Pluralism (informal): elicit, for each system, a full description of the system (as it applies to text encoding), including

list of features

examples of each feature

structural properties of each feature (especially combinations with other features and the like)

test criteria for the recognition of the feature

formalizations available for the system

Pluralism (formal): generate an SGML formalization for the feature set of each system. If SGML syntax does not suffice, the metalanguage committee must be asked to consider or develop extensions to handle the recalcitrant features.

Note that at this low level of formalization, different theories will have separate and incongruent tag sets. The meaning of any tag will be defined only in natural language, and users of different theoretical orientations will be responsible for any translation into their own terms. The same generic identifier might be used for textual features postulated by different theories, and thus be ambiguous when viewed in isolation. The two following approaches handle this ambiguity differently.

Eclecticism: define a single tag set created by the union of all system-specific tag sets, eliminating ambiguity by giving each theoretically distinct textual feature a unique generic identifer (tag). Users of the scheme will be expected to tag some subset of the features in the set, mixing and matching as they wish.

Controlled semantics: eliminate ambiguity of generic identifiers by the explicit formal definition of the linguistic and computational meaning of the tags. Different usages of the same term must at this level be reduced to a finite list of questions with enumerable sets of possible answers. Ultimately, of course, the meaning of these questions and their answers will be expressed in natural languages, so that this cannot amount to a full specification of meaning. The formal system, however, will be constrained more fully by these formal definitions than at the lower levels of conflict resolution. .*

As an example of formalization at this level, let us consider the preparation of lemmatized frequency lists. Lemmatizing practice varies on issues such as the forms of the definite article (are French le, la, l', les, one lemma, two, three, or four? do the tokens au, aux contain one lemma each or two?), but range over a finite number of possible solutions to a finite number of standard problems. The lemmatization practice of existing lists may thus be fully defined by the answers to a finite number of questions (e.g. Are articles of different gender reduced to separate lemmata, or the same lemma? or Are words which combine a preposition and an article reduced to one lemma or two?), such that each answer is drawn from a well-defined range of possible answers. In practice, these questions and answers provide a sufficient definition of the meaning of the tag &lit tag='lemma'>.

Now, for any given finite list of questions, with well-defined ranges for each answer, a set of SGML tags and attributes can be created which specifies answers to each question. An SGML processor could then ascertain that answers are specified for each question (although it could not necessarily check the consistency of the answers with the actual practice in the encoding).

Polytheoretical consensus: provide the smallest possible single set of features which includes as a subset each set of features used by an existing system, or else define explicit mappings from the tag sets provided for one system into the tag sets provided for another. Unlike an eclectic tag/feature set, a polytheoretical set avoids all redundancy and is expected to be used as a unit, with all features tagged, so that a text tagged with such a set is useful to researchers of widely varying theoretical persuasions.