Multitabqa Base Geoquery

🚀 MultiTabQA (基礎規模模型)

MultiTabQA是一個用於多表問答的模型，它能從多個輸入表中生成答案表，解決了傳統單表問答模型在處理複雜多表查詢時的侷限性，為複雜的表格數據查詢提供了有效的解決方案。

🚀 快速開始

模型描述

MultiTabQA由Vaishali Pal、Andrew Yates、Evangelos Kanoulas和Maarten de Rijke在論文 MultiTabQA: Generating Tabular Answers for Multi-Table Question Answering 中提出。原始代碼倉庫可在 這裡 找到。

MultiTabQA是一個表格問答（tableQA）模型，可從多個輸入表中生成答案表。它能夠處理多表操作符，如UNION、INTERSECT、EXCEPT、JOINS等。該模型基於TAPEX（BART）架構，包含一個雙向（類似BERT）的編碼器和一個自迴歸（類似GPT）的解碼器。

預期用途

你可以使用該原始模型對多個輸入表執行SQL查詢。該模型已在GeoQuery數據集上進行了微調，能夠回答關於多個輸入表的自然語言問題。

使用方法

以下是在transformers庫中使用該模型的示例代碼：

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 = "How many departments are led by heads who are not mentioned?"
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]

# flatten the model inputs in the format: query + " " + <table_name> : table_name1 + flattened_table1 + <table_name> : table_name2 + flattened_table2 + ...  
#flattened_input = question + " " + [f"<table_name> : {table_name} linearize_table(table) for table_name, table in zip(table_names, tables)]
model_input_string = """How many departments are led by heads who are not mentioned? <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'

微調方法

請在 這裡 找到微調腳本。

BibTeX引用和引用信息

@inproceedings{pal-etal-2023-multitabqa,
    title = "{M}ulti{T}ab{QA}: Generating Tabular Answers for Multi-Table Question Answering",
    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 = "Recent advances in tabular question answering (QA) with large language models are constrained in their coverage and only answer questions over a single table. However, real-world queries are complex in nature, often over multiple tables in a relational database or web page. Single table questions do not involve common table operations such as set operations, Cartesian products (joins), or nested queries. Furthermore, multi-table operations often result in a tabular output, which necessitates table generation capabilities of tabular QA models. To fill this gap, we propose a new task of answering questions over multiple tables. Our model, MultiTabQA, not only answers questions over multiple tables, but also generalizes to generate tabular answers. To enable effective training, we build a pre-training dataset comprising of 132,645 SQL queries and tabular answers. Further, we evaluate the generated tables by introducing table-specific metrics of varying strictness assessing various levels of granularity of the table structure. MultiTabQA outperforms state-of-the-art single table QA models adapted to a multi-table QA setting by finetuning on three datasets: Spider, Atis and GeoQuery.",
}

📄 許可證

本項目採用MIT許可證。