Glosa Lexicon Builder

The Glosa Lexicon Builder is a morphological database tool that allows the rapid development and testing of descriptions of natural language words and word formation rules. It consists of two components: a lexicon component and a rules component. Either one can be developed first; the lexicon component automatically determines derivations and inflections from the rules component, allowing overriding for special cases, while the rules component can be developed automatically from individual lexicon examples, using a special machine-learning technology.

Features of the Lexicon Builder:

Hierarchical approach to rule building; Use two-level phonological rules to specify language-wide relationships, context-dependent affixation rules for each derivational relationship, and exceptions for individual word special cases.
User-defined character classes with regular expressions allow much more compact rule definition.
Option to automatically generate a new rule each time you add a non-comforming case.
Graphical User Interface lets you see generated forms as you build.
Multiple levels of inflection for agglutinative languages, using finite-state morphotactics.
Can store abstract forms that are not in lexicon.
Store user-defined sets of properties with each form. User can define default properties for each derivation, and override them when necessary. Properties can also control the application of affixing rules.
Export lexicons for Glosa Inflected Forms Generator, or a morphological front-end for custom applications written in C or C++.

Defining the Rules Component

The Lexicon Builder offers three different approaches to defining a language's morphology. These approaches can be used separately or in combination.

Affix-Specifc Surface Rules

This paradigm combines phonological rules and morphotactics into a single rule formalism, using only surface symbols. Each set of rules shows the relationship between surface configurations with and without a particular affix; a single rule shows the relationship for a subset of the word forms that can take the affix. For example, the English Noun Plural suffix -s could be represented by this series of rules (B represents sibilants s, x, and z): Affix-Specific Surface Rules

The most specific rule that applies in a given case is used. Specificity is defined as

Greater length of matching terminal string.
Smaller size of character class (for instance B, which contains only three members, would be more specific than C, the set of all consonants). A single character is more specific than any character class.

General Surface Rules

To capture language-wide phonological phenomona, special rules can be applied at morpheme boundaries These occur conceptually after an affix attachment or before an affix removal, but in reality they are composed with the affix-specific rules to form a new set of affix-specific rules. Thus we could define phonological rules similar to the plural suffix rules above, that would apply in any situation where the suffix consists of a single s (i.e. Noun Plural and Verbal Third Person Singular). The plural suffix specification would then be reduced to a single rule.

General Lexical Symbol Rules

The Lexicon Builder supports special Lexical Symbols in affixes. As in two-level rules[1], these lexical symbols can be realized as different surface formations in different contexts. For example, one could define a lexical symbol S that is realized as es after B, o, ch, or sh, and as s the rest of the time. Some Lexical Symbol Rules The plural suffix specification is then reduced to

...<->...S.

The corresponding surface only rules are automatically computed, and can be viewed in the transformation window. Computed Surface Rules

Comparison of the Lexicon Builder with Two-Level Finite State Transducer based Implementations

Similarities:

In a particular context, one and only one rule is applied. No need to worry about ordering of multiple rules as in classical generative morphology.
Rules are bi-directional. The same specification is used to direct both generation and recognition.
Both allow use of rules to isolate and remove predictable phonological phenomena.

Differences:

The Lexicon Builder compiles all rules to a TRIE-based surface-only transformation algorithm. This algorithm operates in constant time (independent of the size of the lexicon, number of rules, or the length of the input word) for each derivational step. Backtracking is required only when using regular expressions or character classes, and then only in certain situations. While very efficient at verifying an acceptable form, finite-state transducers can run in exponential time when trying to produce a form from scratch[2], largely because of the need for extensive backtracking. Execution time also increases linearly with the number of automata used.
Generation proceeds from the root to the affixes (the left and right ends of the input string), rather than from left to right or right to left. The exact position of newly generated material depends on whether the affix is a prefix or suffix, in some cases being both (German past participles) or neither (infixes for Arabic finite verbs). Recognition proceeds in the opposite manner, from the affixes inward to the root, using a breadth-first search to enumerate all possible morphological parses. This simplifies the construction of sophisticated morphotactics.
Exceptions can be specified without affecting any other part of the lexicon.
Word properties can be specified separately from the word's lexical form. This avoids a complicated interaction between symbols representing properties (such as stress) and symbols that are actually part of the word. Words themselves are stored as surface forms.
User-friendly interface allows the user to verify derived forms immediately. No need for traces or complex debugging.
WYSIWYG rule editor simplifies creation of rules for unusual morphological phenomena like negative affixing and reduplication.

References

[1] Koskenniemi, Kimmo. 1983. Two-level morphology: a general computational model for word-form recognition and production. Publication No. 11. Helsinki: University of Helsinki Department of General Linguistics.

[2] Barton, G. Edward, Robert C. Berwick, and Eric Sven Ristad. 1987. Computational complexity and natural language. Cambridge, MA: The MIT Press. (see chapter 5, "The complexity of two-level morphology").

Demos of Lexicon Builder applications:

Spelling Alternations in English (available soon)
Morphotactics and Vowel Harmony in Turkish (available soon)
A Hyphenator for Greek (available soon)

A portion of an English lexicon.

For further information and pricing, contact:

Glosa International
4538 Winona Ct.
Denver, CO 80212-2513
USA
+1-303-458-1496 (voice and fax)

[Next] [Home]