AVENUE Automatic Machine Translation for lowdensity languages

1
AVENUEAutomatic Machine Translation for
low-density languages

• Ariadna Font Llitjós
• Language Technologies Institute
• SCS
• Carnegie Mellon University

2
2 HCI project proposals

• Interface to Online Bilingual and Multilingual
  Dictionaries
• Translation Correction Tool interface design,
  implementation and user studies

3
Online Bilingual and Multilingual Dictionaries

• bilingual and multilingual dictionaries for
  indigenous languages (Mapudungun Chile, Inupiaq
  Alaska, Aymara, Quechua and Aguaruna Peru)
• For each bilingual/multilingual dictionary, we
  (will) have an excel database created by the
  local teams (Mapudungun from a spoken corpus
  transcribed and translated into Spanish)

4
(No Transcript)
5
Online Bilingual and Multilingual Dictionaries
(cont.)

• For each entry, we give the translation in
  Spanish, some other linguistic information (POS),
  and a link to the actual sentence where it
  appears in the corpus. For example
• Püñpüñkünuukey se manifiesta en forma de
  ronchas
• nmlch-nmfhp1_x_0031_nmfhp_00
•       Mapu Fey itrofillpüle kuerpu, ta pichike
  püñpüñkünuukey ta kalül may, peñi.  Sp Así
  es en todas partes del cuerpo, pequeñas ronchas
  se forman en el cuerpo pues, hermano

6
(No Transcript)
7
Online Bilingual and Multilingual Dictionaries
(cont.)

• Currently, users can search for
• Mapudungun words
• Spanish words
• all the words starting with a letter
• all the words containing a word or a string of
  characters

8
Online Bilingual and Multilingual Dictionaries
(cont.)

• Primary users
• people in the indigenous communities
• researchers in these countries, inside and
  outside the indigenous communities
• Chilean case
• product of the Ministry of Education.
• students and teachers, mostly Mapuche, but maybe
  some Spanish users as well

9
Online Bilingual and Multilingual Dictionaries
(cont.)

• Secondary users
• Linguistic, Lexicography and Anthropology
  researchers from all over the world
• random people browsing the www

10
Online Dictionaries Tasks for HCII project

• analyze design of the basic web interface
• given a query for a word in either language, it
  presents the information for that entry to the
  user in the other language.
• how to incorporate an audio file with the word as
  it was pronounced in the spoken corpus.
• how to make it interactive, i.e. have bilingual
  users comment on the entries and possibly add new
  entries (need profile info)

11
Translation Correction Tool (TCTool)

• AVENUE is a project which developsAutomatic
  Machine Translation Systems for low-density
  languages
• Since translations are automatic, i.e. not
  perfect, we need to refine them.
• instead of having a professional translator, we
  want to find an automatic way to refine the
  output of the MTS -gt TCTool

12
TCTool

• We can use the TCTool to automatically learn a
  refinement of the Transfer rules in our MTS, from
  users input
• Challenges
• users most likely not familiar with computers -gt
  user-friendly and Intuitive interface
• bilingual informants cant be assumed to have any
  linguistic knowledge

13
Automatic Machine Translation
Interlingua
interpretation
Transfer rules
Corpus-based methods
generation
analysis
14
Automatic Learning of a Transfer-based MTS
tentative Transfer rules
SVS algorithm
Elicitation corpus
Transfer module
Rule Refinement module
(tentative) TL sentences
SL sentences
15
Interactive and Automatic rule refinement

• Interactive step (TCTool)
• Given an MTS, translate sentences and
• present them to the users for minimal
• correction (interface design, MT error
  classification)
• Automatic step
• Machine learning DS and algorithms to map user
  input with refined transfer-rules

16
User studies snapshot
17
(No Transcript)
18
(No Transcript)
19
TCTool Tasks for HCII project

• analyze design of the basic web interface
• given a translated sentence, it asks the user to
  minimally correct it, if incorrect, and to
  classify the error(s).
• how to explain what minimally correction is
• what is the right error classification for
  non-expert and non-linguist users
• Can naïve users reliably pinpoint the source of
  errors?
• design User Studies to show reliability of user
  input (Spanish English, English Spanish,
  English Chinese)

20
AVENUE project members

• LTI team
• Researchers  Ph. D. students
• Jaime Carbonell Ariadna Font Llitjós
• Lori Levin Christian Monson
• Alon Lavie Erik Peterson
• Ralf Brown Katharina Probst
• Avenue External Project Coordinator 
• Rodolfo M Vega,
• Chilean team
• Eliseo Cañulef Luis Caniupil Huaiquiñir
• Hugo Carrasco Marcela Collio Calfunao
• Rosendo Huisca Cristian Carrillan Anton
  Hector Painequeo Salvador Cañulef
• Flor Caniupil Claudio Millacura

21
Questions?

• For more information
• http//www.cs.cmu.edu/aria/avenue/