Field Class

Can be one of the following types:

The Value Type property determines the data type and formatting rules for values produced by this Field Class.

• Choose a Storage Type that matches the kind of data you expect to extract (such as text, number, date, or boolean).
• Each storage type exposes its own set of configuration options, accessible by expanding the property in the property grid.
• Any extracted value that cannot be converted to the selected Storage Type will be discarded.

Typical Usage

• Use a date storage type to ensure only valid dates are output.
• Use a numeric storage type to restrict results to numbers and apply formatting or validation.
• Use a text storage type for general string values, with optional delimiters for multi-value fields.

> Tip:
> Setting the correct value type helps ensure data quality and consistency in downstream processing.

Can be one of the following types:

The Value Extractor property defines the extraction logic for locating all possible values of interest on a document.

• Select an extractor type (such as pattern match, list match, OMR, barcode, or zonal extraction) to match your data and document structure.
• After selecting a type, configure its settings using the dynamic panel or property grid.
• The extractor should be set up to return all possible candidates for the value you want to classify (for example, all dates on a page).

Extractor Type Categories

• Text Parsing Extractors:
  Use regular expressions or value lists to extract text-based values.
• OMR Extractors:
  Extract values from checkboxes or bubbles using optical mark recognition.
• Barcode Extractors:
  Read values encoded in barcodes.
• Zonal Extractors:
  Extract text from a defined region of the page.
• Reference Extractor:
  Reuse extraction logic from another Value Reader, Data Type, or Field Class.

> Note:
> Once the extractor type is changed from 'none', the property must be reset before another type can be selected.

Best Practices

• Configure the extractor to return all possible candidates, even if some are not correct; the classifier will select the best match based on context.
• Use the 'Tester' tab to validate and refine your extractor configuration before training the classifier.

Can be one of the following types:

The Feature Extractor property defines how contextual features (such as words, phrases, or labels) are identified near each candidate value found by the 'Value Extractor'. These features are used to train the classifier and to score candidates during extraction.

• Select an extractor type that returns the kind of features most useful for distinguishing the correct value (for example, general words, labels, or specific patterns).
• Configure the extractor to return features that appear in the context of the value you want to classify.
• The context in which features are collected is determined by the 'Context Scope' property (e.g., flow, zonal, nearest, or self).

Usage Example

• For a contract where you want to identify the "Effective Date", configure the feature extractor to return general words (excluding stop words). During training, features near the correct date (such as "lease", "entered", "between") will be positively weighted, while features near incorrect dates (such as "clerk", "filed", "archives") will be negatively weighted.

Best Practices

• Choose a feature extractor that best captures the distinguishing context for your use case.
• Use the 'Tester' tab to review which features are being returned and refine the extractor as needed.
• Remember that the scope and number of features included are controlled by properties such as 'Context Scope' and 'Max Instances'.
• Use a Lexicon of stop words to filter out words which are not helpful in classification (of, the, etc.)

> Note:
> Once the extractor type is changed from 'none', the property must be reset before another type can be selected.

Can be one of the following values:

Overview

The ContextScopes enumeration controls the method used to gather context features for each candidate value in a Field Class. The context scope determines the spatial or logical relationship between a value and its surrounding features, which are used for machine learning-based classification and extraction.

Choosing the right context scope is essential for accurately identifying the correct value among multiple candidates, especially in documents where similar values may appear in different locations or contexts.

Available Modes

• Zonal: Extracts features from one or more rectangular regions (zones) defined relative to each value.
• Flow: Extracts features that occur before and/or after the value in the text flow (reading order).
• Self: Extracts features that are contained within the value itself.
• Nearest: Extracts a limited number of features that are spatially closest to the value, with optional direction and distance filtering.

Usage Guidance

• Use Zonal for semi-structured documents where labels or key features consistently appear in fixed locations relative to the value (e.g., invoices).
• Use Flow for natural language or unstructured documents where context is determined by surrounding words or phrases (e.g., contracts).
• Use Self when the features that define a value are contained within it (e.g., paragraphs or clauses).
• Use Nearest for flexible, proximity-based context extraction, especially when document layouts vary or context is not strictly zonal or flow-based.

The selected context scope affects which additional properties are visible and configurable, such as 'Context Zones', 'Max Instances', 'Geo Filter', 'Flow Filter', 'Max Distance', and 'GeoTag Features'.

> Tip:
> Experiment with different context scopes and review the extracted features in the 'Tester' tab to determine which mode best captures > the distinguishing context for your use case.

For more information, see the documentation for Field Class, Value Extractor, and related context extraction properties.

The Max Instances property limits how many nearby context features are considered for each candidate value during extraction and classification.

• When using a 'Context Scope' of Flow or Nearest, this property controls the maximum number of features (such as words, labels, or patterns) that will be included from the surrounding context.
• Features are selected based on their proximity to the candidate value, either in text flow order or spatial distance, depending on the context scope.
• Reducing this value can help focus the classifier on the most relevant features and improve performance, especially on documents with dense or noisy data.

