Field-level Accuracy Targets

Field-Level Accuracy Targets (FLAT) let you set a higher or lower Target Accuracy for individual fields and table columns (entries) than the target accuracy set for the rest of the fields in the layout, set in the flow settings. Use FLAT when some entries require a different level of human Supervision compared to others. For example, you can require a higher target accuracy for an account number or email address than for less critical fields, without applying that stricter standard to every entry in the flow.

In this article, you'll learn:

• How FLAT works.

• Which target accuracy the system applies to an entry.

• How to set transcription targets in the Field Dictionary and in a flow.

• How to set identification targets in a flow.

• How to monitor entries that have a custom target accuracy.

How FLAT works

By default, the accuracy target on the flow-level applies the same accuracy threshold to every entry of the same type. For example, all text fields in Structured documents use the Structured Text Target Accuracy value. FLAT replaces that value for the specific entries you select.

FLAT is available for two model types:

Entry-level transcription targets in flow settings are available in v40.1 and later. Entry-level identification targets are available in v41 and later.

A higher target accuracy sets a stricter Threshold, so more values are sent for review. A lower target accuracy results in a less strict threshold and higher Automation for that entry.

Use both directions to match review effort to business impact. For example, in a logistics use case, you can raise the target for the origin and destination fields that downstream processing depends on, and lower the target for item descriptions that do not justify the same review effort.

What FLAT does not do

• FLAT does not guarantee that an entry reaches its target accuracy. It applies a higher or lower threshold to the entry than the general target accuracy would.

• FLAT does not create a separate model for the entry. The logic behind Transcription Supervision uses the Accuracy thresholds of the whole Transcription Model, which apply across all entries.

• FLAT does not generate additional QA tasks. The QA sampling rates in the flow's settings still apply.

If every occurrence of a Field requires human review regardless of machine confidence, configure Transcription Supervision for that field in the layout or in the flow instead of using FLAT.

Entry-level targets accept values from 0.1 to 99.9 because the layout already covers the two extremes. To always send a field to Supervision, set its supervision setting in the layout to Always. To never send it based on confidence, set it to Auto-transcribe for transcription or Auto-identify for identification.

Which target accuracy applies to an entry

The system applies the first value that exists, in the following order:

A target accuracy set in a flow applies only to that flow. Because the same layout can be used in several flows, you can set different targets for the same field in each flow, depending on the business need that each flow serves. A target accuracy set in the Field Dictionary applies to every layout field linked to that dictionary entry.

Before you begin

To set transcription target accuracies, enable the Transcription Automation Training setting for the document type (Structured, Semi-structured, or both). To learn more, see Transcription Settings.

Transcription Automation Training uses a field type's target accuracies only after the minimum number of QA records for that field type is available:

• Structured text fields: 5000 records

• Structured checkbox and signature fields: 2000 records each

• Semi-structured text fields, table cells, and checkboxes: 2000 records each

Until the minimum is met, the system uses the manually configured confidence thresholds for that field type. The Period of records to use flow setting applies to entry-level accuracy calculations in the same way it applies to document-level calculations.

Set a transcription target in the Field Dictionary

Use the Field Dictionary to set a transcription target for a Structured field wherever that field appears:

1. Go to Library > Field Dictionary.

2. Click Add Field.

   • The Create New Field tab opens.

3. Enter the Display Name, and select a data type.

4. In Transcription Target Accuracy (Optional), enter the target as a percentage between 0.1 and 99.9.

5. Click Create.

   • To create another field right away, click Save and Create New instead.

The target applies to a layout field when that field is linked to the dictionary entry. To link a field, select the dictionary entry in the field's Field Name box in the Layout Editor.

Enable the dictionary field before you link it

Disabled fields do not appear in the Layout Editor when you search for Field Dictionary entries. To link a layout field to a dictionary entry and apply its target, make sure the dictionary field is enabled.

Edit multiple fields with a CSV file

1. Prepare a CSV file with a Display Name column that matches the existing fields and a Transcription Target Accuracy column.

2. In the Transcription Target Accuracy column, enter each target as a decimal between 0.001 and 0.999 (for example, 0.985 for 98.5%).

   • Leave the cell empty to set no target.

   • The CSV file uses decimals, but the Field Dictionary page and the flow settings use percentages.

3. Go to Library > Field Dictionary, click Add Field, and select the Import Fields tab.

4. Upload the CSV file, and select the option to update existing fields.

The Transcription Target Accuracy column in the Field Dictionary list shows each field's current target.

To learn more, see Editing Fields in the Field Dictionary.

Set a transcription target in a flow

1. Go to Flows, and click the name of the flow.

2. In Settings Type, select General Transcription.

3. Under Transcription Target Accuracy (Entry Level), click the pencil icon.

   • The setting shows how many entries in the flow have a custom target accuracy.

4. In Customize Transcription Target Accuracy, find the entry.

   • The dialog lists every field and table column in the layouts of the flow's release.

   • An icon next to each Entry Name shows whether the entry is a field or a table column.

5. To narrow the list, do one of the following:

   5. Enter a layout or entry name in Search.

   6. Click Filter, set Layout Name, Layout Type, Entry Name, or Entry Type, and click Apply.

      • To reset the filters, click Clear All.

6. In the entry's Target Accuracy cell, enter a percentage between 0.1 and 99.9.

   • You can enter up to one decimal place.

7. Click outside the cell to save the value.

The Field Dictionary Target Accuracy column shows the current Field Dictionary value for each Structured field. A value in Target Accuracy overrides it for this flow only.

To remove a custom target, clear the Target Accuracy cell. The entry then uses the next value in the order described in Which target accuracy applies to an entry section of this article.

Set an identification target in a flow

A value in a Semi-structured document can be incorrect because the system identified the wrong location or because it transcribed the value incorrectly. For the fields that matter most, set both an identification target and a transcription target, so that each value is both correctly located and correctly transcribed.

1. Go to Flows, and click the name of the flow.

2. In Settings Type, select Identification.

3. Under Identification Target Accuracy (Entry Level), click the pencil icon.

4. In Customize Identification Target Accuracy, find the entry. The dialog lists only entries in Semi-structured layouts, so it has no Layout Type column or filter and no Field Dictionary column.

5. To narrow the list, use Search, or click Filter, set Layout Name, Entry Name, or Entry Type, and click Apply.

6. In the entry's Target Accuracy cell, enter a percentage between 0.1 and 99.9, and click outside the cell to save the value. This value overrides Field Identification Target Accuracy or Table Identification Target Accuracy for the entry in this flow only.

Identification targets behave differently from transcription targets

• If a field's identification supervision setting in the layout is set to Auto-identify, the field is not sent to Identification Supervision based on accuracy. An entry-level target set for that field has no effect, even though the field still appears in Customize Identification Target Accuracy.

• When you test one document at several target accuracies, an entry sent to Supervision at one target accuracy may not be sent at every higher target accuracy. Identification selects combinations of confidence thresholds independently for each accuracy level.

• Across varied submissions over time, a higher target accuracy for an entry results in lower automation and more Supervision tasks for that entry.

To learn more, see Identification Settings.

Monitor entries with custom targets

• To view accuracy for each entry that has a custom target, go to Reporting > Accuracy.

• In System Transcription Sampled Errors, you can filter results by field or table column only if field-level targets are configured.

• Accuracy for an entry is calculated from the flow's normal QA sampling. FLAT does not add QA samples for entries with custom targets, so entries that appear less often in your documents may take longer to collect enough QA data to report on.

• To evaluate a target change, compare automation and accuracy across varied submissions over a period of time. Reprocessing a single document at different targets does not give reliable results, especially for identification.