NICE: Native Language Interpretation and Communication Environment

1
NICE Native Language Interpretation and
Communication Environment

• Lori Levin, Jaime Carbonell, Alon Lavie, Ralf
  Brown, Erik Peterson, Katharina Probst, Rodolfo
  Vega, Hal Daume
• Language Technologies Institute
• Carnegie Mellon University
• April 12, 2001

2
NICE

• Rapid development of machine translation for low
  and very low density languages

3
Classification of MT by Language Density

• High density pairs (E-F, E-S, E-J, )
• Statistical or traditional MT approaches are O.K.
• Medium density (E-Czech, E-Croatian, )
• Example-based MT (success with Croatian, Korean)
• JHU initial success with stat-MT (Czech)
• Low density (S-Mapudungun, E-Iñupiaq, )
• 10,000 to 1 million speakers
• Insufficient bilingual corpora for SMT, EBMT
• Partial corpus-based resources
• Insufficient trained computational linguists

4
Machine Translation of Very Low Density Languages

• No text in electronic form
• Cant apply current methods for statistical MT
• No standard spelling or orthography
• Few literate native speakers
• Few linguists familiar with the language
• Nobody is available to do rule-based MT
• Not enough money or time for years of linguistic
  information gathering/analysis
• E.g., Siona (Colombia)

5
Motivation for LDMT

• Methods developed for languages with very scarce
  resources will generalize to all MT.
• Policy makers can get input from indigenous
  people.
• E.g., Has there been an epidemic or a crop
  failure
• Indigenous people can participate in government,
  education, and internet without losing their
  language.
• First MT of polysynthetic languages

6
New Ideas

• MT without large amounts of text and without
  trained linguists
• Machine learning of rule-based MT
• Multi-Engine architecture can flexibly take
  advantage of whatever resources are available.
• Research partnerships with indigenous communities
• (Future Exponential models for data-miserly SMT)

7
History of NICE

• Arose from a series of joint workshops of NSF and
  OAS-CICAD.
• Workshop recommendations
• Create multinational projects using information
  technology to
• provide immediate benefits to governments and
  citizens
• develop critical infrastructure for communication
  and collaborative research
• training researchers and engineers
• advancing science and technology

8
Approach

• Machine learning
• Uncontrolled corpus (Generalized Example-Based
  MT)
• Controlled corpus elicited from native speakers
  (Version Space Learning)
• Multi-Engine MT
• Flexibly adapt to whatever resources are
  available
• Take advantage of the strengths of different MT
  approaches

9
Evaluation Objective

• To achieve a given level of translation quality
  for a series of languages L1 to Ln
• Reduce the amount of training data required
• Reduce the amount of language-specific
  development time after language-independent
  software has been developed

10
Evaluation Baseline From Previous Work
(Generalized EBMT)

• High density languages (French, Spanish)
• 1MW parallel corpora (e.g., subset of Hansards)
• Consistent spelling, grammatically correct
• High coverage, gisting-quality translation

11
Evaluation Baseline GEBMT French Hansards
Coverage (in percent) as a function of corpus
size (in millions of words)
12
Long-Term Target Reduction in Linguistic and
Human Resources
13
Work Completed
14
Establishing Partnerships
15
NICE Partners
16
Nice/MapudungunCurrent Products

• Writing conventions (Grafemario)
• Glossary Mapudungun/Spanish
• Bilingual newspaper, 4 issues
• Ultimas Familias memoirs
• Memorias de Pascual Coña
• 6 hours transcribed speech
• 40 hours recorded speech

17
Instructible Rule-Based MT
18
iRBMT Instructible Rule Based MT
19
Elicitation Process

• Purpose controlled elicitation of data that will
  be input to machine learning of translation rules

20
Elicitation Interface Example
21
Elicitation Interface

• Native informant sees source language sentence
  (in English or Spanish)
• Native informant types in translation, then uses
  mouse to add word alignments
• Informant is
• Literate
• Bilingual
• Not an expert in linguistics or in linguistics or
  computation

