CFE User Guide

When working with UIMA, feature extraction is usually implemented by writing a special UIMA component called a CAS Consumer that contains custom code for accessing the annotations and their attributes, outputting them to a file, memory or database as required. The CAS consumer contains explicit logic for traversing the object's structure and examining values of specific attributes. Also, the CAS consumer would likely have code for outputting the accessed values to a particular destination, as required by the application. Writing CAS consumers can be labor intensive and requires Java programming. While this approach allows powerful control and customization to an application's needs, supporting the code can become problematic, especially as application requirements change. This can have a negative effect on many different aspects of code support, such as maintenance, evolution, bug fixing, reusability etc.

CFE is a multipurpose tool that enables feature extraction from a UIMA CAS in a very generalized and application independent way. The extraction process is performed according to rules expressed using the Feature Extraction Specification Language (FESL) that are stored in configuration files. Using CFE eliminates the need for creating customized CAS consumers and writing Java code for every application. Instead, by using FESL rules in XML format, users can customize the information extraction process to suit their application. FESL's rule semantics allow the precise identification of the information that is required to be extracted by specifying precise multi-parameter criteria. The FESL syntax and semantics are defined further in this guide.

A "feature path" is a mechanism used by FESL to identify a particular feature (not necessarily a UIMA feature) of an annotation. The value associated with the feature, indicated by the feature path, can be either evaluated to match a certain criteria or extracted to the final output or both. The syntax of a feature path is an indexed sequence of attribute/method names separated by the colon character. Such a sequence mimics the sequence of Java method calls required to extract the feature value. For example, a value of the EngineAnnotation attribute Cylinders from Figure 2 can be written as CarAnnotation:Engine:Cylinders, where Engine is an attribute of CarAnnotation. The intermediate results of each step of the call sequence can be referred from different FESL structural elements by their zero-based index. For instance, the Parent Tag notation (see below) uses the index to access intermediate values. The feature path can be used to identify feature values that are either primitives or complex object types.

There are two different ways of using feature path notation to identify an object: full path and partial path. The object can be one of the following:

A full path specifies a path to an object starting from its type. For instance, if EngineAnnotation is specified as a full path, it would refer to all instances of annotations of that type. If CarAnnotation:Engine is specified, it would refer only to instances of the EngineAnnotation type that are attributes of instances of the CarAnnotation type. Full path notation is usually used for TA or FA identification.

A partial path specifies a path to an object starting from a previously located annotation object (whether TA or FA). For example, if an instance of CarAnnotation is located as a TA, then the size of its engine can be specified as Engine:Size. Partial path notation is usually used for specification of feature values that are being examined or extracted. The distinction between "full path" and "partial path" is very similar to the concepts of "absolute path" and "relative path" when discussing a computer's file system.

Each FESL rule is represented by a1 XML element with the tag targetAnnotation , as specified in the XSD by the TargetAnnotationXML type. Each element of this type is a composition of:

The TAM specifies search criteria for locating Target Annotations ( TA s), while FAM s contain criteria for locating Feature Annotations ( FA s) and the specification of features for extraction from the FA s. The criteria for the search and the features to be extracted are specified using the feature path notation, as explained earlier. The XML tags representing the matchers are detailed below.

Since UIMA annotations may have arrays as attributes, FESL provides the ability to perform feature extraction from array objects. In particular, going back to Figure 2, if the implementation for the Wheels attribute is a UIMA FSArray type, then using feature path notation:

The result of using toArray as an accessor is an array of values. FESL also provides syntax for accessing individual elements of arrays by index.

The specification of individual elements can be mixed for example: CarAnnotation:Wheels:toArray[0][2-3]:Diameter refers to all elements of Wheels attribute except the second. If the index specified falls outside the range of the matched data, a null value will be assigned.

If required, FESL allows sorting extracted features by an offset in the text of the annotations that these features are extracted from. For instance CarAnnotation:Wheels:toArray[sort]:Diameter would ensure such an order.

The parent tag is used to access a specific element of a feature path of a TA or FA by index. If a parent tag is used within a TAM specification, it is applied to the full path of the corresponding TA. Likewise, parent tags contained in FAMs are applied to the full a path of the corresponding FA. The tag consists of __p prefix followed by the index of an element that is being accessed. For instance, __p0 addresses the first element of a feature path. The tag can be a part of a feature path. For example, if a TA is specified as CarAnnotation:Wheels:toArray, corresponding to a concept of "wheels of a car" then the value of the Color attribute of a CarAnnotation object can be accessed by specifying __p0:Color. Such a specification can be used when it is required to examine/extract features of a containing annotation along with features of contained annotations. Samples of using parent tags are provided in the sections that detail FESL syntax, below.

