This chapter is organized as follows. Following this introduction, section introduces the elements fs and f, used to represent feature structures and features respectively, together with the elementary binary feature value. Section introduces elements for representing other kinds of atomic feature values such as symbolic, numeric, and string values. Section introduces the notion of predefined libraries or groups of features or feature values along with methods for referencing their components. Section introduces complex values, in particular feature-structures as values, thus enabling feature structures to be recursively defined. Section discusses other complex values, in particular values which are collections, organized as sets, bags, and lists. Section discusses how the operations of alternation, negation, and collection of feature values may be represented. Section discusses ways of representing underspecified, default, or uncertain values. Section discusses how analyses may be linked to other parts of an encoded text. Section describes the feature system declaration, a construct which provides for the validation of typed feature structures. Formal definitions for all the elements introduced in this chapter are provided in section .

The fundamental elements used to represent a feature structure analysis are f (for feature), which represents a feature-value pair, and fs (for feature structure), which represents a structure made up of such feature-value pairs. The fs element has an optional type attribute which may be used to represent typed feature structures, and may contain any number of f elements. An f element has a required name attribute and an associated value. The value may be simple: that is, a single binary, numeric, symbolic (i.e. taken from a restricted set of legal values), or string value, or a collection of such values, organized in various ways, for example, as a list; or it may be complex, that is, it may itself be a feature structure, thus providing a degree of recursion. Values may be under-specified or defaulted in various ways. These possibilities are all described in more detail in this and the following sections.

Feature and feature-value representations (including feature structure representations) may be embedded directly at any point in an XML document, or they may be collected together in special-purpose feature or feature-value libraries. The components of such libraries may then be referenced from other feature or feature-value representations, using the feats or fVal attribute as appropriate.

We begin by considering the simple case of a feature structure which contains binary-valued features only. The following three XML elements are needed to represent such a feature structure: The attributes feats and the fVal are not discussed in this section: they provide an alternative way of indicating the content of an element, as further discussed in section .

An fs element containing f elements with binary values can be straightforwardly used to encode the matrices of feature-value specifications for phonetic segments, such as the following for the English segment [s]. +--- ---+ | consonantal + | | vocalic - | | voiced - | | anterior + | | coronal + | | continuant + | | strident + | +--- ---+

This representation may be encoded in XML as follows: Note that fs elements may have an optional type attribute to indicate the kind of feature structure in question, whereas f elements must have a name attribute to indicate the name of the feature. Feature structures need not be typed, but features must be named. Similarly, the fs element may be empty, but the f element must have (or reference) some content.

The restriction of specific features to specific types of values (e.g. the restriction of the feature strident to a binary value) requires additional validation, as does any restriction on the features available within a feature structure of a particular type (e.g. whether a feature structure of type phonological segment necessarily contains a feature voiced). Such validation may be carried out at the document level, using special purpose processing, at the schema level using additional validation rules, or at the declarative level, using an additional mechanism such as the feature-system declaration discussed in .

Although we have used the term binary for this kind of value, and its representation in XML uses values such as true and false (or, equivalently, 1 and 0), it should be noted that such values are not restricted to propositional assertions. As this example shows, this kind of value is intended for use with any binary-valued feature.

Features may take other kinds of atomic value. In this section, we define elements which may be used to represent: symbolic values, numeric values, and string values. The module defined by this chapter allows for the specification of additional datatypes if necessary, by extending the underlying class model.featureVal.single. If this is done, it is recommended that only the basic W3C datatypes should be used; more complex datatyping should be represented as feature structures.

The symbol element is used for the value of a feature when that feature can have any of a small, finite set of possible values, representable as character strings. For example, the following might be used to represent the claim that the Latin noun form mensas (tables) has accusative case, feminine gender, and plural number:

More formally, this representation shows a structure in which three features (case, gender, and number) are used to define morpho-syntactic properties of a word. Each of these features can take one of a small number of values (for example, case can be nominative, genitive, dative, accusative, etc.) and it is therefore appropriate to represent the values taken in this instance as symbol elements. Note that, instead of using a symbolic value for grammatical number, one could have named the feature singular or plural and given it an appropriate binary value, as in the following example: Whether one uses a binary or symbolic value in situations like this is largely a matter of taste.

The string element is used for the value of a feature when that value is a string drawn from a very large or potentially unbounded set of possible strings of characters, so that it would be impractical or impossible to use the symbol element. The string value is expressed as the content of the string element, rather than as an attribute value. For example, one might encode a street address as follows: 3418 East Third Street

The numeric element is used when the value of a feature is a numeric value, or a range of such values. For example, one might wish to regard the house number and the street name as different features, using an encoding like the following: East Third Street

If the numeric value to be represented falls within a specific range (for example an address that spans several numbers), the max attribute may be used to supply an upper limit: East Third Street

It is also possible to specify that the numeric value (or values) represented should (or should not) be truncated. For example, assuming that the daily rainfall in mm is a feature of interest for some address, one might represent this by an encoding like the following: This represents any of the infinite number of numeric values falling between 0 and 1.3; by contrast represents only two possible values: 0 and 1.