Usage Example

• If set to 5, only the 5 closest features (by flow or distance) will be used for each candidate value.

> Note:
> This property is only applicable when 'Context Scope' is set to Flow or Nearest.

A combination of the following flags:

Overview

The CompassDirection enumeration is used in Field Class objects to filter context features by their geometric direction relative to a candidate value. This is only applicable when the 'Context Scope' property is set to Nearest. It enables fine-grained spatial filtering, allowing you to include or exclude features based on their position (such as above, below, left, right, or diagonal) with respect to the value being classified.

Multiple directions can be combined using bitwise flags. For example, to include only features above or to the left of a value, use:

The special value All includes features from all directions.

Usage Examples

• To focus on likely label positions, include only features above (North) or to the left (West) of a value.
• To reduce noise, exclude features in certain directions (for example, ignore features below a value).
• Combine with the 'Max Distance' property to restrict the spatial region considered for context features.

Special Behaviors

• Features on a previous page are treated as north; features on a subsequent page are treated as south.
• When 'GeoTag Features' is enabled, features are tagged with their direction (for example, (n) for north).

Related Properties

• 'Geo Filter': Specifies which directions to include when extracting context features in Nearest mode.
• 'Max Distance': Limits the maximum distance (in inches) for included features.
• 'GeoTag Features': When enabled, appends a direction tag to each feature.

For more information, see the documentation for Field Class, 'Context Scope', and related extraction settings.

Can be one of the following values:

Overview

The FlowDirection enumeration is used in Field Class objects to filter context features by their position in the text flow relative to a candidate value. This is only applicable when the 'Context Scope' property is set to Flow. It enables filtering of features that occur before, after, or on either side of the value in the document's reading order.

This is useful for natural language documents, such as contracts or unstructured text, where the meaning of a value is often determined by the words or phrases that appear before or after it in the text.

Multiple directions can be combined using bitwise flags. For example, to include features both before and after a value, use:

Usage Examples

• To focus on features that precede a value (such as labels or introductory phrases), use Before.
• To focus on features that follow a value (such as units, qualifiers, or trailing context), use After.
• To include features on both sides, use Either.

Special Behaviors

• When 'GeoTag Features' is enabled, features are tagged with (b) for before and (a) for after.
• The 'Max Instances' property limits the number of features included from the selected directions.

Related Properties

• 'Flow Filter': Specifies which directions to include when extracting context features in Flow mode.
• 'Max Instances': Limits the maximum number of features included.
• 'GeoTag Features': When enabled, appends a direction tag to each feature.

For more information, see the documentation for Field Class, 'Context Scope', and related extraction settings.

The Context Zones property specifies one or more spatial regions (rectangles) around each located value. Features found within these zones are used for training and classification.

• Each zone is defined by relative coordinates (in inches) to the candidate value's bounding box.
• Multiple zones can be configured to capture features from different areas (e.g., left, right, above, or below the value).
• This is especially useful for semi-structured documents where labels or key features consistently appear in predictable locations relative to the value.

Usage Example

• To extract features from a 2-inch region to the left of a value, add a zone with left = -2, right = 0, top = -0.1, bottom = 0.1.

> Note:
> This property is only applicable when 'Context Scope' is set to Zonal.

Can be one of the following types:

The Exclusion Extractor property allows you to specify a Value Extractor that identifies features which, if present within any context zone of a candidate value, will cause that value to be excluded from the results.

• Use this to filter out values that are near unwanted or disqualifying features (such as "void", "sample", or other exclusionary terms).
• The exclusion extractor is run for each candidate value, and if any matches are found within the defined context zones, the candidate is removed from consideration.

Usage Example

• To prevent extraction of values labeled as "void", configure an exclusion extractor that matches the word "void" and add a context zone covering the label area.

> Note:
> This property is only applicable when 'Context Scope' is set to Zonal.

The Max Distance property limits how far away (in inches) a context feature can be from a candidate value to be considered for training or classification.

• Only features within this distance from the center of the candidate value will be included.
• Setting this value helps focus the classifier on features that are most likely to be relevant, reducing noise from distant or unrelated features.
• A value of 0 means there is no distance limit; all features, regardless of distance, may be considered.

Usage Example

• Set to 1.0 to only include features within 1 inch of the candidate value.

> Note:
> This property is only applicable when 'Context Scope' is set to Nearest.

The GeoTag Features property controls whether each context feature is annotated with a suffix that describes its spatial or flow-based relationship to the candidate value.

• When enabled, features are tagged with a code (such as (b), (a), (n), (s), (e), (w), (ne), (nw), (se), or (sw)) to indicate their position:
  • In Flow context scope, (b) means the feature is before the value in text flow, (a) means after.
  • In Zonal or Nearest context scopes, tags indicate geometric direction: (n) for north/above, (s) for south/below, (e) for east/right, (w) for west/left, and combinations for diagonals.
  • In Self context scope, this property is not applicable (features are always within the value).
• Tagging features with their relative position can improve classifier accuracy by distinguishing the same feature appearing in different locations.
• This is especially useful when the meaning of a feature depends on its position (e.g., a label to the left vs. above a value).

