Label documents in Tagging-over-Document Business Process

Overview

The Tagging over Document (TOD) Business Process (BP) allows users to label the original document to avoid facing OCR results.

Before: The user tagged the OCR result, which can differ from the original document, and struggled to find a proper value.

After: The user labels the original, taking no time to recognize the familiar document, its structure, the position, font, and font size of the required value.

Result: For classification models, the labeling speed increase is at least 50%, for extraction models: at least 30%. Labeling speed increase, combined with model training time reduction, decreases the time required to deliver an operation to the customer.

Demo

The TOD BP offers several workflows:

1. Tag: select and label values, chunks of sentences, or entire table rows and columns.
2. Search: find other occurrences of a specific value across the document.
3. Evaluate: check the tagged values. Click the extracted value to jump to it even if the document has many pages.

Features

The out-of-the-box TOD Use Case is intended for the easy setup of the TOD BP. The Use Case utilizes preprocessed OCR XML output to:

• Split text into words.
• Calculate absolute coordinates of each symbol.
• Group them into words, lines, columns.
• Generate HTML where a tag over the original document corresponds to an OCRed text underneath.

According to the execution engine architecture, The BP has no limitations for the number of pages and records or the number of documents on any step. Note that each step must complete the processing of all records before moving to the next step.

Limitations

In Intelligent Automation Cloud, the TOD BP has the following limitations:

• The BP does not include a Manual Task and an ETL step.
• The BP does not support predefined autoselect OCR parameters.
• The BP supports OCR with basic or no authentication only. There is no OCR JWT authentication support.
• When tagging a document, including one token in two overlapping or exactly matching sections, the BP produces a resulting empty XML and should be avoided. However, it is possible to include a smaller tag inside the bigger one. For example, tag a City inside an Address.

Configure environment

• AutoML 10.0 or later.
• ImageMagick (Windows: v7.1.0, for example, ImageMagick-7.1.0-17-Q16-x64-static.exe) and Ghostscript (Windows: v9.20) installed to perform a multi-page document split in the case when the TOD OCR type is used.
• Global variables configured according to the Global variables section.

OCR REST API

The OCR BP implemented in AutoML extends the OCR API by retrieving a preprocessed document to use for labeling an IE Manual Task. It wasn't a difficulty before the users tagged an OCRed document. But if the users label the original document, there is a problem with the coordinates of recognized symbols.

The reason is that OCR performs internal preprocessing of a document regardless of the parameters, such as rotating, improving the quality, and so on. The coordinates of the recognized symbols relate to the internally preprocessed document and mismatch to the very original document. Hence, the users do not actually label original document, but the differences are so tiny that they do not notice them.

You can retrieve an OCR preprocessed document using the following methods:

• POST /processDocument or /submitImage with the storePreprocessedDoc=true parameter

  curl

  curl -s --form "file=@FILENAME" "http://ocr.hostname:8080/api/v1/cloud/processImage?correctSkew=true&xml:writeRecognitionVariants=false&profile=documentConversion&exportFormat=txt&language=English&storePreprocessedDoc=true"

  Response

  <?xml version="1.0" encoding="UTF-8" standalone="yes"?><response><task id="5693d8f77b78005fcbbfbe84" message="OK" processEndTime="2016-01-11T16:32:00" processStartTime="2016-01-11T16:31:56" registrationTime="2016-01-11T16:32:28" resultUrl="http://ocr.hostname:8080/api/v1/cloud/download?taskId=5693d8f77b78005fcbbfbe84&amp;result=1" resultUrl2="http://ocr.hostname:8080/api/v1/cloud/download?taskId=5693d8f77b78005fcbbfbe84&amp;result=2" resultUrl3="http://ocr.hostname:8080/api/v1/cloud/download?taskId=5693d8f77b78005fcbbfbe84&amp;result=3" statusChangeTime="2016-01-11T16:32:28" status="Completed"/></response>

• GET /download/preprocessed with the following parameters:

  • taskId: OCR task ID (returned from /processDocument or /submitImage)

  • page: the number of document pages

    curl

    curl -s --form "file=@FILENAME" "http://ocr.hostname:8080/api/v1/cloud/download/preprocessed?taskId=5693d8f77b78005fcbbfbe84&page=1"