Some communities of practice, notably those with a strong computer-science bias, prefer to dissociate the information on the value of the given feature from the specification of the data type that this value represents. In such cases, feature values can be provided directly as textual content of f, with the assumption that the data type is specified by the schema. The following is an example taken from ISO 24612, presenting the symbolic values for Active Voice and Simple Present Tense in the untyped form:active SimPre

As noted above, additional processing is necessary to ensure that appropriate values are supplied for particular features, for example to ensure that the feature singular is not given a value such as symbol value="feminine"/. There are two ways of attempting to ensure that only certain combinations of feature names and values are used. First, if the total number of legal combinations is relatively small, one can predefine all of them in a construct known as a feature library, and then reference the combination required using the feats attribute in the enclosing fs element, rather than give it explicitly. This method is suitable in the situation described above, since it requires specifying a total of only ten (5 + 3 + 2) combinations of features and values. Similarly, to ensure that only feature structures containing valid combinations of feature values are used, one can put definitions for all valid feature structures inside a feature value library (so called, since a feature structure may be the value of a feature). A total of 30 feature structures (5 × 3 × 2) is required to enumerate all the possible combinations of individual case, gender and number values in the preceding illustration. We discuss the use of such libraries and their representation in XML further in section below.

However, the most general method of attempting to ensure that only legal combinations of feature names and values are used is to provide a feature-system declaration discussed in .

Whether at the level of feature-system declarations, feature- and feature-value libraries, or individual features, it is possible to align both feature names and their values with standardized external data category repositories such as ISOcat. See section for more discussion of the need and rationale for ISOcat references. In the following example, both the feature part_of_speech and its value #commonNoun are aligned with the respective definitions provided by ISO DCR (Data Category Registry), as implemented by ISOcat.

As the examples in the preceding section suggest, the direct encoding of feature structures can be verbose. Moreover, it is often the case that particular feature-value combinations, or feature structures composed of them, are re-used in different analyses. To reduce the size and complexity of the task of encoding feature structures, one may use the feats attribute of the fs element to point to one or more of the feature-value specifications for that element. This indirect method of encoding feature structures presumes that the f elements are assigned unique xml:id values, and are collected together in fLib elements (feature libraries). In the same way, feature values of whatever type can be collected together in fvLib elements (feature-value libraries). If a feature has as its value a feature structure or other value which is predefined in this way, the fVal attribute may be used to point to it, as discussed in the next section. The following elements are used for representing feature libraries and feature-value libraries:

For example, suppose a feature library for phonological feature specifications is set up as follows.

Then the feature structures that represent the analysis of the phonological segments (phonemes) /t/, /d/, /s/, and /z/ may be defined as follows.

The preceding are but four of the 128 logically possible fully specified phonological segments using the seven binary features listed in the feature library. Presumably not all combinations of features correspond to phonological segments (there are no strident vowels, for example). The legal combinations, however, can be collected together, each one represented as an identifiable fs element within a feature-value library, as in the following example:

Once defined, these feature structure values can also be reused. Other f elements may invoke them by reference, using the fVal attribute; for example, one might use them in a feature value pair such as: rather than expanding the hierarchy of the component phonological features explicitly.

Feature structures stored in this way may also be associated with the text which they are intended to annotate, either by a link from the text (for example, using the TEI global ana attribute), or by means of stand-off annotation techniques (for example, using the TEI link element): see further section below.

Note that when features or feature structures are linked to in this way, the result is effectively a copy of the item linked to into the place from which it is linked. This form of linking should be distinguished from the phenomenon of structure-sharing, where it is desired to indicate that some part of an annotation structure appears simultaneously in two or more places within the structure. This kind of annotation should be represented using the vLabel element, as discussed in below.

Features may have complex values as well as atomic ones; the simplest such complex value is represented by supplying a fs element as the content of an f element, or (equivalently) by supplying the identifier of an fs element as the value for the fVal attribute on the f element. Structures may be nested as deeply as appropriate, using this mechanism. For example, an fs element may contain or point to an f element, which may contain or point to an fs element, which may contain or point to an f element, and so on.

