECT AVG(r.fuel_consumption) AS average_fuel_consumption
    FROM refueling_order r
    WHERE r.year = 2023
    ORDER BY r.fuel_consumption ASC NULLS LAST
    LIMIT 10;    		错误
    【结果评价】8错1对，返回SQL存在方言问题，且无法通过prompt解决

  3. 通义千问-Ma

    通义千问-Max
    语句生成的SQL语句准确性
    1SELECT v.license_plate, SUM(ro.fuel_consumption) AS total_fuel_consumption, SUM(dd.mileage) AS total_mileage
    FROM vehicle v JOIN refueling_order ro ON v.id = ro.vehicle_id
    JOIN driven_distance dd ON v.id = dd.vehicle_id AND dd.year = [指定年份] WHERE v.organization_id = [某个组织ID]
    GROUP BY v.license_plate;    		错误
    2SELECT    v.license_plate,    SUM(ro.fuel_consumption) AS total_fuel_consumption,    SUM(dd.mileage) AS total_mileageFROM    vehicle vJOIN    refueling_order ro ON v.id = ro.vehicle_idJOIN    driven_distance dd ON v.id = dd.vehicle_id AND dd.year = [指定年份]WHERE    v.organization_id = [某个组织ID]GROUP BY    v.license_plate;错误
    3SELECT    v.license_plate,    SUM(ro.fuel_consumption) AS total_fuel_consumption,    SUM(dd.mileage) AS total_mileageFROM    vehicle vJOIN    refueling_order ro ON v.id = ro.vehicle_idJOIN    driven_distance dd ON v.id = dd.vehicle_id AND dd.year = [指定年份]WHERE    v.organization_id = [某个组织ID]GROUP BY    v.license_plate;错误
    4SELECT    v.license_plate,    SUM(ro.fuel_consumption) AS total_fuel_consumption,    MAX(dd.mileage) AS year_mileageFROM    vehicle vJOIN    refueling_order ro ON v.id = ro.vehicle_idJOIN    driven_distance dd ON v.id = dd.vehicle_id AND dd.year = 2021WHERE    ro.refueling_time BETWEEN '2021-01-01' AND '2021-12-31'GROUP BY    v.license_plate;错误
    5SELECT    v.license_plate,    SUM(ro.fuel_consumption) AS total_fuel_consumption,    SUM(dd.mileage) AS total_mileageFROM    vehicle vJOIN    refueling_order ro ON v.id = ro.vehicle_idJOIN    driven_distance dd ON v.id = dd.vehicle_id AND dd.year = [指定年份]WHERE    v.organization_id = [某个组织ID]GROUP BY    v.license_plate;错误
    6SELECT    v.license_plate,    SUM(ro.fuel_consumption) AS total_fuel_consumption,    SUM(dd.mileage) AS total_mileageFROM    vehicle vJOIN    refueling_order ro ON v.id = ro.vehicle_idJOIN    driven_distance dd ON v.id = dd.vehicle_id AND dd.year = [指定年份]WHERE    v.organization_id = [某个组织ID]GROUP BY    v.license_plate;错误
    7SELECT    v.license_plate,    SUM(ro.fuel_consumption) AS total_fuel_consumption,    SUM(dd.mileage) AS total_mileageFROM    vehicle vJOIN    refueling_order ro ON v.id = ro.vehicle_idJOIN    driven_distance dd ON v.id = dd.vehicle_id AND dd.year = [指定年份]WHERE    v.organization_id = [某个组织ID]GROUP BY    v.license_plate;错误
    8SELECT    v.license_plate,    SUM(ro.fuel_consumption) AS total_fuel_consumption,    SUM(dd.mileage) AS total_mileageFROM    vehicle vJOIN    refueling_order ro ON v.id = ro.vehicle_idJOIN    driven_distance dd ON v.id = dd.vehicle_id AND dd.year = [指定年份]WHERE    v.organization_id = [某个组织ID]GROUP BY    v.license_plate;错误
    9SELECT    v.license_plate,    SUM(ro.fuel_consumption) AS total_fuel_consumption,    SUM(dd.mileage) AS total_mileageFROM    vehicle vJOIN    refueling_order ro ON v.id = ro.vehicle_idJOIN    driven_distance dd ON v.id = dd.vehicle_id AND dd.year = [指定年份]WHERE    v.organization_id = [某个组织ID]GROUP BY    v.license_plate;错误
    【结果评价】9题全错，无论如何更换question，都无法准确识别prompt问题

