The Hindi Surprise Language Exercise

1
The Hindi Surprise Language Exercise

• Language Technologies Institute
• Carnegie Mellon University

2
Agenda

• Overview of the DARPA TIDES project
• What is a surprise language exercise
• Quick intro to Machine Translation
• List of job tasks

3
Language Technologies Institute

• A department in the School of Computer Science
• Language Technologies translating one natural
  language to another, speech recognition,
  information retrieval, information summarization,
  automatic question answering
• We have courses that undergraduates can take, a
  masters degree program, and a Ph.D. program.

4
Overview of the DARPA TIDES Project

• DARPA Defense Advanced Research Projects Agency
• Defense and non-defense research ARPA Net was
  the pre-cursor of the internet
• TIDES Translingual Informantion Detection
  Extraction and Summarization

5
Overview of the DARPA TIDES Project

• Make news sources in other languages available to
  English speakers
• Automatically translate
• Automatically extract specific facts
• Automatically summarize
• How well does it work?
• Promising so far
• Participants
• Over 20 universities, government contractors, and
  government agencies
• Probably between 100 and 200 people across the US

6
What is a surprise language exercise?

• Commercially available language technologies
  (such as machine translation) take many years to
  develop.
• What if you needed language technologies for a
  new language quickly (in a matter of days or
  weeks)?
• The surprise language is a test of the TIDES
  research to see if an translation, extraction,
  and summarization can be built in one month for a
  new language.

7
Surprise Language Timeline

• June 2, a.m., The name of the surprise language
  is announced Hindi
• June 30 Exercise ends.

8
Quick intro to machine translation

• Machine Translation automatically translating
  from one human language to another human language
• The ideal system
• High quality as good as a human translator
• Broad coverage any topic, any type of text or
  speech
• There is no ideal system currently

9
Tradeoffs based on the task

• Dissemination you want people to read something
  that you wrote (e.g., a users manual) requires
  high quality, but not broad coverage
• Conversation medium quality is ok. People can
  compensate. Conversational systems are currently
  do not have broad coverage.
• Assimilation you want to gather information
  from documents in another language broad
  coverage is critical gisting quality is
  acceptable.

10
Methods for Machine Translation

• Humans write translation rules
• English be NOUN (e.g., be a teacher)
• Hindi NOUN he
• Example-Based Machine Translation
• Rules are learned automatically from parallel
  corpora.
• Statistical Machine Translation
• Probabilities are learned automatically from
  parallel corpora.

11
What is a parallel corpus?

• The same text in two languages.
• Corresponding sentences are aligned.
• EBMT and SMT require huge amounts of parallel
  text millions of words long.

12
Job Tasks

• For someone who is not a linguist or a
  programmer
• Align sentences in parallel text
• Make sure that parallel texts are really parallel
• Align words in parallel sentences
• Translate sentences
• Add linguistic information to dictionaries e.g.,
  whether nouns are masculine or feminine.
• Check the quality of the translations that the
  system produces
• For someone who is a linguist Work with Lori
  and some graduate students to identify rules.
• For someone who is a programmer
• Help run alignment algorithms
• Help to automatically identify and normalize
  temporal expressions (e.g., last year, on
  Tuesday, in 1993)
• Help to automatically identify and normalize
  proper names, etc.

13
To get employed

• Go to the payroll office at 407 S. Craig St.
• Ask to fill out an I9 form.
• Bring a copy of the I9 form to Cheryl Webber
  (assistant business manager of LTI)
• Or fax a copy of the I9 form to Cheryl Webber at
  312-268-6298.
• Fill out an information sheet.