CLMS Career Exploration Talk

Oct 9, 2012

About Me

• BA, Linguistics from UW in 2008
• Graduated from CLMA in 2010
• Currently works as a software developer at Substantial

Overview

• Kiha Software
• Aro Mobile, Kiha's product
• Aro and NLP
• Hackathon!
• What was most useful from CLMA
• What I learned about NLP at Kiha
• Questions?

• Founded in 2008
• Launched Aro Mobile in 2010
• I joined in Sept, 2010

Aro Mobile

• Service and personal information manager (PIM) for Android that:
  • Understands your data
  • Makes you faster and more efficient at managing email, contacts, calendar

Aro Mobile -- how does it work?

Aro Mobile -- how does it work?

Hi Josh,

Could you give Mark a call at 3pm tomorrow?

Thanks,

Bob

Bob the Builder
206 123 1234
bob@bigbricks.com
1337 Pro Builder Lane
Seattle, WA 98122
                    

Aro Mobile's Email Client

Aro and NLP

• Extracted entities:
  • Names, places, businesses
  • Phone numbers
  • Tracking numbers (USPS, UPS, Fedex)
  • Flight numbers
  • Email signatures
  • Calendar events
  • Addresses
  • Email addresses

Aro and NLP

• Models
  • CRF trained IOB sequence labeler
  • Minorthird
• Data
  • Newswire corpora
  • Annotated email (in-house and Mechanical Turk)

What I did at Kiha

• Automating model training through continuous integration with Maven
• Feature engineering for NER models (POS tags, NP chunks)
• Improving performance, while maintaining precision/recall

Hackathon!

• Extending EES: recognizing conference call meeting numbers
• Hacking the Android app to auto-dial conference calls

Hi Bob,

Dora the Explorer invites you to attend this online meeting.

Topic: Computational Linguistics
Date: Monday, January 10, 2013
Time: 2:15 pm, Pacific Standard Time (San Francisco, GMT-08:00)
Meeting Number: 736 771 237
CCP:+16504293300x736771237#

What was most useful from CLMA

• 570, 572
• i2b2 Medical Extraction Challenge
• Experience with MALLET, libSVM
• Familiarity with what algorithms and techniques are available

What I learned about NLP at Kiha

• Use existing libaries -- most of my time was spent thinking about features (not algorithms)
• Sometimes boring techniques are good enough (e.g. name recognition from list of names)
• Learn to have a sense of what "good enough" performance is

Final words

• Kiha has since rebranded and changed their product direction
• For current job seekers: Hacker News Meetup, Seattle.rb, Beer && Code

Questions?