EUREKA: Dialogue-based IR

1
EUREKA Dialogue-based IR

• Antoine Raux
• IR Seminar and Lab 11-743 Fall 2003
• Mid-term Presentation

2
Goal

• Build a light-weight, natural language
  dialogue-based interface to refine queries and
  select relevant documents.

3
Overview

• Backend Vivisimo (using web scraper)
• Dialogue Management RavenClaw (successor of CMU
  Communicator)
• Language Understanding Light Open Vocabulary
  Parser, created for this project. Very simple so
  far (first wordcommand, other wordsarguments)
• NLG/TTS template-based Festival

4
Vivisimo

• Available clustering meta-search
  enginewww.vivisimo.com
• Hand-written Perl web scraper (hope Vivisimo
  doesnt change their page design by the end of
  the semester)

5
RavenClaw

• Semi-declarative language (C macros)
• Hierarchical agent architecture

EUREKA
Greet User
Prompt Query
New Query
Open Cluster
Close Cluster

Submit Query
Get Cluster List
Get Doc List
Inform of Results
6
LOV Parser

• Problem traditional NL parsers require a
  dictionary of all valid words ? not applicable to
  open domain IR
• Solution (implemented in C)
• fix a small number of one-word commands
  (new_query, open, list_clusters)
• parse each line as command arguments or
  command or arguments

7
NLG/TTS

• Template-based Language Generation (e.g. I
  found ltn_docgt documents.)
• General purpose Festival voice for TTS
• NB Browsing through lists is not efficient
  with speech, even for lists of clusters

8
Already Implemented

• Working prototype
• Commands
• new_query
• list_clusters, list_documents
• open, close (cluster)
• more, back (list of clusters/documents)

9
Demo
10
Future Work

• More commands summaries, query modification
  (with , -, , )
• Based on literature (e.g. Belkin), design search
  strategies
• Design/conduct evaluation experiment

11
To be continued
12
Possible Actions by User

• Make a new query
• Refine/reformulate a query
• Select a cluster
• Ask for a document summary
• Ask for a cluster summary

13
Possible Actions by the System

• Give a list of doc titles
• Give a list of clusters
• Give the summary of a document
• Give the summary of a cluster
• Propose a query refinement/reformulation