Questions answered, a lesson learned

1
Questions answered, a lesson learned

• CS2731 NLP
• Final Project Report
• Brano Kveton
• Tomas Singliar

2
System architecture

• Cascade of scoring modules
• Each stage adds a score for sentence
• Last stage picks the best one
• Perl/Java implementation
• Using LingPipe NE tagger
• Efficiency forgone to modularity
• Shallow processing
• No parse tree, coref resolution, semantics,
  inference

3
Modules

• Q-type recognized (pattern match)
• NE tagger
• Improves WHO and WHERE questions
• BOW
• All questions, strongest module
• Dateline recognizer
• WHEN questions, correlated words
• DateTime recognizer

4
Modules (continued)

• DateTime tagger
• Why questions correlated words
• Because, due to,
• Why questions pre/post
• Sentence before or after best BOW-match is a
  likely answer

5
Performance

• Overall, we did a so-so job
• 61, 57, 62 - Tr, T1, T2
• Final preliminary (Test1)
• WHEN 9
• WHERE 8
• WHAT 5
• WHY 2
• WHO 0
• HOW 1
• Improvements occur where easy
• Blame shallow processing hard questions have
  more structure

6
Lessons

• Children are smarter than PhDs
• ML folk wisdom on overfit
• helps in start
• hurts in finish
• Difficult questions need deeper processing

7
Thank you! (the exclamation mark slide!)

• Congratulations to the winner!
• Festivities begin!
• Questions most welcome!