Georgian Verbal Morphology: A Learners Perspective

1
Georgian Verbal MorphologyA Learners
Perspective

• Olya Gurevich
• NTL Lunch Talk
• January 31, 2003

2
Introduction

• The Georgian verb presents problems for
  traditional morphological approaches
• Many of these problems can be resolved in an
  approach compatible with language acquisition

3
Language Learners Perspective

• Analyzing whole words
• Paying attention to frequency and to paradigmatic
  contrasts
• Parallel processing
• Can this be done in NTL?

4
Phonology

• br?gvna - To tear into pieces
• gvbr?gvni - You are tearing us into pieces
• prckvna To peel, to rip off
• gvprckvni You are ripping us off

5
The organization of the Georgian verb

• preverbAGR1theme Vowelstemthematic suffix
  tense/aspect AGR2
• Ex. xat-av you (sg.) draw X
• v- xat-av I draw X
• m-xat-av you (sg.) draw me
• m-xat-av-t you (pl.) draw me
• m-xat-av-di-t you (pl.) drew me
• da-v- xat-av I will draw X

6
Challenges Presented by Data

• General problem the meaning of a word is not
  straightforwardly composed from the meanings of
  the morphemes.
• The particular combination of parts determines
  the meaning of the form, in contrast with other
  possible combinations.
• This general property is manifested in several
  ways

7
Challenges (cont.)

• Agreement slots are not consistently associated
  with argument properties (slot competition).
• Agreement markers do not have a one-to-one
  association with morphosyntactic properties (the
  status of t).

8
A sample paradigm
Forms of the verb to draw in the Present
9
Problem 1 Slot Competition

• AGR1 Morphemes
• m 1SG Object
• gv 1PL Object
• g 2SG or 2PL Object
• v 1SG or 1PL Subject
• The AGR1 slot does not consistently indicate
  properties of the same argument
• But the morphemes correspond to particular
  morphosyntactic properties

10
Problem 2 The meaning of -t

• AGR2 morphemes
• en 3PL Subject
• s 3SG Subject except when 2PL Object
• t 1PL Subject, 2PL Subject, 1SG or 2SG Subject
  with 2PL Object.
• The morpheme t does not have a unique association
  with a morphosyntactic property of an argument

11
More on Slot Competition

• Even if we assign a meaning to each morpheme,
  that meaning does not predict which morpheme will
  show up in a form.
• E.g. v vs. g, t vs. s.

12
Problem 3 Inversion

• The agreement markers do not always correspond to
  the same syntactic arguments.
• In some tenses, m-, gv-, g- mark properties of
  the subject, and v- marks properties of the
  object.
• There is a complete reversal of agreement markers

13
Inversion (cont.)

• Ex. present tense (non-inverted)
• me en g-xatav en me m-xatav
• I you.SG draw you.SG me draw
• I am drawing you. You are drawing me.
• Perfect / evidential tense (inverted)
• me en mo-m-i-xatav-xar en me
  mo-g-i-xatav-var
• I you.SG has-drawn you.SG me
  has-drawn
• I have (apparently) drawn you. You have
  (apparently) drawn me

14
Inversion (cont.)

• Problem agreement markers cannot be uniquely
  associated with properties of syntactic
  arguments.
• More on semantics of inversion later?

15
Problem 4 Differences between tenses

• There is no unique morpheme that indicates the
  tense of the verb.
• Ex.
• v-xat-av I draw present
• da-v-xat-av I will draw future
• da-v-xat-e I drew aorist (perfective past)
• da-xat-a He drew aorist
• v-xat-av-di I was drawing past continuous
• da-v-xat-av-di I would draw past
  frequentative

16
General problem

• Many morphemes do not uniquely determine a
  morpho-syntactic property. The properties of a
  given form are determined by the combination of
  morphemes used, in contrast to other possible
  combinations.

17
Why is this a problem for traditional approaches?

• Morphemic approaches morphemes contribute
  individual meanings the whole is composed of the
  parts
• Rule-based approaches the addition of a morpheme
  consumes or realizes part of the meaning.
  Again, a unique association between an element of
  form and an element of meaning.

18
How can we avoid these problems?

• Analyze entire words and relationships between
  words (i.e. paradigm structure)
• Associate an entire word (i.e. a collection of
  form elements) with a collection of meaning
  elements.
• Produce and recognize words using example
  paradigms rather than specific rules.

19
More evidence for a word-based approach

• Inflectional classes used to predict variable
  selection of inflectional endings.
• In Georgian, there are many variations for the
  selection of endings
• We would end up with thousands of inflectional
  classes if it was the only relevant parameter

20
Predicting a paradigm (cont.)

• Ex. Formation of the aorist
• 1st and 2nd person endings - 2 possibilities e,
  i
• 3rd person endings 2 possibilities a, o
• Thematic suffix 3 possibilities same as in
  present, drop entirely, change
• Stem changes 3 possibilities stay the same,
  add the vowel e or a
• Most, if not all, combinations of the above are
  possible. If a single parameter had to account
  for everything, there would be MANY inflectional
  classes.
• Add even more classes to account for forms in
  other tenses.