CFE allows comparing feature values for equality to null. The root XML element CFEConfig has a string attribute nullValueImage that sets a literal representation of a null value. If an extracted feature value is null, it will be converted to a string that is assigned the nullValueImage attribute. The example below illustrates the usage of this attribute.

While all FAM specifications for a single TAM are independent from each other, there is an implicit dependency between TAMs. In particular, they are dependent on the order in which they are specified in a configuration file. Annotations corresponding to certain concepts that were identified by a TAM that appear earlier in the configuration file will be excluded from further processing by FESL. This rule only applies to TAMs that use the fullPath attribute in their specification (see PartialObjectMatcherXML ). Having the implicit exclusion helps to separate the processing of same type annotations in the case when these annotations have different semantic meaning. For instance, the set of features that is required to be extracted from annotations of type EngineAnnotation that are attributes of CarAnnotation objects can be different than a set of features that is required to be extracted from annotations of the same EngineAnnotation type that are attributes of some other type or are not attached to any annotations of other types. To implement such a behavior in FESL, the fist TAM would contain criteria for locating EngineAnnotation objects that are attached to objects of the CarAnnotation type, while the second TAM would not specify any restriction on containment of objects of the EngineAnnotation type. If such a specification is given, all EngineAnnotation objects located according to the rule in the first TAM will be excluded from further processing and, hence, will not be available for processing by rules given in the second TAM

The specification enables comparing a feature value to an integer bitmask. The feature value is considered to be matched if it is of an Integer type and:

Example:

<bitsetFeatureValues bitmask="3" exact_match="false" />

<bitsetFeatureValues bitmask="3" exact_match="true" />

The first line of the example specifies a test whether either of the two less significant bits of a feature value is set. To be successful, the test specified by the second line requires both less significant bits to be set.

EnumFeatureValuesXML element allow to test if a feature value belongs to a finite set of values. According to EnumFeatureValuesXML specification, if a feature value is equal to either one of the elements of values then the feature is considered to be successfully evaluated. The caseSensitive attribute indicates whether the comparison between the feature value and members of the values element is case sensitive. The FESL fragment below shows how to specify such a comparison:

<enumFeatureValues caseSensitive="true">

<values>red</values>

<values>green</values>

<values>blue</values>

</enumFeatureValues>

This fragment specifies a case sensitive comparison of a feature value to a set of strings: red, green and blue.

Special processing occurs when the array has only a single element that starts with file://, enabling the use of external dictionaries for comparison. In this case, the text within the values element is treated as a URI. The contents of the file referenced by the URI will be loaded and used as a set of values against which the feature value is going to be tested. The file should contain one dictionary entry per line, with each line starting with the # character considered to be a comment and thus will not be loaded. The dictionary handling is implemented in org.apache.uima.tools.cfe.EnumeratedEntryDictionary. The default implementation supports single token (whitespace separated) dictionary entries. If a more sophisticated dictionary format is desired, then either the constructor's parameters can be changed or methods for initializing and loading the dictionary from a file can be overridden.

According to ObjectPathFeatureValuesXML specification, the TA or FA itself (depending on whether this element is in TAM or in FAM ) is tested whether it is at the location defined by the objectPath. This ability to evaluate whether a feature belongs to some CAS object is useful specifically in the cases where a particular feature value is the property of several different objects. For instance, this element can be used when features from annotations should be extracted only if they are attributes of other annotations. The FESL fragment below specifies a test that checks if an object's full path is org.apache.uima.tools.cfe.sample.CarAnnotation:Wheels:toArray. Such a test, for instance, can be used to check if an instance of a WheelAnnotation belongs to an instance CarAnnotation:

<objectFeatureValues objectPath="org.apache.uima.tools.cfe.sample.CarAnotation:Wheels:toArray"b>

The PatternFeatureValuesXML element enables comparing a feature value against a regular expression specified by the pattern attribute using Java Regular Expression syntax and considered to be successfully evaluated if the value matches the pattern.