To illustrate the use of complex values, consider the following simple model of a word, as a structure combining surface form information, a syntactic category, and semantic information. Each word analysis is represented as a fs type='word' element, containing three features named surface, syntax, and semantics. The first of these has an atomic string value, but the other two have complex values, represented as nested feature structures of types category and act respectively: love

This analysis does not tell us much about the meaning of the symbols verb or transitive. It might be preferable to replace these atomic feature values by feature structures. Suppose therefore that we maintain a feature-value library for each of the major syntactic categories (N, V, ADJ, PREP):

This library allows us to use shortcut codes (N, V, etc.) to reference a complete definition for the corresponding feature structure. Each definition may be explicitly contained within the fs element, as a number of f elements. Alternatively, the relevant features may be referenced by their identifiers, supplied as the value of the feats attribute, as in these examples: <!-- ... --> <fs xml:id="ADJ" type="adjective" feats="#F1 #F2"/> <fs xml:id="PREP" type="preposition" feats="#F1 #F3"/> <!-- ... -->

This ability to re-use feature definitions within multiple feature structure definitions is an essential simplification in any realistic example. In this case, we assume the existence of a feature library containing specifications for the basic feature categories like the following:

With such libraries in place, and assuming the availability of similarly predefined feature structures for transitivity and semantics, the preceding example could be considerably simplified: love

Although in principle the fVal attribute could point to any kind of feature value, its use is not recommended for simple atomic values.

Sometimes the same feature value is required at multiple places within a feature structure, in particular where the value is only partially specified at one or more places. The vLabel element is provided as a means of labelling each such re-entrancy point:

For example, suppose one wishes to represent noun-verb agreement as a single feature structure. Within the representation, the feature indicating (say) number appears more than once. To represent the fact that each occurrence is another appearance of the same feature (rather than a copy) one could use an encoding like the following:

In the above encoding, the features named vb-num and nm-num exhibit structure sharing. Their values, given as vLabel elements, are understood to be references to the same point in the feature structure, which is labelled by their name attribute.

The scope of the names used to label re-entrancy points is that of the outermost fs element in which they appear. When a feature structure is imported from a feature value library, or referenced from elsewhere (for example by using the fVal attribute) the names of any sharing points it may contain are implicitly prefixed by the identifier used for the imported feature structure, to avoid name clashes. Thus, if some other feature structure were to reference the fs element given in the example above, for example in this way: then the labelled points in the example would be interpreted as if they had the name NVAL1.

Complex feature values need not always be represented as feature structures. Multiple values may also be organized as sets, bags or multisets, or lists of atomic values of any type. The vColl element is provided to represent such cases:

A feature whose value is regarded as a set, bag, or list may have any positive number of values as its content, or none at all, (thus allowing for representation of the empty set, bag, or list). The items in a list are ordered, and need not be distinct. The items in a set are not ordered, and must be distinct. The items in a bag are neither ordered nor distinct. Sets and bags are thus distinguished from lists in that the order in which the values are specified does not matter for the former, but does matter for the latter, while sets are distinguished from bags and lists in that repetitions of values do not count for the former but do count for the latter.

If no value is specified for the org attribute, the assumption is that the vColl defines a list of values. If the vColl element is empty, the assumption is that it represents the null list, set, or bag.

To illustrate the use of the org attribute, suppose that a feature structure analysis is used to represent a genealogical tree, with the information about each individual treated as a single feature structure, like this: Daniel Edouard

In this example, the vColl element is first used to supply a list of name feature values, which together constitute the forenames feature. Other features are defined by reference to values which we assume are held in some external feature value library (not shown here). For example, the vColl element is used a second time to indicate that the persons's siblings should be regarded as constituting a set rather than a list. Each sibling is represented by a feature structure: in this example, each feature structure is a copy of one specified in the feature value library.

If a specific feature contains only a single feature structure as its value, the component features of which are organized as a set, bag, or list, it may be more convenient to represent the value as a vColl rather than as a fs. For example, consider the following encoding of the English verb form sinks, which contains an agreement feature whose value is a feature structure which contains person and number features with symbolic values.

If the names of the features contained within the agreement feature structure are of no particular significance, the following simpler representation may be used:

The vColl element is also useful in cases where an analysis has several components. In the following example, the French word auxquels has a two-part analysis, represented as a list of two values. The first specifies that the word contains a preposition; the second that it contains a masculine plural relative pronoun:

The set, bag, or list which has no members is known as the null (or empty) set, bag, or list. A vColl element with no content and with no value for its feats attribute is interpreted as referring to the null set, bag, or list, depending on the value of its org attribute.

If, for example, the individual described by the feature structure with identifier p027 (above) had no siblings, we might specify the siblings feature as follows.

A vColl element may also collect together one or more other vColl elements, if, for example one of the members of a set is itself a set, or if two lists are concatenated together. Note that such collections pay no attention to the contents of the nested vColl elements: if it is desired to produce the union of two sets, the vMerge element discussed below should be used to make a new collection from the two sets.