  4. 讯飞星火-3.5

    讯飞星火-3.5
    语句生成的SQL语句准确性
    1SELECT SUM(cost) FROM refueling_order WHERE YEAR(refueling_time) = 2023;正确
    2SELECT QUARTER(refueling_time) AS quarter, SUM(cost) AS total_cost FROM refueling_order WHERE YEAR(refueling_time) = 2023 GROUP BY quarter;正确
    3SELECT MONTH(refueling_time) as month, SUM(cost) as total_cost FROM refueling_order WHERE YEAR(refueling_time) = 2023 GROUP BY month正确
    4SELECT MONTH(refueling_time) AS month, SUM(cost) AS total_cost FROM refueling_order WHERE YEAR(refueling_time) = 2023 GROUP BY month错误
    5SELECT t1.month, ((t1.total_cost - t2.total_cost) / t2.total_cost) * 100 AS month_on_month_percentage FROM (SELECT MONTH(refueling_time) AS month, SUM(cost) AS total_cost FROM refueling_order WHERE YEAR(refueling_time) = 2023 GROUP BY month) t1 JOIN (SELECT MONTH(refueling_time) AS month, SUM(cost) AS total_cost FROM refueling_order WHERE YEAR(refueling_time) = 2023 AND MONTH(refueling_time) < 12 GROUP BY month) t2 ON t1.month = t2.month + 1;正确
    6SELECT vehicle_id, SUM(cost) AS total_cost FROM refueling_order WHERE YEAR(refueling_time) = 2023 GROUP BY vehicle_id ORDER BY total_cost DESC LIMIT 10正确
    7SELECT v.license_plate, v.usage_type FROM vehicle v JOIN refueling_order r ON v.id = r.vehicle_id WHERE YEAR(r.refueling_time) = 2023 GROUP BY v.license_plate ORDER BY SUM(r.cost) DESC LIMIT 10正确
    8SELECT MONTH(refueling_time) AS month, license_plate, SUM(cost) AS total_cost FROM refueling_order WHERE YEAR(refueling_time) = 2023 GROUP BY month, license_plate ORDER BY total_cost DESC LIMIT 10;错误
    9SELECT vehicle_id, AVG(fuel_consumption) as avg_fuel_consumption FROM refueling_order WHERE YEAR(refueling_time) = 2023 GROUP BY vehicle_id ORDER BY avg_fuel_consumption ASC LIMIT 10;错误
    【结果评价】6对7错，准确率较高，复合问题无法处理

