Training Batch (Concept): Difference between revisions

Revision as of 16:43, 16 April 2020

The Training Set batch is more convenient way to work with all of the samples a Content Model has been trained against

A Content Model and accompanying set of Batches can be found by following this link and downloading the provided file. It is not required to download to understand this article, but can be helpful because it can be used to follow along with the steps in this article. This file was exported from and meant for use in Grooper 2.9

About

During the development and training of Classification of a Grooper Content Model, it can be challenging to keep track of all of the samples you have trained TF-IDF against. In previous versions, each trained sample was stored under each content type in the Grooper Design Studio node tree. In 2.9, the trained samples are stored both under each content type and in the Training Set batch.

How To

Following is an example of how to perform TF-IDF classification that creates the Training Set batch. In the example content model, there are five different content types from three different batches.

!	Some of the tabs in this tutorial are longer than the others. Please scroll to the bottom of each step's tab before going to the step.

PrerequisitesTrain Content TypesReview the Training Set batch

Prerequisites

Following these steps assumes you already have a content model created up with Lexical set as the Classification Method and the appropriate Text Feature Extractor selected. In the example content model, this property is set to Words(Stemmed)

Train Content Types

1. Browse to the Content Model' node and select the Classification Testing tab on the right. 2. Select the appropriate batch in the Batch drop down. 3. Select the document to be trained and select Train Document
4. Repeat these steps for remaining Content Types. In the example Content Model provided, train all five Content Types from all three example batches

Review the Training Set batch

As you train your content types you will see a Training Set batch begin to populate under the Local Resources folder.
A Grooper engineer can review and keep track off all of the documents that have been used for 'TF-IDF Classification training. As the development cycle of Classification continues and more content types are training, the Grooper Engineer now has a single place to review, test and perform regression testing for Classification

It is important to understand that the Training Set is not tied to the actual TF-IDF Weightings that is associated with the Content Type or Content Category. Purging the training from a Content Model does not delete any or all of the documents in the Training Set. Conversely, deleting a document from the Training Set does not remove or purge anyTF-IDF Weightingsfrom a Content Type or Content Category.

Version Differences

Versions prior to Grooper 2.9 do not automatically generate a Training Set batch in the local resources folder

@@ Line 44: / Line 44: @@
 </tab>
 <tab name="Review the Training Set batch" style="margin:25px">
-====Seeing the Results====
+====Review the Training Set batch====
 {|
 | style="padding:25px; vertical-align:top" |
@@ Line 50: / Line 50: @@
 A Grooper engineer can review and keep track off all of the documents that have been used for '''TF-IDF'' Classification training.  As the development cycle of Classification continues and more content types are training, the Grooper Engineer now has a single place to review, test and perform regression testing for Classification <br/>
 <p/><br/>
-It's important to note that because the '''Data Element Overrides''' are applied to a '''Content Type''' a document must be properly classified in order for the '''Data Model''' to know that overrides would be used for extraction for that document. You may be able to successfully test results from the '''Data Element Overrides''' interface without a classified document, but doing so on the '''Data Model''' will result in no extraction.
+|| [[File:Training Batch04.PNG]]
-|| [[File:File:Training Batch04.PNG]]
 |}
 </tab>
 </tabs>
 <br/>
-It is worth noting that one could have accomplished the above by simply making another extractor and set it up for OMR, then have the '''Value Extractor''' '''Data Types''' for each '''Data Field''' simply reference a third element. Overrides would not be necessary in that case. This example, however, sufficed to provide something to show. As with many things in '''Grooper''' there isn't always a ''right'' or ''wrong'' way. There is perhaps a ''best practice'', and in this case, making the third extractor would be the better thing to do.
+It is important to understand that the '''Training Set''' is not tied to the actual '''TF-IDF Weightings''' that is associated with the '''Content Type''' or '''Content Category'''.  Purging the training from a '''Content Model''' does not delete any or all of the documents in the '''Training Set'''.  Conversely, deleting a document from the '''Training Set''' does not remove or purge any'''TF-IDF Weightings'''from a '''Content Type''' or '''Content Category.'''
-</p>
-A simpler, perhaps more common, example of where '''Data Element Overrides''' very much come in handy is with the visibility of '''Data Elements'''. On of the properties of a '''Data Element''' is the '''Visible''' property which is default ''True''. Imagine a '''Data Model''' that has five '''Data Fields''', and the '''Content Model''' has 3 '''Document Types'''. '''Document1''' uses '''Data Fields''' 1-3, '''Document2''' uses '''Data Fields''' 2-4, and '''Document3''' uses '''Data Fields''' 3-5. In '''Data Review''' you want to simplify the job for the person reviewing, so you do not want them to concern themselves with fields that are not relevant. To accomplish this you could use '''Data Element Overrides''' on each of the aforementioned hypothetical '''Document Types''' and set the '''Visibility''' property to ''False'' on all the fields you don't need. This would keep only relevant '''Data Fields''' visibile upon review.
 ==Version Differences==
-Versions prior to '''Grooper 2.9''' had an initial concept version of overrides in the '''Data Element Profiles''' tab located on the '''Content Model''' or '''Document Type'''. These profiles only allowed modification to a limited number of properties on the data element, as opposed to '''Grooper 2.9''' where all properties can be overridden.
+Versions prior to '''Grooper 2.9''' do not automatically generate a '''Training Set''' batch in the local resources folder
-===Where Did Zonal Properties Go?===
-All the zonal extraction properties are now set directly on the '''Data Element'''.