An autonomous email replier

1
An autonomous e-mail replier

• Charlie Westergren
• CSI5386 Natural Language Processing
• March 31, 2006

2
Agenda

• Introduction
• Previous work
• The idea
• My implementation
• Issues
• Results
• Conclusions
• Questions

3
Introduction

• System that automatically answers e-mails
• Not
• Mr. Westergren is currently out of town, he will
  be back again on Monday.
• Extended version
• Answers questions
• Receives information

4
Who (why)?

• Possible users
• Stressed employees
• Stressed students
• People who dislikes writing e-mails?
• Ethical?
• Add notice, with reason

5
Other approaches

• No found systems
• Question answering systems
• Chatterbox (similar in some ways)
• ELIZA, Jabberwacky,

6
The idea

• Create a knowledgebase
• Simplified algorithm
• Structure e-mail
• Extract key words (nouns, pronouns)
• Look for entities in knowledgebase
• Extract information for answer
• Create answer
• For non-questions, add to knowledgebase?

7
My implementation

• Python nltk (Natural Language Tool-Kit)
• Define a simple Context Free Grammar (CFG)
• Tag the e-mail with a Brown bigram tagger
• Add tagged words to the CFG lexicon
• Build parse trees from sentences and CFG
• Extract key-words and find entries in the
  knowledge base

8
Implementation - issues

• Generation of answers
• Proper knowledgebase
• Synonym handling
• Connections
• Anaphora
• Adding info to the knowledgebase

9
Implementation - knowledgebase

• personal
• ('keys', ('Tim', 'name'), ('you',
  'relation')), ('relation', ('brother', john),
  ('sister', lisa), ('mother', sue), ('father',
  tom)),
• ('recent events', christmas, new_year,
  reading_week, last_weekend),
• ('upcoming events', this_weekend, easter)

10
Implementation - knowledgebase

• very ad-hoc solution
• christmas ('keys', ('Christmas', 'name')),
• ('what', 'skiing', 'resting'),
• ('where', 'Mont Tremblant'),
• ('who', 'me', 'my girlfriend'),
• ('preps', 'for', 'were', 'in')

11
Implementation - result

• Question found ' What did you do for Christmas '
  parsing... Done.
• QWORD gtgt 'What'
• VERB(S) gtgt 'do'
• KEY(S) gtgt 'Christmas', 'you'
• MATCH Christmas name
• MATCH II What 'skiing', 'resting'
• ANS gtgt For Christmas me and my girlfriend were
  skiing and resting in Mont Tremblant.

12
Conclusions and further work

• Difficult to structure up knowledgebase
• Hard to create psychologically realistic answers
• E-mails are often very unstructured
• Adding information to the database

13
Thank you

• Questions?
• send them to me in an e-mail )