语言: 中文 标签:

• 多表问答
• 多表问题回答 许可证: MIT 管道标签: 表格问答 数据集:
• vaishali/geoQuery-tableQA

MultiTabQA (基础规模模型)

MultiTabQA由Vaishali Pal、Andrew Yates、Evangelos Kanoulas和Maarten de Rijke在论文MultiTabQA: 为多表问答生成表格答案中提出。原始代码库可在此处找到。

模型描述

MultiTabQA是一个表格问答模型，能够从多个输入表格生成答案表格。它可以处理多表操作，如UNION、INTERSECT、EXCEPT、JOINS等。

MultiTabQA基于TAPEX(BART)架构，包含一个双向（类似BERT的）编码器和一个自回归（类似GPT的）解码器。

预期用途

您可以使用该原始模型对多个输入表格执行SQL操作。该模型已在GeoQuery数据集上进行了微调，能够回答涉及多个输入表格的自然语言问题。

使用方法

以下是如何在transformers中使用该模型：

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import pandas as pd

tokenizer = AutoTokenizer.from_pretrained("vaishali/multitabqa-base-geoquery")
model = AutoModelForSeq2SeqLM.from_pretrained("vaishali/multitabqa-base-geoquery")

question = "有多少个部门的负责人未被提及？"
table_names = ['department', 'management']
tables=[{"columns":["Department_ID","Name","Creation","Ranking","Budget_in_Billions","Num_Employees"],
                  "index":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14],
                  "data":[
                          [1,"State","1789",1,9.96,30266.0],
                          [2,"Treasury","1789",2,11.1,115897.0],
                          [3,"Defense","1947",3,439.3,3000000.0],
                          [4,"Justice","1870",4,23.4,112557.0],
                          [5,"Interior","1849",5,10.7,71436.0],
                          [6,"Agriculture","1889",6,77.6,109832.0],
                          [7,"Commerce","1903",7,6.2,36000.0],
                          [8,"Labor","1913",8,59.7,17347.0],
                          [9,"Health and Human Services","1953",9,543.2,67000.0],
                          [10,"Housing and Urban Development","1965",10,46.2,10600.0],
                          [11,"Transportation","1966",11,58.0,58622.0],
                          [12,"Energy","1977",12,21.5,116100.0],
                          [13,"Education","1979",13,62.8,4487.0],
                          [14,"Veterans Affairs","1989",14,73.2,235000.0],
                          [15,"Homeland Security","2002",15,44.6,208000.0]
                        ]
                  },
                  {"columns":["department_ID","head_ID","temporary_acting"],
                    "index":[0,1,2,3,4],
                    "data":[
                            [2,5,"Yes"],
                            [15,4,"Yes"],
                            [2,6,"Yes"],
                            [7,3,"No"],
                            [11,10,"No"]
                          ]
                  }]

input_tables = [pd.read_json(table, orient="split") for table in tables]

# 将模型输入格式化为：查询 + " " + <table_name> : table_name1 + 扁平化表格1 + <table_name> : table_name2 + 扁平化表格2 + ...  
model_input_string = """有多少个部门的负责人未被提及？ <table_name> : department col : Department_ID | Name | Creation | Ranking | Budget_in_Billions | Num_Employees row 1 : 1 | State | 1789 | 1 | 9.96 | 30266 row 2 : 2 | Treasury | 1789 | 2 | 11.1 | 115897 row 3 : 3 | Defense | 1947 | 3 | 439.3 | 3000000 row 4 : 4 | Justice | 1870 | 4 | 23.4 | 112557 row 5 : 5 | Interior | 1849 | 5 | 10.7 | 71436 row 6 : 6 | Agriculture | 1889 | 6 | 77.6 | 109832 row 7 : 7 | Commerce | 1903 | 7 | 6.2 | 36000 row 8 : 8 | Labor | 1913 | 8 | 59.7 | 17347 row 9 : 9 | Health and Human Services | 1953 | 9 | 543.2 | 67000 row 10 : 10 | Housing and Urban Development | 1965 | 10 | 46.2 | 10600 row 11 : 11 | Transportation | 1966 | 11 | 58.0 | 58622 row 12 : 12 | Energy | 1977 | 12 | 21.5 | 116100 row 13 : 13 | Education | 1979 | 13 | 62.8 | 4487 row 14 : 14 | Veterans Affairs | 1989 | 14 | 73.2 | 235000 row 15 : 15 | Homeland Security | 2002 | 15 | 44.6 | 208000 <table_name> : management col : department_ID | head_ID | temporary_acting row 1 : 2 | 5 | Yes row 2 : 15 | 4 | Yes row 3 : 2 | 6 | Yes row 4 : 7 | 3 | No row 5 : 11 | 10 | No"""
inputs = tokenizer(model_input_string, return_tensors="pt")

outputs = model.generate(**inputs)

print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
# 'col : count(*) row 1 : 11'

如何微调

请在此处找到微调脚本here。

BibTeX条目和引用信息

@inproceedings{pal-etal-2023-multitabqa,
    title = "{M}ulti{T}ab{QA}: 为多表问答生成表格答案",
    author = "Pal, Vaishali  and
      Yates, Andrew  and
      Kanoulas, Evangelos  and
      de Rijke, Maarten",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.348",
    doi = "10.18653/v1/2023.acl-long.348",
    pages = "6322--6334",
    abstract = "近年来，基于大型语言模型的表格问答（QA）进展受限于其覆盖范围，仅能回答涉及单个表格的问题。然而，现实世界中的查询通常涉及关系数据库或网页中的多个表格，复杂性较高。单表问题不涉及常见的表格操作，如集合操作、笛卡尔积（连接）或嵌套查询。此外，多表操作通常会产生表格形式的输出，这就要求表格QA模型具备表格生成能力。为了填补这一空白，我们提出了一个新的任务：回答涉及多个表格的问题。我们的模型MultiTabQA不仅能回答涉及多个表格的问题，还能推广到生成表格形式的答案。为了有效训练，我们构建了一个包含132,645条SQL查询和表格答案的预训练数据集。此外，我们通过引入不同严格程度的表格特定指标来评估生成的表格，这些指标评估表格结构的各个粒度级别。MultiTabQA在三个数据集（Spider、Atis和GeoQuery）上微调后，优于适应多表QA设置的最先进单表QA模型。",
}