Configure Data Stores

OCR cache Data Store

Cache storage is optional for TOD but highly recommended, as it allows to decrease requests to the OCR cluster. Create Data Store for OCR cache in Control Tower. The Data Store with the name you specified is created automatically if it doesn't exist. The most important Data Store parameters are as follows:

Parameter name	Description	Mandatory	Example
ocr_json	Cached response from OCR	yes	{ "original_document_url": "https://services-rnd.s3.amazonaws.com/testing%20OCR/new_fiels/timo0002125.tif", "ocr_export_format": "xmlForCorrectedImage", "original_document_name": "timo0002125", "ocr_document_name": "timo0002125-xmlForCorrectedImage", "ocr_document_format": "xml", "ocr_document_url": "https://vision-automl-db1.workfusion.com:8443/doc-upload/timo0002125-xmlForCorrectedImage.xml" }
key	Document hash	yes	208682471691263be19a1fc0909d3efa
ocr_type	OCR flow used in BP run	yes	tod
export_format	List of formats used in BP	yes	xmlForCorrectedImage, xmlWithoutRecognitionVariants
meta_info_json	Contains all required data for TOD Manual Task, including chars positions, preprocessed document URL, and so on.	yes	{ "pages": "https://vision-s3.workfusion.com:9000/doc-upload/tod/images/preprocessed/links/4e5d19df-b421-4638-8eee-0b723df92438.json", "ocrXmlUrl": "https://vision-s3.workfusion.com:9000/doc-upload/6b674b22-6825-4383-8e37-adea5e6bf54d/39203488-1102-4a5f-acc0-de0d347e9716/1916260-xmlWithoutRecognitionVariants.xml", "imageUrl": "https://bep-vision-db1.workfusion.com/doc-upload/1916260.pdf" }

Global variables

Global Variables are used for the TOD global configuration. Specify the following parameters:

Parameter name	Description	Mandatory	Example
convert_lib_path	Specify the path to the ImageMagick library if it was not added to PATH	no	/usr/bin/convert
ocr_url	URL to OCR service with context path	yes	https://velcom-ocr1.workfusion.com/api/v1/cloud
ocr_username	OCR username	no	<some_username>
ocr_password	OCR password	no

You can change the OCR engine version by changing the URL in global variables. For example, https://velcom-ocr1.workfusion.com/api/v1/cloud > https://velcom-ocr1.workfusion.com/api/v2/cloud (v1 > v2).

Configure Business Process

BP structure

The BP must include the following steps:

• OCR settings
• OCR Workflow
• TOD Manual Task

OCR settings

If the ETL settings step is inserted using drag-and-drop from the bot library to the right of the workflow field, the settings are rewritten for all BPs using this step. To avoid this, drag an empty Bot Task from the toolbar above the workflow field and define it using the OCR settings case as described below.

Add the OCR settings step to the BP before the out-of-the-box TOD BP:

1. On the Workflow tab, drag an empty bot step to the BP designer.

2. Double-click the Bot Task and select an ELT settings Use Case.

   

3. Specify OCR settings according to the instructions below:

Setting	Description	Required	Example or Options
S3 bucket	S3 bucket for TOD and OCR results	yes	doc-upload
OCR Type	OCR flow for the BP specifies the OCR BP use mode	yes	Enriched output (default) Standard
Export Format	Specifies the storage format for the OCR result	yes	html pdfSearchable txt xml xmlForCorrectedImage xmlWithoutRecognitionVariants
Document Language		no	English (default)
Change DPI	DPI value for documents	no	from 50 to 400
Invert image		no	true or false
Discard Color Image		no	true or false
Remove Color Objects	Remove color objects according to the selected option	no	Blue Green Red Yellow
Remove Color Objects Type	Remove color objects type	no	Background Full Stamp
Custom Regions	Custom regions for OCR	no	<some-region>
Allowed Region Types	Allowed region types for OCR	no	BT_Autoanalysis BT_ Barcode BT_Checkmark BT_CheckmarkGroup BP_RasterPicture BT_Separator BT_SeparatorGroup BT_Table BT_Text BT_VectorPicture
Cache Data Store	OCR cache Data Store	no	ocr_cache_datastore described in the OCR cache Data Store section

When the Enriched output OCR Type is selected, the Export Format field is hidden, and xmlWithoutRecognitionVariants, xmlForCorrectedImage are selected.

