Overview

• Dependency grammar
• Extensible Dependency Grammar
• Relevant features of Amharic morphosyntax
• XDG for Amharic
  • Incorporating morphology
  • Null subjects and objects
  • Agreement in relative clauses

Dependency grammar

• Analysis consists of a set of directed, labeled arcs connecting word nodes; no phrasal nodes.
• Constraints on analysis graph, e.g., each node other than the root node has a head node.
  
• Constraints are defined for words, lexemes, or lexical classes.

Extensible Dependency Grammar (Debusmann, 2006)

• Modular: separate dimensions represent different levels of grammatical abstraction.
• Each dimension has its own set of arcs joining the nodes. An analysis is a multigraph over the nodes.
• Parallel, rather than syntacto-centric
• Declarative: supports both parsing and generation
• Processing through constraint satisfaction
• Lexical entries arranged in an inheritance hierarchy: word, lexeme, grammatical class

XDG: principles and constraints (1)

• Grammatical analysis: one that conforms to a set of constraints, each generated by one or another principle.
• Examples
  • Principles concerned with the structure of the graph; for example, on the Syntax dimension, each node other than the root must have exactly one head.
  • Valency Principle, governing labels on arcs into and out of a node

    -  lexeme: ~V~take
       syn:
          out: {sbj: !, obj: !}
    

  • Agreement Principle, constraining how certain features within some words must match features in other words

    -  gram: V_FIN_PRES
       syn:
          agree: [sbj]
    
    -  gram: V_3SP
       syn:
          agrs: [3ps]
    

XDG: principles and constraints (2)

• Examples
  • Order Principle, concerned with order of words in the sentence

    -  gram: V_T
       syn:
          order: [sbj ^ obj]
    

  • Linking principles, concerned with how arc labels correspond across dimensions; apply on interface dimensions, which join dimensions such as Syntax and Semantics ("lb12s" is "LinkingBelow1or2Start").

    -  gram: V_T
       synsem:
          lb12s: {agt: [sbj], pat: [obj]}   
    

Parsing with XDG

Lexicalization: assigning entries to nodes Creation of variables and instantiation of constraints Constraint satisfaction

Amharic morphosyntax: verb morphology

Verb: stem and up to 4 prefixes and 4 suffixes Stem: root and template Template represents tense/aspect/mood (TAM) and one of a set of 8 derivational categories Affixes represent subject and object agreement, negation, relativization, definiteness, (accusative) case, prepositions, conjunctions Lexeme: root + derivational category Null subjects and null objects

Amharic morphosyntax: noun phrases

• Nouns without modifiers take suffixes indicating definiteness and accusative case and prepositional prefixes

  ሐኪም hakim doctor 'a doctor'	ሐኪሙ hakimu doctor-DEF 'the doctor'	ሐኪሙን hakimun doctor-DEF-ACC 'the doctor (obj.)'	ለሐኪሙ lehakimu for-doctor-DEF 'for the doctor'

• When a noun is modified by one or more determiners, adjectives, or relative clauses, it is the first modifier that takes these prefixes or suffixes.

  ሰነፉ ሐኪም senefu hakim lazy-DEF doctor 'the lazy doctor'	ለሰነፉ ሐኪም lesenefu hakim for-lazy-DEF doctor 'for the lazy doctor'	ያንን ሰነፍ ሐኪም yann senef hakim that-ACC lazy doctor 'that lazy doctor (obj.)'

Amharic morphosyntax: relative clauses (1)

• Relative clauses consist of a relative verb (imperfective or perfective with a relativization prefix) and zero or more arguments of the verb.
• As with other verbs, a relative verb agrees with its subject and may agree with its object if it has one. It may also agree with an argument of the matrix verb.
• Both subject and object relativization are possible.

  የሚወዳት ሴት yemiwedat sEt REL-he-likes-her woman 'the woman that he likes'

  የሚወዳት ወንድ yemiwedat wend REL-he-likes-her man 'the man who likes her'

• As with other modifiers, when the relative clause is the first modifier of a noun, it takes the definiteness and accusative suffixes and prepositional prefixes.

  የተመረቀው ልጅ ወንድሜ ነው yetemereqew lj wendmE new REL-he-graduated-DEF boy my-brother is 'the boy who graduated is my brother'

  የተመረቀውን ልጅ አላውቅም yetemereqewn lj alawqm REL-he-graduated-DEF-ACC boy NEG-I-know 'I don't know the boy who graduated'

Amharic morphosyntax: relative clauses (2)

• The head noun of a noun phrase with a relative clause (or adjective) modifier is optional.

  የሚወዳት አልደረሰም yemiwedat alderesem REL-he-likes-her NEG-he-arrived '(he) who likes her didn't arrive'

• Because relative clauses (and adjectives) (1) agree with the matrix verb, (2) determine the syntactic role and definiteness of the noun phrase, and (3) do not require a head noun, we consider them to be the syntactic heads of their noun phrases.
• The "head" noun plays the role of semantic (but not syntactic) head.
• The relationship between relative clause modifier and modified noun is analogous to that between auxiliary verb (syntactic head) and main verb (semantic head) in XDG.

XDG for Amharic: incorporating morphology

• It's impractical to store all Amharic wordforms (a single verb lexeme can appear in more than 100,000 wordforms).
• We take the lexeme as the basic lexical unit.
• During parsing, we first run the HornMorpho morphological analyzer (Gasser, 2009) over the words in the sentence.
• Given a word, HornMorpho returns the lemma and the grammatical analysis in the form of a feature structure description.