  5. 智普GLM-4

智普GLM-4
语句生成的SQL语句准确性
1SELECT SUM(cost) AS total_cost
FROM refueling_order
WHERE YEAR(refueling_time) = 2023		正确
2SELECT  QUARTER(refueling_time) AS quarter,  SUM(cost) AS total_costFROM  refueling_orderWHERE  YEAR(refueling_time) = 2023GROUP BY  quarterORDER BY  quarter正确
3SELECT    DATE_FORMAT(refueling_time, '%Y-%m') AS month, -- 获取年-月    SUM(cost) AS total_cost -- 计算每个月的加油金额总和FROM    refueling_orderWHERE    YEAR(refueling_time) = 2023 -- 确保只选取2023年的数据GROUP BY    monthORDER BY    month正确
4SELECT  DATE_FORMAT(refueling_time, '%Y-%m') AS month,  SUM(cost) AS total_costFROM  refueling_orderWHERE  DATE_FORMAT(refueling_time, '%Y') = '2023' -- 仅选择2023年的数据GROUP BY  monthORDER BY  month正确
5SELECT  DATE_FORMAT(refueling_time, '%Y-%m') AS month, -- 格式化日期到年-月  SUM(cost) AS monthly_cost, -- 每个月的加油总金额  LAG(SUM(cost)) OVER (ORDER BY DATE_FORMAT(refueling_time, '%Y-%m')) AS previous_month_cost, -- 上个月的加油总金额  (    (SUM(cost) - LAG(SUM(cost)) OVER (ORDER BY DATE_FORMAT(refueling_time, '%Y-%m'))) / NULLIF(LAG(SUM(cost)) OVER (ORDER BY DATE_FORMAT(refueling_time, '%Y-%m')), 0)  ) * 100 AS percentage_change -- 计算环比变化百分比FROM  refueling_orderWHERE  YEAR(refueling_time) = 2023 -- 筛选2023年的数据GROUP BY  monthORDER BY  month正确
6SELECT    ro.license_plate,    SUM(ro.cost) AS annual_fueling_costFROM    refueling_order roJOIN    vehicle v ON ro.vehicle_id = v.idWHERE    YEAR(ro.refueling_time) = 2023GROUP BY    ro.license_plate ORDER BY    annual_fueling_cost DESCLIMIT 10正确
7SELECT    vo.license_plate,    v.usage_type,    SUM(vo.cost) AS total_costFROM    refueling_order voINNER JOIN    vehicle v ON vo.license_plate = v.license_plate WHERE  YEAR(vo.refueling_time) = 2023 GROUP BY    vo.license_plateORDER BY    total_cost DESCLIMIT 10正确
8SELECT  MAX(total_cost_by_month.month) AS max_cost_month,  vehicles_with_max_cost.license_plate,  vehicles_with_max_cost.total_costFROM (  SELECT    EXTRACT(MONTH FROM refueling_time) AS month,    SUM(cost) AS total_cost  FROM    refueling_order  WHERE    EXTRACT(YEAR FROM refueling_time) = 2023  GROUP BY    month) AS total_cost_by_monthJOIN (  SELECT    license_plate,    SUM(cost) AS total_cost,    EXTRACT(MONTH FROM refueling_time) AS month  FROM    refueling_order  WHERE    EXTRACT(YEAR FROM refueling_time) = 2023  GROUP BY    license_plate,    month) AS vehicles_with_max_cost ON total_cost_by_month.month = vehicles_with_max_cost.monthWHERE  total_cost_by_month.total_cost = (    SELECT      MAX(total_cost)    FROM (      SELECT        EXTRACT(MONTH FROM refueling_time) AS month,        SUM(cost) AS total_cost      FROM        refueling_order      WHERE        EXTRACT(YEAR FROM refueling_time) = 2023      GROUP BY        month    ) AS max_cost_per_month  )ORDER BY  vehicles_with_max_cost.total_cost DESCLIMIT 10错误
9SELECT     v.id AS vehicle_id,        v.license_plate,        AVG(ro.fuel_consumption) AS average_fuel_consumption    FROM        vehicle v    JOIN        refueling_order ro ON v.license_plate = ro.license_plate    WHERE        YEAR(ro.refueling_time) = 2023 -- 确保只考虑2023年的数据    GROUP BY        v.id, v.license_plate-- 然后选择油耗较低的前十辆车SELECT    vehicle_id,    license_plate,    average_fuel_consumptionFROM    VehicleAverageFuelConsumptionORDER BY    average_fuel_consumption ASC -- 升序排列，以获取油耗最低的车辆LIMIT 10正确