Usage Example

• If the feature "lease" appears before a value in flow, it will be stored as "lease (b)".
• If "total" appears above a value in zonal context, it will be stored as "total (n)".

> Note:
> This property is only available when 'Context Scope' is not set to Self.

The Minimum Feature Count property sets a threshold for the number of context features that must be present near a candidate value.

• If a candidate value does not have at least this many context features, it will be excluded from training and output.
• This helps ensure that only values with sufficient contextual evidence are considered, improving the quality and reliability of the classifier.
• Setting this property to 0 disables the requirement, allowing values with no context features to be included.

Usage Example

• Set to 3 to require at least three context features for a value to be eligible for training or extraction.

> Tip:
> Increase this value to reduce noise from weakly supported candidates, especially in documents with sparse or inconsistent context.

The Training Threshold property controls when features not near a positively trained value are treated as negative examples.

• At 0%, all features not near a positive value during training are treated as negative, strengthening the classifier's ability to distinguish correct from incorrect values.
• At 100%, no features are trained negatively; only positive associations are learned.
• Intermediate values allow for more nuanced control, which can be useful when similar context appears near both correct and incorrect values.

Usage Example

• Set to 0.0 to maximize negative training (default for most scenarios).
• Set to 1.0 (100%) to disable negative training, useful when context is ambiguous or similar between candidates.

> Tip:
> Use a higher threshold if negative training is causing confusion due to highly similar context among candidates.

The Use Class Frequency property changes the feature weighting model from standard TF-IDF to TF-IDF-CF, where class frequency (the number of classes a feature appears in) is considered.

• This can improve classification accuracy in scenarios where features are common across multiple classes.
• When enabled, features that appear in many classes are weighted differently than those that are unique to a single class.
• For more details, see:
  An improvement of TF-IDF weighting in text categorization (Mingyong Liu and Jiangang Yang, 2012).

Usage Example

• Enable when your training data contains features that are shared among several classes and you want to reduce their influence.

> Note:
> Changing this property will reset the classifier and require retraining.

The Sublinear TF Scaling property modifies the calculation of term frequency in the TF-IDF model by applying a logarithmic scale.

• This reduces the impact of very frequent features, preventing them from dominating the classification decision.
• Useful when repeated occurrences of a feature should not be linearly more significant (e.g., 20 occurrences are not 20 times as important as one).
• Disabling this property uses the standard (linear) term frequency calculation.

Usage Example

• Enable to dampen the effect of highly repetitive features in your documents.

> Tip:
> Use in combination with stop word filtering and other feature selection techniques for best results.

The Smooth IDF property determines how the inverse document frequency (IDF) is calculated for features:

• When disabled, standard IDF is used: IDF = log(Classes / ClassesWithFeature). Features present in all classes have an IDF of 0.
• When enabled, smoothing is applied: IDF = log(1 + Classes / ClassesWithFeature). This prevents the IDF from reaching zero, even for ubiquitous features.
• Smoothing can improve stability and prevent rare features from being over-weighted, especially in small or unbalanced training sets.

Usage Example

• Enable to avoid extreme weighting for features that appear in every class, or when working with a small number of classes.

> Note:
> Changing this property will reset the classifier and require retraining.

The Minimum Confidence property sets a threshold for the classifier's confidence score. Only candidate values with a confidence score greater than or equal to this value will be output.

• Confidence is calculated based on the similarity between the candidate's context features and the trained model.
• Use this property to filter out low-confidence results and improve data quality.
• Setting this value to 0 will include all candidates, regardless of confidence.

Usage Example

• Set to 0.8 to only output candidates with at least 80% confidence.

> Tip:
> Adjust this value to balance recall and precision based on your application's requirements.

Can be one of the following values:

The Collation Type determines how multiple qualifying results from a Field Class are presented in the output.

> Tip:
> The available collation method may affect which additional properties are visible, such as 'Separator' or 'Maximum Page Distance'.

The Separator property specifies the delimiter to use when the collation method is set to Combine.

• This string is inserted between each value when multiple results are joined into a single output.
• Common separators include a space, comma, semicolon, or custom string.
• Has no effect unless the collation method is Combine.

Usage Example

• Set to ", " to join values as "A, B, C".
• Set to "\n" to join values on separate lines.

> Note:
> Choose a separator that does not conflict with the content of your values.

The Maximum Page Distance property limits how far apart (in page numbers) candidate values can be to be included together when the collation method is Combine.

• Only values within this page distance from the first qualifying instance will be combined.
• Set to 0 to allow combining values from any page.
• Useful for restricting combined results to a single page or a limited range of pages.

Usage Example

• Set to 1 to combine values only if they are on the same page or the next page.

> Note:
> This property is only applicable when the collation method is Combine.

Can be one of the following values:

The Sort Order enumeration specifies how a collection of Data Instance results will be sorted for output, display, or further processing.

Sorting can be based on position, frequency, confidence, extractor, value length, value content, or geometric coordinates.

Can be one of the following values: