Einat Minkov Ramnath Balasubramanyan William W. Cohen

1
Einat Minkov Ramnath
Balasubramanyan William W.
Cohen
einat_at_cs.cmu.edu
rbalasub_at_cs.cmu.edu
wcohen_at_cs.cmu.edu
Novel

• Inter-Entity Similarity
• Graph Walks
• Graph edge l weighted by its type
• Pr(x y) the edge weight, normalized by
  total outgoing weight from x.
  
• Starting with initial distribution V, walk for K
  time steps
  
• Where M is the transition matrix and --
  decay rate.
  
• TF-IDF
• Vector similarityTask, Person, Folder Average
  (TFIDF vectors of related messages)

Person-Activity Prediction Given a folder that re
presents a project activity, at point in time T,
predict (rank) email-addresses to be associated
with the folder in the future tT.
Activity-Centric Search

• Given an underlying activity
• Short Term Tasks Schedule a meeting, prepare a
  report.. Or an ongoing Project activity e.g.,
  teaching a course. Search for related email
  messages, relevant persons, email-addresses,
  documents, forms, links etc.
• See also Activity-Centered Task Assistant
  (ACTA) BellottiThornton 06

This task is hard!Many possible candidates.
Recall (persons) at the top 10 ranks
Recall (persons) at the top 20 ranks
Activity Email Representation
Novel
To-do Email Search Find relevant emails for a giv
en to-do item.
Email Foldering and Message Tracking
Foldering For a given message, rank user
folders. Message tracking For a folder, rank
messages (relevant, but not associated..)
Evaluated on the Cspacecorpus (CMU). Manual
annotations.
Evaluated on folders representing a project
activity from the Enron corpus (4 users).
GW-U Graph walk, uniform weightsGW-M Manually
assigned weightTF-IDF-1 vector over words in
task plus headers of email that spawned the
taskRF-IDF-2 vectors of only words in task

• Nodes denote entities Person, Email-address,
  Message, Date, Terms and TASK.
  
• Directed edges represent relations such as
  sent-by, contains-term etc.
  
• Task linked to relevant Messages.

Precision at 5
Conclusions Graph walk performs well when task
uses mainly social network information graph
walk may improve with doc. Length normalization
We explored activity-centric search tasks in
email.
Foldering (MAP)
Msg tracking (MAP)
Machine Learning Dept. Language Technologies
Institute, Carnegie Mellon University
EMAIL Workshop - AAAI 08