Hermit Crab Parsing Engine Specification



Download 403.76 Kb.
Page3/20
Date31.07.2017
Size403.76 Kb.
#25627
1   2   3   4   5   6   7   8   9   ...   20

2.3Phonological Rules


Phonological rules are of the general form

X ® Y / W __ Z

where X, Y, W and Z represent (possibly optional) sequences of phonetic segments (specified as phonetic feature matrices) and/or boundary markers. These sequences cannot include any segments of morphemes which have not yet been attached by a morphological rule, i.e. which are outside of the “current” stem. The application of phonological rules may also be restricted by requiring the presence or absence of specified MPR features and/ or part of speech. Syntactic features (i.e. head and foot features) are invisible to phonological rules.

“Phonological rules”, as here defined, encompass the allophonic and morpho-phonemic rules of structuralist linguistics, as well as the phonological rules of generative phonology.

Phonetic features are specified in phonological rules by a unique value (such as ‘+’ or ‘ ’), or by alpha variables (as used e.g. in Chomsky and Halle 1968).

As mentioned above, phonological rules are allowed to apply at one or more strata. Within each stratum, phonological rules may be specified as being ordered linearly, or as applying simultaneously; disjunctive subsets of rules may also be defined.


2.3.1Reverse Application of Phonological Rules


One peculiarity of Hermit Crab's application of phonological rules is due to its operation in analysis mode. Rather than starting with an underlying form and applying phonological and morphological rules to unambiguously synthesize a surface form, Hermit Crab begins with a known surface form, and attempts to analyze it into one or more underlying forms. As with morphological rules, phonological rules are neutralizing in the synthesis mode: they assign a value to a feature regardless of what the previous value (if any) of that feature may have been. In undoing the application of such a rule, i.e. applying it in an analysis manner, one doesn't know what feature values the vowel had prior to the rule's application, at least until lexical lookup. The rule may have applied by changing the feature values in question, but it is equally possible the rule applied vacuously. When un-applying a phonological or morphological rule, Hermit Crab un-instantiates such features. When another rule needs to know the value of an un-instantiated feature, the morpher simply assumes that uninstantiated features have the required value. This may result in over-application of phonological rules, but any incorrect application will be discovered when the rules are applied in synthesis mode to looked-up lexical entries.

At any rate, a deeper rule will require the value of a feature which was uninstantiated by a shallower rule only in the case of opaque rule orderings, i.e. when one rule counterbleeds or counterfeeds another rule. For instance, consider two rules ® B /C__D and C ® E /false__G, applying in that (synthesis) order, and suppose that there are at least some forms in which both rules will apply. When Hermit Crab undoes the application of the second (shallower) rule, it will uninstantiate the features of C. But when Hermit Crab comes to the first rule, the value of those features required by that rule are unknown, so Hermit Crab simply assumes that the rule applies. (For further background, see Maxwell 1991.)

The situation is more complex in the case of length-changing rules (e.g. deletion rules, epenthesis rules, and diphthongization rules). Due to the internal representation used by Hermit Crab, the morpher will need to explore two search paths: one for which the length-changing rule is unapplied, and one for which it is not. This is true regardless of the interaction of the length-changing rule with other rules.

When doing lexical lookup of a partially instantiated phonetic feature matrix, Hermit Crab looks for all lexical forms which would match the feature matrix, ignoring unspecified feature values. For instance, suppose a language has both front rounded and front unrounded vowels, and Hermit Crab has undone a rule which rounds vowels in some environment. Then given that the rule has been “un-applied” to a form with surface [ü], the morpher will create a high front vowel with unknown rounding, and attempt to find a lexical entry with either an [i] or an [ü] in that position. (If the phonological analysis which Hermit Crab were modeling in this example used archiphonemes, the morpher would also attempt to find a lexical entry with the appropriate archiphoneme in that position.)


2.3.2Cyclic Phonological Rules


Cyclic phonological rules apply once at the beginning of a cyclic stratum, and once after each application of a cyclic morphological rule; they are ordered among themselves as specified by the user. The morpher further constrains the application of cyclic phonological rules on all but the first cycle by Kiparsky's (1982) Strict Cycle Condition, given below:

Cyclic phonological rules apply only to derived representations.

A representation X is derived with respect to phonological rule R in cycle j iff X meets the structural analysis of R by virtue of a combination of morphemes introduced in cycle j or by virtue of the application of a previous cyclic phonological rule in cycle j (even if that application was vacuous).

2.3.3Non-cyclic rules


Non-cyclic phonological rules are applied as a block after any applicable morphological rules of the same stratum have applied. Their order among themselves may be specified by the user.

2.3.4Boundary Markers


Theories differ as to the number and kind of boundary markers they countenance. Hermit Crab makes no commitment to any of these theories, save that there is no provision for treating boundary markers as segments with features (as in Chomsky and Halle 1968).

Boundary markers are inserted as strings (not phonetic feature matrices).

Boundary markers in the phonetic shape of a lexical entry are ignored when matching that lexical entry against a phonological rule, unless the rule explicitly requires the boundary.

Boundary markers are erased at the end of each cycle and stratum.

Both morphological rules and phonological rules may insert boundary markers. However, the use of phonological rules to insert or alter boundary markers (i.e. readjustment rules) is discouraged, as it may lead to computational intractability.

2.3.5Deletion Rules


In the absence of other restrictions, the fact that phonological rules can delete segments puts phonology into the domain of an unrestricted rewrite grammar. Since such a grammar would be impossible to parse, Hermit Crab places arbitrary (i.e. nonlinguistic) restrictions on deletion rules. (This is not to say that we have placed sufficient restrictions on such rules—a sufficiently ingenious linguist may still find some way of putting the morpher into an infinite loop, perhaps by including a deletion rule in a cyclic stratum!)

For the purposes of this discussion, a deletion rule is any rule which deletes part of its input, i.e. where the number of segments in the output of the rule is less than the number of segments in its input.

Understanding the arbitrary restriction that Hermit Crab uses requires an understanding of the way in which unapplication of deletion rules proceeds. Unlike all other rules, deletion rules are always unapplied as if they had been applied simultaneously. That is, during unapplication to a form X, X is scanned for all places where the deletion rule could be unapplied, and the rule is unapplied to those places, resulting in the new form X'. By default, that is the end of it; deletion rules cannot be unapplied again. However, if the user is brave, he can set the variable *del_re_apps* to some number greater than zero (its default); then the deletion rule is unapplied to X', and to X'', etc. *del_re_apps* times.

The above definition is couched in terms of simultaneous (un )application of the deletion rule. However, if *del_re_apps* is set to a sufficiently large number, unapplication of a deletion rule will generate from a surface form all the underlying forms (and more) from which iterative application of the deletion rule might have generated the surface form (I think!).




Download 403.76 Kb.

Share with your friends:
1   2   3   4   5   6   7   8   9   ...   20




The database is protected by copyright ©ininet.org 2024
send message

    Main page