The FESL fragment below defines a test that checks if a feature value conforms to the hex number format:

<patternFeatureValues pattern="(0[Xx][0-9A-Fa-f]+)" />

According to RangeFeatureValuesXML specification the feature value is evaluated whether it is of a Comparable type and belongs to the interval specified by the attributes lowerBoundary and upperBoundary. The attributes lowerBoundaryInclusive and upperBoundaryInclusive indicate whether the corresponding boundaries should be included in the range for comparison. FESL fragment below specifies a test that checks if feature value is in the numeric range between 1 and 5, including 1 and excluding 5:

<rangeFeatureValues lowerBoundary="1.8" upperBoundaryInclusive="true" upperBoundary="3.0" />

The SingleFeatureMatcherXML defines rules for matching of a feature value to the featureValues element. The featureValues can be one of the elements in the bullet list above. The previous section detailed rules for matching a feature value to each of these elements. According to the specification for matching of a single feature value, first, a value of a feature denoted by the required featurePath attribute is located. For features that have arrays in their featurePath multiple values can be found. If such value(s) is found and optional featureTypeName attribute specifies a type name of the feature value, every found feature value is tested to be of that type. If the test is successful, then feature values are evaluated according to a specification given in featureValues. After the evaluation is performed a single feature is considered to be successfully evaluated if:

For SingleFeatureMatcherXML elements that are parts of TAM element only evaluation of feature values is performed. If a SingleFeatureMatcherXML element is a part of FAM then the feature value is output only if the quiet attribute is set to false. If the value of the quiet attribute is set to true, then, even if the feature is matched, only an evaluation is performed, but no value is written into the final output. A featurePath attribute uses feature path notation explained earlier.

FESL fragment below defines a test that checks if a value of the Size attribute is in a range defined by rangeFeatureVulues element:

<featureMatchers featurePath="Size" featureTypeName="java.lang.Float">

<rangeFeatureValues lowerBoundary="1.8" upperBoundaryInclusive="true" upperBoundary="3.0"/>

</featureMatchers>

In addition it is allowed to use the parent tag (see Parent tag ) in the featurePath attribute. A sample in the PartialObjectMatcherXML section detail on how use the parent tag notation.

This is a specification for matching a group of features. It can be applied to both types of annotations, TAs and FAs. Each element in featureMatchers is evaluated against either a TA or a FA annotation. The group is considered to be matched if:

The FESL fragment below defines a group with the two features Color and Wheels:Size to be matched. The entire group is to be successfully evaluated if both features are matched. The first feature is successfully evaluated if its value is one of the values listed by its enumFeatureValues element and the second feature is matched if its value is not in the set contained in its enumFeatureValues element, as specified by its exclude attribute. It should be noted that if the optional attribute featureTypeName is omitted then a feature value is assumed to be of a string type. Otherwise a feature value's type will be evaluated if it is the same or derived from the type specified by the featureTypeName attribute. Assuming the groupFeatureMatcher is specified for the CarAnnotation type, the test defined by a FESL fragment below is successful is a car is ether red, green or blue and it does not have 1 or 3 wheels:

<groupFeatureMatchers>

<featureMatchers featurePath="Color" featureTypeName="java.lang.Stting">

<enumFeatureValues caseSensitive="true">

<values>red</values>

<values>green</values>

<values>blue</values>

</enumFeatureValues>

</featureMatcher>

<featureMatchers featurePath="Wheels:Size" exclude="true">

<enumFeatureValues caseSensitive="true">

<values>1</values>

<values>3</values>

</enumFeatureValues>

</featureMatchers>

<groupFeatureMatchers>

This is a base specification for an annotation matcher that will search annotations of a type specified by annotationTypeName located on a path specified by fullPath. If fullPath is omitted or just contains the type name of an annotation (same as annotationTypeName attribute) then all instances of that type are considered for further feature value evaluation. If fullPath contains a path to an object from an attribute of a different object, then only instances of annotationTypeName that located on that path will be considered for further evaluation Once an annotation is successfully evaluated to match a type/path, its features are evaluated according to specification given in all elements of groupFeatureMatchers. If evaluation of any groupFeatureMatchers is successful or if no groupFeatureMatchers is given, then the annotation is considered to be successfully evaluated. The fullPath attribute should be specified using syntax described in the feature path section above, with the exception that it can not contain any parent tags. For instance, a specification where a value of the fullPath attribute is CarAnnotation:Engine and a value of the annotationTypeName is EngineAnnotation would address only engines that are car engines. PartialAnnotationMatcherXML is used to specify search rules in TAM specifications. To illustrate the use of parent tag notation let's consider an example where it is required to identify engines of blue cars that have a size more than 1.8 l but not greater then 3.0 l. According to a class diagram in Figure 2, the FESL fragment below defines rules for the task. It should be noted that the second feature matcher uses the parent tag notation to access a value of the CarAnnotation's attribute Color:

<targetAnnotationMatcher annotationTypeName="EngineAnnotation" fullPath="CarAnnotation:EngineAnnotation" >