22
The Learning Process

• Learning Instance
• English the big boy Hebrew ha-yeled ha-gadol
• Acquired Transfer Rule
• Hebrew NP N ADJ ltgt English NP the ADJ
  N
• where (HebrewN ltgt English N)
• (HebrewADJ ltgt EnglishADJ)
• (HebrewN has ((def )))
• (HebrewADJ has ((def )))

23
Standard Version Space Learning

• Hypothesis Space of all possible rules consistent
  with data seen so far
• Represented by a generalization lattice bounded
  by S (most specific) and G (most general)
  boundaries
• New positive instances (translation pairs)
  generalize S
• New negative instances (incorrect translations)
  specialize G
• Converge when S and G intersect
• Problem worse case exponential blow-up

24
Locally-Constrained, Seeded Version Spaces

• Preferred generalization level (e.g.
  Parts-of-speech linguistic features semantic
  features)
• First translation pair generalized to preferred
  level gt seed the VS
• Define P max levels of seed generalization or
  specialization (i.e. how close is initial guess)
• Generate S/P and G/P boundaries, and apply VS
  learning
• Allow mutation operator if S/P and G/P prove
  incorrect

25
Advantages of Seeded Version Spaces

• Worst case polynomial with degree P gt
  "tractable"
• Generalization level can be estimated reasonably
  well for MT transfer rules gt good seeds
• Faster convergence, requiring less training data

26
Version Space Abstraction Lattice
27
The Elicitation Corpus

• List of sentences in a major language
• English
• Spanish
• Dynamically adaptable
• Different sentences are presented depending on
  what was previously elicited
• Compositional
• Joe, Joes brother, I saw Joes brother, I told
  you that I saw Joes brother, etc.
• Aim for typological completeness
• Cover all types of languages

28
Pilot Version of Elicitation Corpus

• Approximately 800 sentences
• Tested on Swahili
• Vocabulary
• Include a variety of semantic classes e.g.,
  animate, inanimate, man-made objects, natural
  objects, etc.
• Noun phrases
• Detect number, gender, types of possessives,
  classifiers, etc.
• Basic sentences
• Detect agreement between verb and subject and/or
  object, basic word order, problems with
  indefinite or inanimate subjects, etc.
• Complex constructions
• Currently relative clauses. Later, comparatives,
  questions, embedded clauses, etc.

29
Detection of Grammatical Features

• Each language uses a different inventory of
  grammatical features tense, number, person,
  agreement.

Swahili The hunter kill-ed the animal Mwindaji
a-li-mu-ua mnyama a class-one subject li past
tense mu class-one object ua kill
Fox (Algonquian) Ne-waapam-aa-wa I-see-direct-him
Ne-waapam-ek-wa me-see-indirect-he
30
Organization of Tests
Dual
Plural
Paucal
Diagnostic Tests

Subj-V Agr


31
Demo of Elicitation Interface and Feature
Detection
32
Data Collection
33
Mapudungun Data

• Spanish-Mapudungun parallel corpora
• Total words 223,366
• Spanish-Mapudungun glossary
• About 5500 entries
• 40 hours of speech recorded
• 6 hours of speech transcribed
• Speech data will be translated into Spanish

34
Progress and Plans
35
Summary of Year 1Partnerships

• Establishment of a partnership with the Institute
  for Indigenous Studies at the Universidad de la
  Frontera (UFRO) in Chile.
• Establishment of a partnership with the Chilean
  Ministry of Education.
• Identified partners in Alaska and Colombia.
  Details of the partnership are being discussed.

36
Summary of Year 1 Data

• Spanish-Mapudungun parallel corpus over 200,000
  words
• Standardization of orthography Linguists at UFRO
  have evaluated the competing orthographies for
  Mapudungun and written a report detailing their
  recommendations for a standardized orthography
  for NICE.
• Training for spoken language collection In
  January 2001 native speakers of Mapudungun were
  trained in the recording and transcription of
  spoken data.
• Mapudungun spoken language corpus 40 hours
  recorded, 6 hours transcribed (as of end of
  February).

37
Summary of Year 1 iKBMT

• Preliminary design of transfer rule formalism for
  machine translation.
• Design and pilot testing of prototype elicitation
  corpus.
• First prototype of feature detection
• Morphological processing in PC Kimmo covering
  about 40 Mapudungun morphemes.
• Preliminary version of new parser for run-time
  translation component.

