Overview

• Rule-based machine translation
• Extensible Dependency Grammar
• XDG for translation: L³
  • Shallow and deep translation
  • Node mismatch
  • Structural divergences
• Project status
• Ongoing work
• Conclusions

Context and long-term goals

• Linguistic Digital Divide: dearth of documents in many African languages
• Computer-assisted translation and machine translation: production of publication-quality documents in languages of Horn of Africa

Rule-based vs. statistical machine translation

• SMT requires bilingual corpora: few available for under-resourced languages like Amharic
• Morphologically complex languages like Amharic: further data sparsity
• RBMT offers straightforward treatment of significant structural differences (as with English-Amharic)
• RBMT appropriate when the goal is publication-quality documents in narrow content domains

Dependency grammar (1)

• Simplicity: only lexical nodes
  

Dependency grammar (2)

• Separate dimensions; sentence analysis a multigraph over a single set of word nodes
  • Immediate Dominance and Linear Precedence dimensions: handling word order variation and long-distance dependencies
    

Dependency grammar (3)

• Separate dimensions
  • Syntax and semantics dimensions
    

Extensible Dependency Grammar (XDG)

• Debusmann (2007), Debusmann et al. (2004)
• Flexible, modular
• Proven capacity to handle complex syntax
• Interface dimensions that relate dimensions such as ID and LP or ID and Semantics
• Processing by constraint satisfaction

XDG: interface dimensions (1)

• The linking end constraint: constrains node n to have as the daughter on its l1 arc in dimension d1 a node which is the daughter of some node on dimension d2, not necessarily n, on an arc with label l2
  

XDG: interface dimensions (2)

XDG: grammatical features

XDG: lexicon

• Lexical entries, attributes

  - lemma: cure
    ID:
      out: {sbj: !, obj: !}

XDG: parsing

L³

• An adaptation of XDG for translation
• A multilingual sentence "analysis"
  
  

L³: cross-lingual links

• Cross-lingual links: join lexical entries in different languages
• Optionally include a constraint attribute on an interface dimension

- lemma: cure
  cross:
    sem:
      lex: CURE
      IDSem:
        linkend: {arg1: [sbj]}
    am:
      lex: adane
      IDID:
        linkend: {sbj: [sbj]}

L³: translation

Shallow and deep translation (1)

• Deep translation: language-independent semantic level (interlingua), rules mapping SL lexical and grammatical units to semantic units, rules mapping semantic units to TL lexical and grammatical units
• Shallow translation: (transfer) rules map SL lexical and grammatical units directly to TL lexical and grammatical units
• Advantage of deep MT
  • Possible to translate from any of system's SLs to any TLs
• Advantages of shallow MT
  • No need for abstract general representations
  • No need to update entire system when new language is added
  • More efficient than deep MT

L³: shallow and deep translation (2)

L³: shallow and deep translation (3)

• Translation of the doctor cured the patient into Amharic
  • Deep: 3544 constraints
  • Shallow: 2879 constraints, 28.1% savings in time required
• Translation of ሐኪሙ በሽተኛውን አዳነው into English
  • Deep: 5594 constraints
  • Shallow: 4371 constraints, 30.3% savings in time required

L³: node mismatch (1)

• Words in input sentence do not correspond to explicit semantic units: del arcs in XDG
• Another solution required for case where nodes are inserted
  • Generation: node in output that has no corresponding node in semantic input
  • Translation: source language has no node corresponding to semantic and/or target-language word

L³: node mismatch (2)

• Zero subjects and objects in Amharic

  1. ሐኪሙ	በስተኛዋን	አዳናት
     hakimu	bǝštǝňňawan	adanat
     the-doctor(mas.)	the-patient(fem.,acc.)	he-cured-her
     'The doctor cured the patient.'

  2. (እሱ)	(እሷን)	አዳናት
     ɨssu	ɨsswan	adanat
     he	her	he-cured-her
     'He cured her.'

