Update README.md

README.md

@@ -1,3 +1,96 @@
 ---
+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}
+}
+```