Overview

• Project goals
• Dependency grammar and XDG
• Multilingual XDG
• Translation in XDG
• Project status, ongoing work

Why RBMT?

• Some reasons for preferring RBMT
  • Lack of bilingual (or even monolingual) data for training SMT
  • Goal of publication quality output over a restricted domain (Ranta et al., 2010)
  • Morphologically complex languages
• Our goals
  • Translation among
    • English, Amharic, Oromo, Tigrinya
    • Spanish, Quechua
  • Restricted domains related to science and health (currently water); relatively small grammars/lexicons

Dependency grammar

• Features
  • Units: words and word classes; there are no phrasal units
  • Analysis: dependency graph consisting of labeled arcs joining word nodes
  • Grammar: constraints on possible arcs joining nodes
• Advantages
  • Simplicity
  • Ease of integration of syntax and semantics
  • Handling of word order variation and long-distance dependencies
• In MT: Bick (2007), Čmejrek et al. (2003), Dianescu (2004), Mel'čuk & Wanner (2006)
• Extensible Dependency Grammar (Debusmann et al., 2004; Debusmann, 2007)
  • Flexibility
  • Proven capacity to handle complex syntactic constraints
  • But untested with wide coverage grammars and unconstrained input

XDG: nodes and arcs

• Simplest sentence analysis
  • Weakly connected, directed graph over
    • a set of word nodes and
    • a distinguished root node representing end-of-sentence punctuation
    
  • A set of features for each node
    • Node 2: person=3, number=1

XDG: dimensions (1)

• Multiple dimensions, each corresponding to a level of grammatical abstraction
• Analysis is a multigraph
• Some dimensions
  • Syntax: Linear Precedence, Immediate Dominance
  • Semantics: Predicate-Argument, Scope

XDG: dimensions (2)

• Interface dimension: relates two dimensions, no arcs of its own
  

XDG: principles

• A grammatical analysis is one that conforms to a set of constraints
• Each constraint is generated by a principle
• Each dimension has its own set of principles
• All principles are defined in a language-independent fashion
• Some principles
  • Tree Principle: constrains the topology of the graph
  • Valency Principle: constrains the labels on arcs into and out of nodes
  • Agreement Principle: constrains agreement between feature values for nodes related by arcs with particular labels
  • Agr Principle: constrains the possible values for a given feature for a node
  • Order Principle: constrains the order of nodes
  • Linking End Principle: constrains how arcs in two dimensions associate with one another

XDG: grammar and lexicon

• An XDG grammar
  • A set of dimensions, each with its characteristic principles and possible arc labels
  • A lexicon
• All specific grammatical constraints are stored in word-level units
• Lexicon
  • A set of entries organized in an inheritance hierarchy
  • Wordforms (contaminated), lexemes (contaminate), lexical classes (v)
• Entry
  • Lexemes or lexical classes that the unit inherits from
  • One or more grammatical constraint specifications, each invoking one principle on a given dimension
  • An example: main clause transitive verb: Syntax dimension, Valency principle
    

Lexicon/grammar as an inheritance hierarchy

XDG: processing

• (Morphological analysis)
• Lexicalization
  • Search for matching entries
  • Invokation of principles referenced in entries
    • Creation of variables; examples:
      • Each node n has a daughters variable whose value is the set of indices of daughter nodes of n
      • Each node that must agree with another has an agr variable for the possible values of the relevant agreement feature (for example, person-number-gender)
    • Instantiation of constraints relating variables
  • Creation of a disambiguation variable for each node; its value is the index of an entry all of whose constraints are satisfied
• Constraint satisfaction
  • If it succeeds, it returns all possible complete variable assignments, each corresponding to a single grammatical analysis of the sentence (a multigraph across the nodes)
• XDG grammars are declarative, so they can be used for generation as well as analysis
  • Each node needs an explicit position variable representing its output position in the generated sentence
  • (Morphological generation)

Analysis example

Multilingual XDG (1)

• For each language
  • Constraints (by way of principles) on one or more syntactic dimensions
  • Interface dimension relating one syntactic dimension to one semantic dimension, constraining how arc labels in syntax relate to arc labels in semantics
• Semantics shared by all languages
• An Amharic translation of the water is contaminated
  

Multilingual XDG (2)

• The English predicate adjective contaminated and the Amharic present perfect passive verb ተበክሏል play the same role in the semantics despite their very different syntax
• These differences are reflected in the English entries for predicate adjectives and main clause copulas and the Amharic entry for present perfect intransitive verbs
  

Synchronous XDG

• One big multilingual grammar with syntactic dimensions for each language, as well as semantics and the syntax-semantics interface dimension for each language
• A sentence and its translation into one or more languages: just a multigraph connecting the nodes of the sentence
• A multilingual analysis
  

Connecting the languages

• Entries in one language associated with entries in the other through cross-lingual links that behave like lexical inheritance links
• Partial Amharic and English lexicons with three cross-lingual links
  

Translation

• Analysis + generation
• Ordinary constraint satisfaction with additional target language constraints
• Additional generation steps
  • Ordering of output words
  • Morphological generation

Translation example

Not discussed

• Mismatch in number of words across dimensions
• Disambiguation
• Multi-word units

Project status

• Our own (mostly completely) Python implementation of XDG
• Morphological analyzers/generators for Amharic, Oromo, Tigrinya, Quechua, and Spanish (see Gasser (2009) for the approach)
• Tiny XDG grammars for English and Amharic plus cross-lingual links
• Downloadable at http://www.cs.indiana.edu/~gasser/Research/software.html

Ongoing research and concerns

• Grammatical details of six languages
• water semantics
• Approaches to lexical disambiguation
• Machine learning
  • Cross-lingual links
  • Weighted constraints
  • Active learning, reinforcement learning

Conclusions

• Properties of XDG, especially multi-dimensionality and modularity, lend themselves to grammar-driven MT
• Synchronous XDG grammar
  • Modular lexicons of two or more languages
  • Simple set of cross-lingual inheritance links joining entries
• Source and target sentence viewed as a single abstract "sentence" consisting of bilingual word nodes
• Lexicons are declarative so translation is possible in both directions
• More in common with interlingua than transfer (though a transfer version is also possible)
• But unlike in usual interlingua translation, no separate analysis and generation phases
  • Constraint satisfaction seeks a representation of the sentence at all levels simultaneously
  • Syntax constrains semantics, semantics constrains syntax, and the evolving syntactic representations of source and target constrain one another