【结果评价】8对1错，准确率很高，有概率识别复合问题

三. 综合评价

根据您以上测试结果，对市面上几款商业大模型在SQL生成方面的性能进行了比较，可以看出各个模型的表现存在明显差异。

  • 千帆Qianfan-Chinese-Llama-2-13B的表现显示其在准确性方面存在挑战，只有3个问题回答正确，而且回答方式啰嗦，这可能表明其在理解复杂查询和准确生成SQL方面还有改进的空间。
  • SQLCoder-34B-Beta的表现则表明了一个极端的情况，其中8个问题回答错误，仅1个问题回答正确。这可能指出模型在理解特定数据库方言（如Oracle）方面遇到了困难，或者它在根据提示信息生成准确SQL方面存在局限性。
  • 通义千问-Max的测试结果显示了最大的挑战，所有问题都回答错误，这可能意味着模型在理解问题和生成对应SQL方面有根本性的缺陷，或者需要更多的定制化训练以提高其性能。
  • 讯飞星火-3.5展示了较好的性能，尽管在处理复杂问题时遇到了困难，但其准确率相对较高，这表明它在理解简单至中等复杂度问题方面具有一定的优势。
  • 智普GLM-4的性能最为突出，准确率非常高，仅1个问题回答错误。这表明该模型在理解问题和生成准确SQL方面具有较强能力，甚至能够一定程度上处理复合问题。

综合评价来看，各个模型在SQL生成任务上的表现各不相同，体现了在自然语言理解、问题转换、以及特定数据库知识方面的差异。智普GLM-4和讯飞星火-3.5在测试中表现较好，显示出较高的准确性和一定程度上处理复杂问题的能力。而其他模型如SQLCoder-34B-Beta和通义千问-Max则在这些测试中遇到了较大挑战。这些差异可能源于模型的设计、训练数据的丰富性、以及对特定数据库语言的理解等方面的不同。未来的发展可能会看到这些模型在性能上的进一步优化和改进。