L³: node mismatch (3)

• Empty nodes created by "trigger" nodes

  - lemma: cure
    empty: [@SBJ, @OBJ]
  

• Empty nodes either used (playing role in another dimension) or deleted if not needed

L³: node mismatch (4)

L³: node mismatch (5)

Structural divergences in MT (1)

• Structural divergences (Dorr, 1994; Mel'čuk & Wanner, 2006)
• English-Amharic example illustrating thematic and categorial divergences
  • Amharic impersonal experiential verbs
    • ደከመ dǝkkǝmǝ 'be tired', ሰለቸ sǝlǝččǝ 'be bored', ራበ rabǝ 'be hungry', ጠማ t'ǝmma 'be thirsty'
    • Verb subject agreement always 3rd person singular masculine ('it')
    • Verb object agreement with experiencer
    • Explicit experiencer argument in "topic" (not accusative) form

Structural divergences in MT (2)

1. (እሷ)	ደከማት
   ɨsswa	dǝkkǝmat
   she	it-tired-her
   'She is tired.'

2. ልጆቹ	ደከማቸው
   lɨjočču	dǝkkǝmaččǝw
   the-children	it-tired-them
   'The children are tired.'

Structural divergences in MT (3)

• Thematic divergence
  • Experiencer is subject in English
  • Experiencer is topic (if explicit), obligatorily cross-indexed as object on verb in Amharic
  • Experiencer is arg1 in Semantics
• Categorial divergence
  • Experience takes the form of an adjective + be (the ID root) in English
  • Experience takes the form of a dedicated verb, ደከመ (the ID root), in Amharic
  • Experience is the root (be is deleted) in Semantics

L³: structure mapping (1)

AMHARIC
- lemma: dekeme
  empty: [@TOP]
  ID:
    out: {top: !}
    agree: [[top, obj, png]]
  cross:
    sem:
      lex: TIRED
      IDSem:
        linkend: {arg1: [top]}

L³: structure mapping (2)

ENGLISH
- lemma: be_padj
  ID:
    out: {padj: !}
  cross:
    sem:
      lex: zero
- lemma: tired
  ID:
    in: {padj: !}
  cross:
    sem:
      lex: TIRED
      IDSem:
        linkend: {arg1: [sbj]}

L³: structure mapping (3)

Project status

• Morphological analyzer/generator for Amharic (HornMorpho)
• Reimplementation of XDG in Python 3
• Lexicon in YAML format
• Limited grammar
• All software open source: http://www.cs.indiana.edu/˜gasser/Research/software.html

Ongoing, future work

• Extension of Amharic and English grammars
• Application to translation in restricted domain of mathematics
• Integration into CAT system
• Machine learning for disambiguation

Conclusions

• Bidirectional RBMT
• Handles structural divergences
• Integrates shallow and deep translation
• Focus on features that distinguish the languages

References

• Debusmann, R. (2007). Extensible Dependency Grammar: a modular grammar formalism based on multigraph description. PhD Dissertation, Universität des Saarlandes.
• Debusmann, R., Duchier, D., and Kruijff, G-J. M. (2004). Extensible Dependency Grammar: a new methodology. Proceedings of the COLING 2004 Workshop on Recent Advances in Dependency Grammar.
• Dorr, B. (1994). Machine translation divergences: a formal description and proposed solution. Computational Linguistics, 20, 597-633.
• Forcada, M. L., Ginestí-Rosell, M., Nordfalk, J., O’Regan, J., Ortiz-Rojas, S., Pérez-Ortiz, J. A., Sánchez-Martínez, F., Ramírez-Sánchez, G., and Tyers, F. M. (2011). Apertium: a free/open-source platform for rule-based machine translation. Machine Translation, 25, 127-144.
• Gasser, M. (2010). A dependency grammar for Amharic. Workshop on Language Resources and Human Language Technologies for Semitic Languages.
• Gasser, M. (2011). Towards synchronous Extensible Dependency Grammar. International Workshop on Free/Open-Source Rule-Based Machine Translation, 2.
• Mel'čuk, I., and Wanner, L. (2006). Syntactic mismatches in machine translation. Machine Translation, 20, 81-138.
• Ranta, A. (2010). Creating linguistic resources with GF. LREC 2012.