|
|
--- |
|
|
language: en |
|
|
tags: |
|
|
- multitabqa |
|
|
- multi-table-question-answering |
|
|
license: mit |
|
|
pipeline_tag: table-question-answering |
|
|
--- |
|
|
|
|
|
# MultiTabQA (base-sized model) |
|
|
|
|
|
MultiTabQA was proposed in [MultiTabQA: Generating Tabular Answers for Multi-Table Question Answering](https://arxiv.org/abs/2305.12820) by Vaishali Pal, Andrew Yates, Evangelos Kanoulas, Maarten de Rijke. The original repo can be found [here](https://github.com/kolk/MultiTabQA). |
|
|
|
|
|
## Model description |
|
|
|
|
|
MultiTabQA is a tableQA model which generates the answer table from multiple-input tables. It can handle multi-table operators such as UNION, INTERSECT, EXCEPT, JOINS, etc. |
|
|
|
|
|
MultiTabQA is based on the TAPEX(BART) architecture, which is a bidirectional (BERT-like) encoder and an autoregressive (GPT-like) decoder. |
|
|
|
|
|
## Intended Uses |
|
|
|
|
|
This pre-trained model can be used on SQL queries over multiple input tables. |
|
|
|
|
|
### How to Use |
|
|
|
|
|
Here is how to use this model in transformers: |
|
|
|
|
|
```python |
|
|
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM |
|
|
import pandas as pd |
|
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained("vaishali/multitabqa-base-sql") |
|
|
model = AutoModelForSeq2SeqLM.from_pretrained("vaishali/multitabqa-base-sql") |
|
|
|
|
|
query = "select count(*) from department where department_id not in (select department_id from management)" |
|
|
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 = query + " " + [f"<table_name> : {table_name} linearize_table(table) for table_name, table in zip(table_names, tables)] |
|
|
model_input_string = """select count(*) from department where department_id not in (select department_id from management) <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' |
|
|
``` |
|
|
|
|
|
### How to Fine-tune |
|
|
|
|
|
Please find the fine-tuning script [here](https://github.com/kolk/MultiTabQA). |
|
|
|
|
|
### BibTeX entry and citation info |
|
|
|
|
|
```bibtex |
|
|
@misc{pal2023multitabqa, |
|
|
title={MultiTabQA: Generating Tabular Answers for Multi-Table Question Answering}, |
|
|
author={Vaishali Pal and Andrew Yates and Evangelos Kanoulas and Maarten de Rijke}, |
|
|
year={2023}, |
|
|
eprint={2305.12820}, |
|
|
archivePrefix={arXiv}, |
|
|
primaryClass={cs.CL} |
|
|
} |
|
|
``` |