梧桐168
关注
专栏目录
NL2SQL进阶系列(4)：ConvAI、DIN-SQL、C3-浙大、DAIL-SQL-阿里等16个业界开源应用实践详解[Text2SQL]
丨汀、的博客
04-13 2318
NL2SQL进阶系列(4)：ConvAI、DIN-SQL等16个业界开源应用实践详解[Text2SQL]
Text2SQL 语义解析数据集、解决方案和学术论文资源整合
qq_38334677的博客
05-21 1053
首先，让我们了解一下Text2SQL。简单来说，Text2SQL就是将人类语言（如英语）的查询转化为数据库查询语言（SQL）。例如，对于一个英语查询 "Who is the oldest employee in the company?"，Text2SQL技术的目标就是将这个英语查询转化为一个SQL查询，如 "SELECT MAX(age) FROM employees"。Text2SQL非常有用，因为它使那些没有数据库查询语言知识的人也能方便地获取数据库中的信息。
大模型LLM在Text2SQL（文本转SQL）上的应用实战
m0_72621177的博客
04-25 7905
大模型LLM在Text2SQL（文本转SQL）上的应用实战
Tool-SQL：基于Agent智能体的Text2SQL解决方案，显著提升Text2SQL效果
最新发布
2401_85378759的博客
09-04 1397
近期，Text-to-SQL 技术通过整合数据库系统的反馈，有效利用了大型语言模型（LLMs）。尽管这些技术能有效纠正 SQL 查询的执行错误，但在处理数据库不匹配问题上仍显不足，这类问题不会引发执行异常。为此，我们设计了一个辅助工具框架，包括检索器和检测器，专门用于诊断并修正 SQL 查询中的不匹配问题，从而提升 LLM 在实际应用中的查询处理能力。同时，我们推出了 Spider-Mismatch 数据集，专门针对现实场景中的条件不匹配问题。实验显示，我们的方法在少样本环境下，不仅在 Spider 系列数
NLP-语义分析（Semantic Parsing）：Text2SQL【“打破人与结构化数据之间的壁垒”，即普通用户可以通过自然语言描述完成复杂数据库的查询工作】
u013250861的博客
04-07 2679
Text to SQL( 以下简称Text2SQL)，是将自然语言文本（Text）转换成结构化查询语言SQL的过程，属于自然语言处理-语义分析（Semantic Parsing）领域中的子任务。，即普通用户可以通过完成复杂数据库的查询工作，得到想要的结果。
大模型LLM在 Text2SQL 上的应用实践
机器学习社区
01-10 4747
目前，大模型的一个热门应用方向Text2SQL，它可以帮助用户快速生成想要查询的SQL语句，再结合可视化技术可以降低使用数据的门槛，更便捷的支持决策。本文将从以下四个方面介绍LLM在Text2SQL应用上的基础实践。· Text2SQL概述· LangChain基础知识· 基于SQLDatabaseChain的Text2SQL实践· 后续计划。
大模型+数据中台（text2sql技术）
haidizym的博客
04-26 2761
大模型+数据中台（text2sql技术）
NL2SQL实践系列(2)：2024最新模型实战效果(Chat2DB-GLM、书生·浦语2、InternLM2-SQL等)以及工业级案例教学
丨汀、的博客
04-14 1592
NL2SQL实践系列(2)：2024最新模型实战效果(Chat2DB-GLM、书生·浦语2、InternLM2-SQL等)以及工业级案例教学
NL2SQL进阶系列(5)：论文解读业界前沿方案（DIN-SQL、C3-SQL、DAIL-SQL、SQL-PaLM）、新一代数据集BIRD-SQL解读
丨汀、的博客
04-15 1421
NL2SQL进阶系列(5)：论文解读业界前沿方案（DIN-SQL、C3-SQL、DAIL-SQL）、新一代数据集BIRD-SQL解读
面向工业生产的中文Text-to-SQL模型.pdf
01-04
"面向工业生产的中文Text-to-SQL模型" 面向工业生产的中文Text-to-SQL模型是指将自然语言查询转换为 Structured Query Language（SQL）语句的技术，旨在解决工业生产中数据查询的需求。本文中提出的模型针对中文...
LLMs之Text2Sql：大语言模型场景实战案例应用—让查询和问答应用更智能(利用LLMs与数据库交互)—Text2Sql任务的简介、多种技术方案(LLM/RAG/ETA等)、排行榜平台(如Spid
近期请国内外头部出版社可尽快私信博主！——心比天高，仗剑走天涯，保持热爱，奔赴向梦想！低调，谦虚，自律，反思，成长，还算是比较正能量的博主，公益免费传播……内心特别想在AI界做出一些可以推进历史进程影响力的东西(兴趣使然，有点小情怀，也有点使命感呀)…
11-26 675
​ LLMs之SQL：大语言模型场景实战案例应用—让查询和问答应用更智能(利用LLMs与数据库交互，比如法律合同/简历/财务数据/客户支持)、多种技术方案(LLM/RAG等) 目录 相关文章 Spider数据集的简介 LLMs场景实战案例：查询(定位到准确目标)、搜索(查找到一个范围) 现有案例实战应用 相关文章 LLMs：《Building LLM applications for production构建用于生产的LLM应用程序》翻译与解读 LLMs：《Building LL