  ይወዳታል ywedatal 'he likes her'
  →
  (wedede, {tam: impf, +rel, -neg,
            sbj: [-p1,-p2,-plr,-fem],
            obj: [-p1,-p2,-plr,+fem]})
  

• Sequence of lemma-structure tuples, rather than wordforms, is the input to XDG lexicalization.

Parsing Amharic with XDG

Morphological analysis Lexicalization: assigning lexemes and grammatical entries to nodes Creation of variables and instantiation of constraints Constraint satisfaction

XDG for Amharic: null subjects and objects (1)

• XDG is grounded in the words in the sentence, but it has to come to grips with the mismatch between syntax and semantics.
• Surface words that don't correspond to semantic units are "deleted" with del arcs in XDG.
• What about semantic units that don't correspond to surface words: null subjects and null objects?
• Empty nodes: each verb creates one for its subject; each transitive verb creates one for its direct object.
• Empty Node Principle: sanctions arc from the verb with the relevant label (sbj or obj) to either an explicit word or the associated empty node, but not both.

XDG for Amharic: null subjects and objects (2)

• Analysis of

  ዮሐንስ     ይወዳታል
  yoHans    ywedatal
  Yohannis  he-likes-her
  'Yohannis likes her'
  

  

XDG for Amharic: subject and object agreement (1)

• In original XDG, agreement applies to two separate verb attributes
  • agrs: list of possible features for the verb form
  • agree: list of arc labels for daughters that must agree with the verb

  -  word: eats
     syn:
        agrs: [3ps]
        agree: [sbj]
  

• This approach fails to address the complexity of a language such as Amharic
  • agrs must distinguish subject, direct object, and indirect object features.
  • agree must specify which agreement feature of the mother verb agrees with the daughter on the specified arc.
  • agree must allow for agreement with different features of the daughter when the daughter is a verb itself (that is, a relative clause).

XDG for Amharic: subject and object agreement (2)

• Entry for transitive verbs

  -  gram: V_T
     syn:
        agree: {sbj: [sbj, [ˆ, sbj, obj, iobj]],
                obj: [obj, [ˆ, sbj, obj, iobj]]}
  

• Analysis of

  አስቴር  ዮሐንስን      ትወደዋለች
  astEr yoHansn     twedewalec
  Aster Yohanis-ACC she-likes-him
  'Aster likes Yohannis'
  

  

XDG for Amharic: subject and object agreement (3)

• Verb agreement feature and arc label need not be the same.
• For one subclass of verbs, the verb object suffix agrees with the topic.

አስቴር  ደክሟታል
astEr dekmWatal
Aster it-has-tired-her
'Aster is tired.'

-  lexeme: dekeme
   syn:
      agree: {top: [obj, [ˆ,sbj,obj,iobj]]}

XDG for Amharic: relative clauses (1)

• Relative verb is the syntactic head of its noun phrase.
• Relative verb is joined to its "head" noun by an arc labeled with the syntactic role that the noun plays in the verb.
• In the Semantics dimension, the head noun has two arcs into it, one from the relative verb, one from the verb of the matrix clause.
• Analysis of sentence in which the relative verb is the subject of the matrix clause, and the "head" noun of the relative clause is the object of the relative verb.

  አስቴር የምትጠላው           ወንድልጅ ታመመ
  astEr yemetTelaw        wendlj  tameme
  Aster REL-she-hates-him boy     he-got-sick
  'The boy that Aster hates got sick.'
  

  

XDG for Amharic: relative clauses (2)

• Relative clauses with no "head" noun have no overt form for the shared semantic argument, so this is introduced as an empty node.
  

XDG for Amharic: relative clauses (3)

• Without further constraints, the grammar assigns multiple analyses to some unambiguous sentences and analyses to some ungrammatical sentences.

  *አስቴር  የምትጠላው          ወንድልጅ ታማመች 
  *astEr yemtTelaw         wendlj  tamemec
  Aster  REL-she-hates-him boy     she-got-sick
  ‘The boy that Aster hates (she) got sick.’
  

• This satisfies the two relevant agreement constraints:
  1. that some feature of the main verb (its subject) agree with some feature of the relative verb (its subject)
  2. that some feature of the relative verb (its object) agree with the modified noun wendlj.

XDG for Amharic: relative clauses (4)

• Further XDG principle needed to prevent such analyses.
• Cross-Agreement Principle: forces the same feature of the relative verb to agree with the main clause verb and the modified noun.
• Represents a fundamental fact about relative clauses in all languages: that the same noun functions as an argument of two different verbs.
• No analysis for the sentence because the feature of the relative verb yemtTelaw that agrees with the modified noun (its object) differs from the feature that agrees with the main verb (its subject).
  

Conclusions

• Modularity of XDG allows us to treat relative verbs as NP heads in the Syntax dimension, but not the Semantics dimension.
• However, XDG requires augmentation to deal with
  • empty subjects and objects: empty nodes and the Empty Node Principle
  • agreement: elaborated agreement features, Cross-Agreement Principle
  • complex morphology: incorporation of morphological analysis into parsing.
• Many of relevant properties of Amharic are shared by other Ethio-Semitic languages.
• Ongoing work
  • Extending Amharic grammar: clefting, infinitive complements
  • Synchronous XDG for Amharic and other languages to support translation