<groupFeatureMatchers>

<featureMatchers featurePath="Size" featureTypeName="java.lang.Float">

<rangeFeatureValues lowerBoundary="1.8" upperBoundaryInclusive="true" upperBoundary="3.0"/>

</featureMatchers>

<featureMatchers featurePath="__p0:Color" featureTypeName="java.lang.String"

<enumFeatureValues caseSensitive="true">

<values>red</values>

<values>green</values>

<values>blue</values>

</enumFeatureValues>

</featureMatcher>

<groupFeatureMatchers>

</targetAnnotationMatcher>

extends PartialAnnotationMatcherXML

The FeatureObjectMatcherXML element contains rules that specify how FeatureAnnotations (FA) should be located and which features should be extracted from them. It inherits its properties from PartialObjectMatcherXML. In addition it has semantics for specifying:

It is done by using boolean attributes windowsizeLeft, windowsizeInside, windowsizeRight, windowsizeEnclosed and the bitmask windowFlags attribute that indicate FA's search rules:

The location of a FA is included in the generated output according to optional orientation and distance attributes. For example, if values of both of these attributes are set to true and the FA is a first annotation of required type to the left from TA, then the generated feature value will start with the prefix L1. If the values are set to false, then the feature value's prefix will be X0. This allows generating unique feature names for model building and evaluation for machine learning.

FeatureObjectMatcherXML is used to specify search rules in FAM specifications.

The FESL fragment below adds rules to the previous sample to extract a number of cylinders from engines of cars whose wheels diameter is at least 20.0":

<targetAnnotationMatcher annotationTypeName="EngineAnnotation" fullPath="CarAnnotation:EngineAnnotation" >

<groupFeatureMatchers>

<featureMatchers featurePath="Size" featureTypeName="java.lang.Float">

<rangeFeatureValues lowerBoundary="1.8" upperBoundaryInclusive="true" upperBoundary="3.0"/>

</featureMatchers>

<featureMatchers featurePath="__p0:Color" featureTypeName="java.lang.String">

<enumFeatureValues caseSensitive="true">

<values>red</values>

<values>green</values>

<values>blue</values>

</enumFeatureValues>

</featureMatcher>

<groupFeatureMatchers>

</targetAnnotationMatcher>

<featureAnnotationMatcher annotationTypeName="EngineAnnotation" fullPath="CarAnnotation:EngineAnnotation" windowsizeInside=1 >

<groupFeatureMatchers>

<featureMatchers featurePath="__p0:Wheels:toArray:Diameter" featureTypeName="java.lang.Float" quiet="true" >

<rangeFeatureValues lowerBoundary="20.0" lowerBoundaryInclusive="true"/>

</featureMatcher>

<featureMatchers featurePath="Cylinders" featureTypeName="java.lang.Float" />

<groupFeatureMatchers>

</featureAnnotationMatcher>

This is a root specification for a class (group) of annotations of all extracted instances, which are assigned the same label (className) in the final output. The label can be a literal string or a feature path in curly brackets or a combination of the two (i.e. SomeText_{__p0:SomeProperty}). If using a feature path in a class name label it is required to use the parent tag notation. In such a case the parent tag refers to the TA specified by the targetAnnotationMatcher element. Annotations that belong to the group are searched within a span of enclosingAnnotation according to the specification given in the targetAnnotationMatcher (TAM) and features from matched annotations are extracted according to specification given in featureAnnotationMatchers (FAM). In general, the annotation that features are extracted from could be different from annotations that are matched during the search This is useful when extracting features for machine learning model building and evaluation where features are selected from annotations that could be located in a specific location relatively to the annotation that satisfy a search criteria. For instance, POS tags of 5 words to the left and right from a specific word. Only if an annotation is successfully evaluated (matched) by a TAM further feature extraction is allowed and rules specified by corresponding FAMs are executed.

The sample configuration file below has been created for extracting features in order to build models for a machine learning application. The type system for this sample defines several UIMA annotation types:

pennTag: String - POS tag of a token

Code: String - specific code assigned to a named entity

SemanticClass: String - semantic class of a named entity

Tokens: FSArray - array of org.apache.uima.tools.cfe.sample.Token annotations, ordered by their offset, that are included in the named entity

The classification task is defined as follows:

To build a model for machine learning it is required to extract features from surrounding tokens for all classes listed above. In particular the following features are required to be extracted:

Line 1 - a standard XML declaration that defines the XML version of the document and its encoding

Line 2, 87 - FESL root element that references the schema and defines global variables, such as nullValueImage (see Null values )

Line 3-32 - rules for extracting features for first tokens of named entities.