Text2SQL 语义解析数据集、解决方案、paper资源整合项目sql
05-14
Text to SQL( 以下简称Text2SQL)，是将自然语言文本（Text）转换成结构化查询语言SQL的过程，属于自然语言处理-语义分析（Semantic Parsing）领域中的子任务。 它的目的可以简单概括为：“打破人与结构化数据之间的壁垒”，即普通用户可以通过自然语言描述完成复杂数据库的查询工作，得到想要的结果。
gap-text2sql:GAP-text2SQL
05-23
GAP-text2SQL：通过生成增强的预训练学习语义表示的上下文表示 代码和模型 更新 [2020/02/05]支持在自己的数据库和查询上运行模型。 签出。 抽象的 最近，人们通过利用大规模文本语料库来训练具有自我监督学习目标的大型神经语言模型，例如屏蔽语言模型（MLM），来学习各种NLP任务的上下文表示，引起了人们的极大兴趣。 但是，根据一项初步研究，当将现有通用语言模型应用于文本到SQL语义解析器时，我们会观察到三个问题：无法检测到话语中的列提及，无法从单元格值推断出列提及，并且无法编写复杂SQL查询。 为了缓解这些问题，我们提出了一个模型预训练框架，即代增广预训练（GAP），该模型通过利用生成模型来生成预训练数据来共同学习自然语言话语和表模式的表示。 GAP模型针对2M话语-模式对和30K话语-SQL-SQL三元组进行训练，它们的话语是由生成模型产生的。 根据实验结果，利用GAP
【Text2SQL】评估 LLM 的 Text2SQL 能力
VLyb
06-06 1328
⭐⭐⭐⭐。
Text2SQL论文-07：Robust Text-to-SQL Generation with Execution-Guided Decoding
iin729的博客
08-18 472
对于NL2SQL任务提出了一个新的机制->execution guidance ：对部分生成的sql执行情况为条件，在解码过程中检测并排除故障程序，从而对整个对于生成的sql语句进行纠正。 该机制可与任何自回归生成模型一起使用. 基于S2S结构的attention 和 copy机制的生成会fail at generatingsyntactically validqueries 作者的核心思想在于这种sql语句一部分生成以后就可以执行，执行的结果反过来可以引导生成过程 示例 执行opp...
大模型在Text-to-sql领域的应用和论文
qq_36001281的博客
11-19 2211
阿里团队名为DAIL-SQL的新的综合解决方案。使用开源LLM对比OpenAI LLM。文档比较了几种prompt engineering方法，探索高效的提示工程。提出了一种新的Text - to - SQL提示工程方法，命名为DAIL - SQL，该方法以86.6 %的执行准确率刷新了Spider排行榜中的最佳性能。
【论文阅读】让数据库听懂人话（Text-to-SQL）
热门推荐
一只热爱奔跑的程序猿
12-30 1万+
论文标题: 在跨领域数据库中引入中间表示来实现复杂Text2SQL 英文标题: Towards Complex Text-to-SQL in Cross-Domain Database with Intermediate Representation 论文下载链接: https://arxiv.org/abs/1905.08205 论文项目代码: GitHub@IRNet 摘要 本文提出一种名为IRNet的神经方法解决复杂的, 跨领域的(cross-domain) Text-to-SQL任务; IR
完全免费的 Text2SQL：智能SQL代码生成器，开发者的得力助手
乌南竹的博客
08-19 130
大家好，我是IT派同学，一名喜欢思考的互联网爱好者。在最近的几个月里，我参与了多个激动人心的项目，这些项目不仅考验了我的技术能力，更激发了我对 AI 的深刻思考。AI的浪潮正以不可阻挡之势席卷整个技术领域，并且我深信，它将是推动未来软件开发的关键力量。在这些项目中，我花费了大量时间编写SQL代码来处理复杂的业务逻辑。虽然对编写 SQL 并不陌生，但面对繁复的SQL语句和无尽的调试，我发现自己仍然需要一种更高效、更智能的方式来简化这一过程。
Text2SQL论文-02：Abstract Syntax Networks for Code Generation and Semantic Parsing
iin729的博客
08-18 553
转载于https://zhuanlan.zhihu.com/p/79486610 这篇文章主要介绍的一种新颖的neural semantic parsing的方法来做code generation，这个特殊设计网络的名字叫Abstract Syntax Network（ASN），是伯克利在2017年的工作。其中有几个亮点： 使用了神经网络的方式生成code 使用了一种语法树来限制神经网络只能生成有效的输出。这种语法树是由一种叫Abstract Syntax Description Language（A
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值