21
Predicting a paradigm (cont.)

• The combination of the following parameters
  uniquely determines all forms of a verb
• Thematic suffix - always there in the
  Present-tense form
• Shape of the root (whether or not it has a vowel,
  which consonant it ends in) - can be seen in the
  Present form
• The pre-verb always there in Future or Aorist /
  Imperative forms
• Stem suppletion if occurs at all, it occurs in
  the Future
• Other irregularities are predictable from the
  Infinitive or the Aorist (Imperative) forms
• The meaning of the verb and its syntactic valence

22
Predicting a paradigm (cont.)

• For predictive value, it does not matter which
  Present or Future forms are used (i.e. 1st, 2nd,
  or 3rd person).
• In language learning, the most frequent forms are
  used (Bybee 2001).
• The forms needed to predict a paradigm
• Present
• Future
• Aorist (Imperative)
• These tenses are among the first learned by
  children and the most frequent in discourse

23
Predicting a paradigm (cont.)

• Ex to draw
• Present tense (2SG) xatav
• Stem xat has a vowel
• Thematic Suffix av
• Future tense (2SG) daxatav
• Preverb da
• No stem suppletion
• Imperative (2SG Aorist) daxate
• Aorist ending e
• Predicted choices
• Formation of 2nd subjunctive (TS drops, ending
  o)
• Formation of 1st and 2nd Resultant (TS stays,
  ending s)
• 3rd person aorist ending a
• No stem changes in any forms

24
How does this work?

• Example paradigms (i.e. paradigms of the most
  common verbs) are stored
• For less frequent verbs, only the most frequent
  (predictive) forms are stored
• The relevant parameters in a form to be
  recognized activate similar example paradigms,
  possibly more than one
• The example paradigm that fits the pattern best,
  wins. The form is interpreted / constructed
  based on that example.
• Best fit is defined by a combination of
  similarity and frequency of the example (Bybees
  lexical strength).

25
Can this be done in NTL?

• Neural nets are very good at associating
  combinations of forms with combinations of
  properties (distributed representations)
• Neural nets are very good at finding best fit
  solutions
• Can a simple PDP net learn paradigm structure?
  Seems like it should be able to.

26
What about ECG?

• ECG cannot express paradigm structure, and refer
  to other forms possible within a paradigm.
• However, it should be able to send off a form
  to the PDP net and get back a combination of
  morphosyntactic properties. This process could
  be reversed for production, as well a
  combination of properties is run through a PDP
  net which gives a likely form
• Any reason why this wouldnt work? Any other
  ideas?
• Concern how do you train a PDP net to produce
  example paradigms that are compatible with actual
  human input?

27
So Georgian is weird. Who cares?

• The difficulty of assigning meaning to morphemes
  (or even words) is not unique to Georgian.
• Periphrasis (e.g. the English auxiliaries) works
  this way too, as well inflectional morphology of
  many other languages.
• Perhaps this could be a general solution to those
  kinds of problems
• Mappings between form and meaning can be
  many-to-many, and it is necessary to know what
  the possibilities are to know what a particular
  combination of form elements means.

28
Semantics of Inversion

• Inversion in Transitive and other Active Verbs
• In Present tense, Subject is in Nominative, Obj
  is in DAT/ACC
• In Evidential tenses, Subj is in DAT, Obj is in
  NOM
• The reversal of agreement markers corresponds to
  a reversal of case markings on the arguments.
• The subject in Inverted tenses is less active
  (less responsible) and is therefore indicated by
  a less prominent case

29
Inversion (cont.)

• Present
• k'ac-i dzaql-s xatav-s
• man.NOM dog-sg.DAT draw.PRES
• 'The man draws / is drawing the dog
• Evidential
• k'ac-s dzaqli da-u-xatav-s
• man.DAT he.NOM draw.1RESULT
• 'The man has (apparently) drawn the dog'

30
Inversion (cont.)

• Present
• k'ac-i me m-xatav-s
• man.NOM me draw.PRES
• 'The man draws / is drawing me
• Evidential
• k'ac-s me da-v-u-xatav-var
• man.DAT me draw.1RESULT
• 'The man has (apparently) drawn me'

31
Inversion (cont.)

• There may be a separation between semantic roles
  (most prominent argument, least prominent
  argument etc.) and syntactic arguments (subject,
  object).
• In non-inverted tenses, subject is most active /
  most prominent
• In inverted tenses, the mappings are reversed
• Morphology is sensitive to semantic roles rather
  than syntactic arguments
• m-, gv- markers mark the less prominent argument,
  v- marks the more prominent argument)

32
Inversion (cont.)

• Indirect Verbs
• Verbs like have, want always have inverted
  markers, their subjects are always in Dative
• Subjects of these verbs are inherently
  non-active, receiving the action
• These verbs are among the first learned by
  children and present no problems
• Regular inversion is learned late and is
  difficult for children to master
• Indirect verbs must have a different
  representation they are high-frequency irregular
  verbs