Two-Level Morphology: A General Model for Word-Form Recognition and - - PowerPoint PPT Presentation

Kimmo Koskenniemi: Two-Level Morphology

Two-Level Morphology:

A General Model for Word-Form Recognition and Production Kimmo Koskenniemi, 1983 Flavia N¨ ahrlich

Eberhard Karls Universit¨ at T¨ ubingen

December 13, 2016

Kimmo Koskenniemi: Two-Level Morphology

Content

• 1. Previous Work on Computational Morphology
• 2. The Two-Level Approach
• 3. Properties and Goals of Two-Level Morphology
• 4. The Lexical Representation
• 5. Two-Level Rules
• 6. Lexical Transducers
• 7. A Critical View on Two-Level Morphology

Kimmo Koskenniemi: Two-Level Morphology

• 1. Previous Work on Computational Morphology

1.1. Generative Phonology 1.2. More General Approaches for Morphological Analysis 1.3. The Model of Kaplan and Kay

Kimmo Koskenniemi: Two-Level Morphology

1.1. Generative Phonology

• traditional phonological grammars formalized by

Chomsky and Halle, 1968

• computationally difficult
• set of ordered rewriting rules
• step by step transformation
• unidirectional
• not parallel
• problem of rule interaction

Kimmo Koskenniemi: Two-Level Morphology

1.2. More General Approaches for Morphological Analysis

• in English systems usually very few true

morphological rules, simply listed in the lexicon

• entirely language specific morphological systems
• formalismus not general enough to handle e.g.

finnish morphology (phenomena like consonant gradation)

• TEXTFIN for Finnish morphological analysis

(Karttunen 1981) relies on concept of linked minilexicons

• complexity of Finnish morphology is not beyond

the power of regular grammars or finite state automata

Kimmo Koskenniemi: Two-Level Morphology

1.3. The Model of Kaplan and Kay, 1982

• representing

rewriting rules as finite state automaton

• successive levels
• f the generative

framework

• cascades of rule

transducers could be merged into single automaton

• problem: size of

the merged automaton

Kimmo Koskenniemi: Two-Level Morphology

• 2. The Two-Level Approach
• follows the lines of

concrete and natural morphology

• role of rules is

restricted to

• ne-segment variations
• alternations wich

affect more than one segment are handled by a lexicon system

• no intermediate stages

Kimmo Koskenniemi: Two-Level Morphology

• 3. Properties and Goals of Two-Level Morphology
• language independent
• parallel and relatively independent rules
• bidirectional (conceptually and processually)
• single rules as finite state automaton
• realistic internal representations
• minimal complexity, small finite state automata
• operational computer program
• computational efficent analysis and generation of

morphologically complex languages

Kimmo Koskenniemi: Two-Level Morphology

• 4. Two-Level Rules
• only lexicon and phonemic surface representations
• lexicon: representations of word entries and

endings, e.g. Finnish

• phonemic surface level: phonemes, or letters of

phonemic alphabet

• parallel rules refer to both representations
• production or analysis by separate mechanism

guided by the rules

Kimmo Koskenniemi: Two-Level Morphology

• 4. Two-Level Rules
• partitive plural of lasi ("glass")
• specify how lexical and surface rules may correspond

to each other

• regard individual rules as equations
• stem final i is realized as e in front of typical

plural forms, e.g. the plural morpheme I

• I is realized as j if it is between vowels
• partitive ending A agrees with the stem with respect

to vowel harmony, i.e. the harmonic value of the system

Kimmo Koskenniemi: Two-Level Morphology

• 4. Two-Level Rules

two-level rules resemble generative rewriting rules

Kimmo Koskenniemi: Two-Level Morphology

• 4. Two-Level Rules

=-= denotes "any other pair"

permit the pair i − e if and only if the plural I follows

Kimmo Koskenniemi: Two-Level Morphology

• 4. Two-Level Rules

Kimmo Koskenniemi: Two-Level Morphology

• 4. Two-Level Rules
• each rule of a two-level description corresponds

to finite state automaton

• rule-automata compare two representations
• automata work in parallel instead of being

cascaded

• rules can be written as tabular automata

Kimmo Koskenniemi: Two-Level Morphology

• 5. Lexical Transducers
• set of two-level

transducers compiled from phonological rewrite rules can be merged into a single

• ne
• size of resulting

single transducer is huge compared to the size of the original rule network

• worst case: exponetial

Kimmo Koskenniemi: Two-Level Morphology

• 5. Lexical Transducers
• intersection of two-level rules is huge because

it constraints realization of all strings in universal language

• typically only strings of particular language are

interesting

• by composing lexicon with the rules spurious

strings can be filtered out

• composing a source lexicon with an intersected

two-level rule system is never sgnificantly larger than original source lexicon

• typically much smaller than intersection of the

rules by themselves

Kimmo Koskenniemi: Two-Level Morphology

• 5. Lexical Transducers
• intersecting and composing two-level rules with a

lexicon results in single lexical transducer

• contains all lexical forms of the source lexicon

and proper surface realisations as determined by the rules

Kimmo Koskenniemi: Two-Level Morphology

• 6. A Critical View on Two-Level Morphology

6.1. Tasks and Limitations 6.2. Bartons Challenge Relating to Complexity

Kimmo Koskenniemi: Two-Level Morphology

6.1. Tasks and Limitations

• in the implementation of 1983 all two level rules

are compiled into finite state automata by hand

• a compiler wich accepts rules directly in

two-level formalisms could automatically transform the rules into finite state automata

• two-level constraints are inviolable
• in Optimality Theory some problems of two-level

morphology can be solved by using vioable rules

• a finite-state formalism with weighted and

vioable two-level constraints would be desireable

Kimmo Koskenniemi: Two-Level Morphology

6.2 Bartons Challenge Relating to Complexity

• two-level morphology has been found an extremely

efficient method for processing Finnish word on very small machines

• Barton has shown the complexity of two-level method to

be NP-hard

• suggestion: words of natural languages are easy to

analyze because morphological grammars are small

• Barton has shown two-level complexity grows rapidly

with the number of harmony processes

• natural grammars do not have more than two harmony

processes

• generation time is linear with length of words and

exponential with number of harmony processes

• in practice processing time is linear with input legth