File size: 5,227 Bytes
4c214e7 eac6ec4 4c214e7 eac6ec4 4c214e7 eac6ec4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 |
---
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}
}
``` |