It is sometimes desirable to express the value of a feature as the result of an operation over some other value (for example, as not green, or as male or female, or as the concatenation of two collections). Three special purpose elements are provided to represent disjunctive alternation, negation, and collection of values:

The value of a feature may be underspecified in a number of different ways. It may be null, unknown, or uncertain with respect to a range of known possibilities, as well as being defined as a negation or an alternation. As previously noted, the specification of the range of known possibilities for a given feature is not part of the current specification: in the TEI scheme, this information is conveyed by the feature system declaration. Using this, or some other system, we might specify (for example) that the range of values for an element includes symbols for masculine, feminine, and neuter, and that the default value is neuter. With such definitions available to us, it becomes possible to say that some feature takes the default value, or some unspecified value from the list. The following special element is provided for this purpose:

The value of an empty f element which also lacks a fVal attribute is understood to be the most general case, i.e. any of the available values. Thus, assuming the feature system defined above, the following two representations are equivalent.

If, however, the value is explicitly stated to be the default one, using the default element, then the following two representations are equivalent:

Similarly, if the value is stated to be the negation of the default, then the following two representations are equivalent:

Text elements can be linked with feature structures using any of the linking methods discussed elsewhere in the Guidelines (see for example sections and ). In the simplest case, the ana attribute may be used to point from any element to an annotation of it, as in the following example: The closest he came to exercise was to open one eye every so often , if someone entered the room

The values specified for the ana attribute reference components of a feature-structure library, which represents all of the grammatical structures used by this encoding scheme. (For illustrative purposes, we cite here only the structures needed for the first six words of the sample sentence): The components of each feature structure in the library are referenced in much the same way, using the feats attribute to identify one or more f elements in the following feature library (again, only a few of the available features are quoted here):

Alternatively, a stand-off technique may be used, as in the following example, where a linkGrp element is used to link selected characters in the text Caesar seized control with their phonological representations. Caesar seized control.

As this example shows, a stand-off solution requires that every component to be linked to must be addressable in some way, by means of an XPointer. To handle the POS tagging example above, for example, each annotated element might be given an identifier of some sort, as follows: The closest he came to exercise It would then be possible to link each word to its intended annotation in the feature library quoted above, as follows:

The Feature System Declaration (FSD) is intended for use in conjunction with a TEI-conforming text that makes use of fs (that is, feature structure) elements. The FSD serves three purposes: It provides a mechanism by which the encoder can list all of the feature names and feature values and give a prose description as to what each represents. It provides a mechanism by which the encoder can define constraints not only what it means to be a well-formed feature structure, but also valid feature structure, relative to a given theory stated in typed feature logic. These constraints may involve constraints on the range of a feature value, constraints on what features are valid within certain types of feature structures, or constraints that prevent the co-occurrence of certain feature-value pairs. It provides a mechanism by which the encoder can define the intended interpretation of underspecified feature structures. This involves defining default values (whether literal or computed) for missing features.

The scheme described in this chapter may be used to document any feature structure system, but is primarily intended for use with the feature structure representation defined by the ISO 24610-1:2006 standard, which corresponds with the recommendations presented in these Guidelines, . This chapter relies upon, but does not reproduce, formal definitions and descriptions presented more thoroughly in the ISO standard, which should be consulted in case of ambiguity or uncertainty.

The FSD serves an important function in documenting precisely what the encoder intended by the system of feature structure markup used in an XML-encoded text. The FSD is also an important resource which standardizes the rules of inference used by software to validate the feature structure markup in a text, and to infer the full interpretation of underspecified feature structures.

The reader should be aware the terminology used in this document does not always closely follow conventional practice in formal logic, and may also diverge from practice in some linguistic applications of typed feature structures. In particular, the term interpretation when applied to a feature structure is not an interpretation in the model-theoretic sense, but is instead a minimally informative (or equivalently, most general) extension of that feature structure that is consistent with a set of constraints declared by an FSD. In linguistic application, such a system of constraints is the principal means by which the grammar of some natural language is expressed. There is a great deal of disagreement as to what, if any, model-theoretic interpretation feature structures have in such applications, but the status of this formal kind of interpretation is not germane to the present document. Similarly, the term valid is used here as elsewhere in these Guidelines to identify the syntactic state of well-formedness in the sense defined by the logic of typed feature structures itself, as distinct from and in addition to the well-formedness that pertains at the level of this encoding standard. No appeal to any notion from formal semantics should be inferred.

We begin by describing how an encoded text is associated with one or more feature system declarations. The second, third, and fourth sections describe the overall structure of a feature system declaration and give details of how to encode its components. The final section offers a full example; fuller discussion of the reasoning behind FSDs and another complete example are provided in .

This elements discussed in this chapter constitute a module of the TEI scheme which is formally defined as follows: Feature Structures Feature structures Structures de traits 功能結構 (Feature Structures) Strutture di configurazione (feature structures)Estrutura das características素性構造モジュール The selection and combination of modules to form a TEI schema is described in .