Informal presentation

1
Informal presentation

• Jian-Yun Nie
• My lab (easy)
• My view of NLP (tough)
• My experience in doing research (may be useful)

2
1. Presentation of RALI lab.

• Dept. of Computer science and operations
  research, University of Montreal, Canada
• RALI Recherche Appliquee en Linguistique
  Informatique (Applied Research in Computational
  Linguistics)
• One of the largest labs in CL in Canada
• Long tradition of MT

3
Personnel

• 3 professors
• G. Lapalme Text generation, Info. Extraction,
  QA, Text summarization
• Ph. Langlais Statistical MT
• J.Y. Nie IR and some NLP
• 5 researchers
• 20 graduate students

4
Research projects

• Translation
• TransType suggest translation words/expressions
• TransTalk Speech input of translation
• TransSearch Search for translations in parallel
  texts (Hansards)
• TransCheck check for trans. errors
• IE
• Automatic email answering (from customers)
• QA
• Text summarization
• Scientific articles simulate a human summarizer
• Law documents
• Arabic texts

5
Research projects

• IR
• IR models (logic models, language models, etc.)
• Inferential IR (how to reason in IR?)
• CLIR (English-French, Chinese, etc.)
• Text mining (mining parallel texts from the Web)
  for query translation
• Business intelligence (application of IR)
• Fining relevant calls for tenders
• Mining info. about companies
• Classification of calls for tenders

6
Current research projects in IR

• Integrating term relationships into language
  models
• CLIR using statistical translation model
• Improving translation models for query
  translation
• Mining parallel texts and filtering noise
• Classification using language models, compound
  terms etc.
• User- and domain-dependent IR (user profile)
• Law document retrieval using hyperlinks
• Image retrieval (cross-media IR)

7
Tools

• SILC (automatic identification of language and
  coding)
• PTMiner (Parallel Text mining)
• TransSearch (transferred to a company)

8
2. My view of NLP

• Many interesting subjects to be explored
• Combining rules with statistics
• Exploring more learning methods for NLP
  (exploiting the mass of data)
• History
• Rule-based
• Statistical approaches
• McKey
• Proper places of man and machine
• Combination of rules and statistics
• Statistics can help choose appropriate rules
• Many rules cannot be learnt through statistics
  (at least for now)

9
Example Machine aided human translation

• Statistics can help select translation
  words/terms
• But is unable to deal with grammars and meaning
• Machine makes suggestions, and human makes the
  final translation

10
Example MT

• Translations made by statistical MT are often
  unreadable
• Why not use rules to improve readability?
• Question what is the proper places of man
  (rules) and machine (statistics)? How to make
  them working together?

11
Exploiting the mass of data

• New translations extract from data (the Web)
• New terms
• New rules
• At least frequency

12
3. My own experience in doing research

• Paper structure
• Introduction of the problem, motivation of the
  research
• Existing methods
• New methods
• Modification of an existing one
• Borrow a method from a different area
• Create a completely new method
• Implementation
• Experiments/analysis
• Conclusion
• Research planning similar to paper structure

13
Two types of researchers

• I have a hammer, and everything looks like a nail
  (opportunistic)
• I have a problem, and I try to use every possible
  method

14
My way of doing research

• Determining an interesting research problem
• Not solved, not trivial
• Possibility to find a solution
• In my area of research
• Being aware of the existing methods
• Reading papers, but with a critical eye
• Finding possible improvements, new solutions
• First intuitively, then formulate with more rigor
• Implementation and experiments
• Analysis of the results (to find possible
  improvements, and to understand better the
  problem)

15
My way of doing research

• Improve the writing skills (very critical for a
  researcher)
• A long learning process
• Observe how the others write papers and correct
  my papers
• Imagine Im a reader, and try to see if the
  description can follow smoothly, and is clear to
  a non-specialist reader
• Is the logic of my description clear, and easy to
  follow?
• Whole structure
• Description of a specific element
• Is a term used ambiguous and not clear? Can I use
  a better term?
• Is the description monotonic/repetitive/boring
• Is there a hole?
• Have I made my idea clear enough? How can I make
  it clearer?
• The reader (reviewer) wants to understand, so try
  to help him/her
• What is not clear is not necessarily profound, it
  can also be obscure
• Complex math. formulas ? good paper

16
Some common problems

• Superficial
• Introducing new concepts without introducing new
  methods
• Too ambitious but without concrete support
• Doing research in isolation
• No literature review
• No comparison with others
• No (not enough) justification of the method
• Logic not clear in description
• Big holes (an important notion not explained,
  two pieces of description are not coherent, )
• Unrelated description
• Experiments not rigorous
• Trace of manipulation of experimental results
• English (but often not the main reason)

17
Recommendation

• Read papers
• Observe the writing
• Write papers
• Read your own papers as a reader
• Ask your colleagues to read your papers and to
  tell you what is not clear
• Read and analyze carefully the corrections by the
  others