Multitabqa Base Geoquery

🚀 MultiTabQA (ベースサイズのモデル)

MultiTabQAは、Vaishali Pal、Andrew Yates、Evangelos Kanoulas、Maarten de RijkeによるMultiTabQA: Generating Tabular Answers for Multi-Table Question Answeringで提案されたモデルです。元のリポジトリはこちらで確認できます。

🚀 クイックスタート

MultiTabQAは、複数の入力テーブルから回答テーブルを生成するtableQAモデルです。UNION、INTERSECT、EXCEPT、JOINSなどの多テーブル演算子を扱うことができます。このモデルは、GeoQueryデータセットで微調整されており、複数の入力テーブルに対する自然言語の質問に回答することができます。

✨ 主な機能

• 複数の入力テーブルから回答テーブルを生成することができます。
• 多テーブル演算子（UNION、INTERSECT、EXCEPT、JOINSなど）を扱うことができます。
• TAPEX(BART)アーキテクチャに基づいています。

📦 インストール

このモデルを使用するには、transformersライブラリをインストールする必要があります。以下のコマンドでインストールできます。

pip install transformers pandas

💻 使用例

基本的な使用法

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'

高度な使用法

このモデルを微調整するには、こちらの微調整スクリプトを使用してください。

📚 ドキュメント

モデルの説明

MultiTabQAは、複数の入力テーブルから回答テーブルを生成するtableQAモデルです。このモデルは、TAPEX(BART)アーキテクチャに基づいており、双方向（BERTのような）エンコーダと自己回帰型（GPTのような）デコーダで構成されています。

想定される使用方法

生のモデルを使用して、複数の入力テーブルに対してSQLを実行することができます。このモデルは、GeoQueryデータセットで微調整されており、複数の入力テーブルに対する自然言語の質問に回答することができます。

BibTeXエントリと引用情報

@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.",
}

📄 ライセンス

このプロジェクトはMITライセンスの下で公開されています。