How to train a custom model to extract data from a PDF with nested grouping?

Wingman Cheung 0

Hi,

I would like to train a custom model to extract data for this type of PDF. It contains nested grouping. I can create the table for name list under organization but I cannot find to create a table within a table. Please suggest any method to extract all data from this kind of dynamic document.

Enclosed the screenshots.

Thanks.

romungi-MSFT 43,656 Reputation points Microsoft Employee

2023-08-30T11:56:55.92+00:00

@Wingman Cheung Nested tables are not supported by the prebuilt layout models. It might work for simple scenarios where the cells are merged but for the above format a custom model has to be trained. Even in this scenario if it is a complex scenario, you will have resort to extracting text of each cell and use the column names to recognize text. I think if you have a chance to simplify the table structure you can use the prebuilt layout model instead of building a custom model from scratch. Thanks!!

Wingman Cheung 0

thanks for reply

however, it is a predefined document, i cannot change the table structure.

For custom model, how to train it to extract data?

I would like to extract the data and then create a json like this:

"headings": 
{
	"heading":
	{
		"heading_name":"",
		"subheadings":
		{
			"subheading":
			{
				subheading_name:"",
				organizations:
				{
					"organization"{
					{
						"organization_name":"XXX",
						participants:
						{
							participant:
							{
								"name_eng":"",
								"name_chi":"",
								"title_eng":"",
								"title_chi":""
							},
							participant:
							{
								"name_eng":"",
								"name_chi":"",
								"title_eng":"",
								"title_chi":""
							}
						}
					},
					"organization"{
					{
						"organization_name":"XXX",
						participants:
						{				
							participant:
							{
								"name_eng":"",
								"name_chi":"",
								"title_eng":"",
								"title_chi":""
							}
						}	
					}
				}
			}
		}
	},
	"heading":
	{
		"heading_name":"",
		"subheadings":
		{
			"subheading":
			{
				subheading_name:"",
				organizations:
				{
					"organization"{
					{
						"organization_name":"XXX",
						participants:
						{
							participant:
							{
								"name_eng":"",
								"name_chi":"",
								"title_eng":"",
								"title_chi":""
							},
							participant:
							{
								"name_eng":"",
								"name_chi":"",
								"title_eng":"",
								"title_chi":""
							}
						}
					},
					"organization"{
					{
						"organization_name":"XXX",
						participants:
						{				
							participant:
							{
								"name_eng":"",
								"name_chi":"",
								"title_eng":"",
								"title_chi":""
							},
							participant:
							{
								"name_eng":"",
								"name_chi":"",
								"title_eng":"",
								"title_chi":""
							},
							participant:
							{
								"name_eng":"",
								"name_chi":"",
								"title_eng":"",
								"title_chi":""
							}
						}	
					}
				}
			}
		}
	}	
}

Wingman Cheung 0 Reputation points

2023-08-31T07:42:08.8733333+00:00

I create a sample of document. Pls suggest how to setup the train model. I have added labels for some fields already. However, for the duplicate items, how to do to extract all data?
romungi-MSFT 43,656 Reputation points Microsoft Employee

2023-09-01T06:48:03.7+00:00

@Wingman Cheung You will have to label all data that needs to be extracted. Since the table is not of standard format, use the entire document to label and train. Ensure the visual structure of the document is maintained in the document that needs to be analyzed. Any deviation is provide incorrect results.

The output structure of the document cannot be changed as per the format you mentioned above, the output will contain a standard format with the fields that you have trained, if the field or label is not recognized there will be no value.
Wingman Cheung 0 Reputation points

2023-09-04T03:05:00.22+00:00

Hi Rmungi,

From the sample that I made above, different "Heading", "SubHeading", "Organization" and name list will be shown in the document. I created a field named "heading_name", how can i label [Heading 1] and [Heading 2] to be "heading_name" too? Same problems for the other labels. Please suggest how to do to label all data?

Share via

How to train a custom model to extract data from a PDF with nested grouping?