38
Goals for Year 2 Data

• Continue collection, transcription, and
  translation of Mapudungun data.
• Take inventory of existing Inupiaq data available
  from the Alaska Native Languages Center and the
  Inupiaq community.
• Focus on the North Slope dialect and other
  dialects that are easily intelligible to North
  Slope speakers. Type and record additional
  Inupiaq data as needed.
• Plans for Siona data collection will be discussed
  at a meeting in Bogota in May.

39
Goals for Year 2 Elicitation Corpus

• Extend the elicitation corpus with more complex
  constructions (such as causatives and
  comparatives) and add diagnostics for complex
  features such as the tense and aspect system.
• Refine elicitation interface based on preliminary
  experiments.
• Preliminary user studies with the corpus and
  interface using at least two languages.
• Refine the linguistic corpus so as to accelerate
  learning of the more common and useful structures
  first.

40
Goals for Year 2 EBMT

• Baseline EBMT systems for Mapudungun and Inupiaq.
• Extend baseline systems with preliminary version
  of linguistic generalization.

41
Goals for Year 2 MT Run-time System

• Develop learnable transfer-rule structure and
  interpreter.
• Unlike existing hand-coded transfer system for
  machine translation, a learnable structure
  requires full compositionality and
  component-wise generalizability/specializability
  for data-driven inductive learning.
• Develop morphological processors and part of
  speech taggers for Mapudungun and Spanish.

42
Goals for Year 2 Version Space Learning

• Develop baseline Seeded-Version-Space (SVS)
  inductive learning method
• Extend the elicitation interface to enable the
  SVS system to generate questions for the native
  informant, so as to speed the transfer-rule
  learning process

43
Future Projects

• Discussion

44
Appendix
45
The IEI Team

• Coordinator (leader of a bilingual and
  multicultural education project)
• Distinguished native speaker
• Linguists (one native speaker, one near-native)
• Typists/Transcribers
• Recording assistants
• Translators
• Native speaker linguistic informants

46
Agreement Between LTI and Institute of Indigenous
Studies (IEI), Universidad De La Frontera, Chile

• Contributions of IEI
• Socio-linguistic knowledge
• Linguistic knowledge
• Experience in multicultural bilingual education
• The use of IEI facilities, faculty/researchers
  and staff for the project
• electronic network support and computer technical
  support

47
Agreement between LTI and Institute of Indigenous
Studies (IEI), Universidad de la Frontera, Chile

• Contributions of LTI
• Equipment four computers and four DAT recorders
• Payment of consulting fees pending funding from
  the Chilean Ministry of Education
• Expertise in language technologies

48
LTI/IEI Agreement

• Cooperate in expanding the project to convergent
  areas, such as bilingual education, as well as in
  pursuing additional funding

49
MINEDUC/IEIAgreement Highlights

• Based on the LTI/IEI agreement, the Chilean
  Ministry of Education got involved in funding the
  data collection and processing team for the year
  2001. This agreement will be renewed each year,
  as needed.

50
MINEDUC/IEI Agreement

• Objectives
• To evaluate the NICE/Mapudungun proposal for
  orthography and spelling
• To collect an oral corpus that represent the four
  Mapudungun dialects spoken in Chile. The main
  domain is primary health, traditional and
  Occidental.

51
MINEDUC/IEI Agreement

• Deliverables
• An oral corpus of 800 hours recorded,
  proportional to the demography of each current
  spoken dialect
• 120 hours transcribed and translated from
  Mapudungun to Spanish
• A refined proposal for writing Mapudungun

52
Mapudungun Morphology

• kudu.le.me.we.la.n
• lay_down.st.Hh.rem.neg.ind.1S
• I am not going to lay down there any more
• illku.faluw.kUle.n
• get_angry.SIM.ST.IND.1s
• I am pretending to be angry
• antU.kUdaw.kiaw.ke.rke.fu.y
• day.work.CIRC.CF.REP.IPD.IND.3s
• he used to work here and there as a day laborer,
  I am told
• wisa.ka.dungu.fe.nge.y.mi
• bad.VERB.FAC.speak.NOM.VERB.IND.2s
• you are someone who always does and says nasty
  things