Line 3 - extracted features for those tokens are assigned a composite label that includes prefix CMBegin_ pl s a value of a Code attribute of the first element of the TA's path. The search for FA is going to be performed within boundaries of enclosing org.apache.uima.tools.cfe.sample.Sentence annotation

Line 4-12 - TAM that defines rules for identifying the fist TA

Line 4 - defines TA's type (org.apache.uima.tools.cfe.sample.Token) and a full path to it (org.apache.uima.tools.cfe.sample.NamedEntity:Tokens:toArray[0]). According to this path notion, the CFE will:

Line 5-11 - defines rules for matching a group of features for TA

Line 6-10 - defines rules for matching a feature for this group

Line 6 - defines that the feature value is of the type java.lang.String and has the feature the path __p0:SemanticClass, which translates to a value of the attribute SemanticClass of the first element of the TA's path (org.apache.uima.tools.cfe.sample.NamedEntity)

Line 7-9 - defines an explicit list of values that the feature value should be in

Line 8 - defines the value Car Maker as the only possible value for the feature

Line 13-17 - FAM that defines rules for identifying first FA and its feature extraction

Line 13 - defines FA's type to be org.apache.uima.tools.cfe.sample.Token; the attribute windowsizeInside with the value 1 tells CFE to extract features from TA itself (TA=FA) and setting orientation and distance attributes to true tells CFE to include position information into the generated feature value

Line 14-16 - defines rules for matching a group of features for the first FA.

Line 15 - defines rules for matching the only feature for this group of the type java.lang.String and with feature path coveredText that eventually will be translated by CFE to a method call of a org.apache.uima.tools.cfe.sample.Token annotation object; according to this specification the feature value will be unconditionally extracted

Line 18-31 - FAM that defines rules for identifying second type of FA and its feature extraction

Line 18 - defines FA's type to be org.apache.uima.tools.cfe.sample.Token; the attributes windowsizeLeft and windowsizeRight with the values 5 tell CFE to extract features from 5 nearest annotations of this type to the left and to the right from TA and having orientation and distance attributes set to true tells CFE to include position information into the generated feature value.

Line 19-30 - defines rules for matching a group of features for the second FA.

Line 20 - defines rules for matching the first feature of the group to be of the type java.lang.String and with the feature path coveredText that eventually will be translated by CFE to a method call of a org.apache.uima.tools.cfe.sample.Token annotation object; according to this specification the feature value will be unconditionally extracted

Line 21-29 - define rules for matching the second feature of the group

Line 21 - defines rules for matching the second feature of the group to be of the type java.lang.String and with the feature path pennTag that eventually will be translated by CFE to getPennTag method call of a org.apache.uima.tools.cfe.sample.Token annotation object; according to this specification the feature will be evaluated against enumFeatureValues and, as the exclude attribute is set to true:

As the quiet attribute is set to true, the feature value extracted by the second matcher will not be added to the generated for this FA output

Line 22-28 - defines an explicit list of values that the value of the second feature should be in

Line 23-27 - defines values IN, CC, DT, CD, null as possible values for the second feature; if the feature value is equal to one of these values, evaluation of the enclosing feature matcher is successful; if the feature value is null it will be converted to the string defined by nullValueImage (null as set in line 2 of this sample) and as null is one of the list's elements, it will be successfully evaluated.

Line 34-63 - rules for extracting features for all tokens of named entities except the first. These rules are the same as the rules defined for first tokens of named entities (lines 3-32) with the following exceptions:

Line 34 - defines that TAs matched by these rules will be assigned a composite label that includes prefix CMInside_ plus a value of the Code attribute of a first element of the TA's path

Line 35 - sets the fullPath attribute to org.apache.uima.tools.cfe.sample.NamedEntity:Tokens:toArray that can be translated as any token of a named entity, but because of implicit TA exclusion , the TAs that were matched for first tokens of named entities by the rules for previous TAM are not included into the set of TAs that will be evaluated by rules for this TAM

Line 65-86 - rules for extracting features for all tokens other than tokens of named entities. These rules are the same as the rules defined for previous categories with the following exceptions:

Line 65 - defines that TAs matched by the enclosed rules will be assigned the string label Other_token

Line 66 - only defines a type of TAs that should be processed by the corresponding TAM without fullPath attribute. Such a notation can be translated as all tokens, but because of the implicit TA exclusion , the TAs, which were matched for tokens of named entities by rules defined by the previous TAMs, are not included into the set of TAs that will be evaluated by rules for this TAM. So, the actual translation will be all tokens other than tokens of named entities.