OCR workflow

To add the OCR Workflow step to your BP, perform the following actions:

1. On the Workflow tab of your BP, select Sub-Process on the right.

2. Expand the OCR section and select OCR Worklow.

3. Drag and drop the OCR Workflow step to the left on the BP Structure field.

   

Manual Task

To label documents after recognition, add Information Extraction Manual Task to the BP after the out-of-the-box TOD step.

The IE Manual Task is generic and accepts both old and new formats. It displays XML or the original document based on the input format and includes an original document if the input file refers to the corresponding JSON. Otherwise, the information extraction task works with any document type as previously.

For more details on the input format, refer to Input.

TOD answer

Configure TOD Answer in the following way:

• Unique Code: document_xml_link (default) or mapped to a column containing XML or link to XML from the previous steps.
• Answer Type: Information Extraction.
• Content Source: Input Data.
• TOD data: meta_info_json (default) or mapped to a column containing JSON from OCR.

The following Unique code values are not allowed:

• block
• formatting
• line
• p
• par
• pre
• rect
• region
• td
• text
• tr
• document
• separator
• page
• regionrect
• textregion
• lineformatting
• textpar
• blocktext
• ptext|pline
• rectrect
• end
• cell
• row
• start
• table
• th
• sec-document
• sec-header
• document
• type
• sequence
• description
• text
• filename
• br

tip

To learn more, see Create IE Answers.

Input

Input data must contain original_document_url.

Input file content sample

original_document_url
https://vision-automl-db1.workfusion.com:8443/doc-upload/d1ceba.tif
https://vision-automl-db1.workfusion.com:8443/doc-upload/b5ytx.tif
https://vision-automl-db1.workfusion.com:8443/doc-upload/p4exv.tif
https://vision-automl-db1.workfusion.com:8443/doc-upload/6ng9y.tif
https://vision-automl-db1.workfusion.com:8443/doc-upload/j7ter.tif

Output

The TOD OCR BP produces meta_info_json in case the Enriched output flow is selected in the OCR settings. Below is a sample of the meta_info_json structure:

meta_info_json

{
 "pages": "https://vision-s3.workfusion.com:9000/doc-upload/tod/images/preprocessed/links/4e5d19df-b421-4638-8eee-0b723df92438.json",
 "ocrXmlUrl": "https://vision-s3.workfusion.com:9000/doc-upload/6b674b22-6825-4383-8e37-adea5e6bf54d/39203488-1102-4a5f-acc0-de0d347e9716/1916260-xmlWithoutRecognitionVariants.xml",
 "imageUrl": "https://bep-vision-db1.workfusion.com/doc-upload/1916260.pdf"
}

• pages: a link to a JSON document containing an array of links to characters recognized on a specific page of the document and the preprocessed image URL

  pages

  [{
      "imgUrl": "https://vision-s3.workfusion.com:9000/doc-upload/tod/images/preprocessed/5694ad66-66c1-4129-848c-9b3e3b66b2ac.png",
      "charsUrl": "https://vision-s3.workfusion.com:9000/doc-upload/tod/images/preprocessed/1-d50e6d9a-5b1b-4887-820b-6280130fadbc.json"
  },
   ...
  ]

• charsUrl content:

  chars

  [{
      "r": "657",
      "b": "80",
      "t": "39",
      "char": "I",
      "l": "652",
      "wordStart": true
  }, {
      "r": "699",
      "b": "80",
      "t": "39",
      "char": "N",
      "l": "667"
  },
   ...
  ]

• ocrXmlUrl: OCR result in the Export Format.

• imageUrl: original document.

Add certificates

The out-of-the-box TOD BP works in all environments. In the case of untrusted certificates, download the SSL DB server certificate and add it to the trusted list. Otherwise, you cannot access the document image in a TOD Manual Task in Workspace.