Conversational Sales Assistant for Online Shopping
Margo Budzikowska, Joyce Chai, Sunil Govindappa, Veronika Horvath, Nanda Kambhatla,
Nicolas Nicolov & Wlodek Zadrozny
Conversational Machines Group
IBM T. J. Watson Research Center
30 Saw Mill River Rd, Hawthorne, NY 10532, U.S.A.
{sm1, jchai, govindap, veronika, nanda, nicolas, wlodz}@us.ibm.com
ABSTRACT
Websites of businesses should accommodate both customer
needs and business requirements. Traditional menu-driven
navigation and key word search do not allow users to describe
their intentions precisely. We have developed a conversational
interface to online shopping that provides convenient,
personalized access to information using natural language
dialog. User studies show significantly reduced length of
interactions in terms of time and number of clicks in finding
products. The core dialog engine is easily adaptable to other
domains.
1. INTRODUCTION
Natural language dialog has been used in many areas, such as
for call-center/routing application (Carpenter & Chu-Carroll
1998), email routing (Walker, Fromer & Narayanan 1998),
information retrieval and database access (Androutsopoulos &
Ritchie 1995), and for telephony banking (Zadrozny et al. 1998).
In this demonstration, we present a natural language dialog
interface to online shopping. Our user studies show natural
language dialog to be a very effective means for negotiating
user's requests and intentions in this domain.
2. SYSTEM ARCHITECTURE
In our system, a presentation manager captures queries from
users, employs a parser to transform the user's query into a
logical form, and sends the logical form to a dialog manager.
The presentation manager is also responsible for obtaining the
system's response from the dialog manager and presenting it to
the user using template-based generation. The dialog manager
formulates action plans for an action manager to perform back-
end tasks such as database access, business transactions, etc. The
dialog manager applies information state-based dialog strategies
to formulate responses depending on the current state, discourse
history and the action results from the action manager.
The Data Management Subsystem maintains a “concept”
repository with common sense “concepts” and a phrasal lexicon
that lists possible ways for referring to the concepts. Business
Rules map concepts to business specifications by defining
concepts using a propositional logic formula of constraints over
product specifications. Thus, the Business Rules reflect business
goals and decisions. The Extended Database combines product
specifications and precompiled evaluations of the concept
definitions for each product to provide a representation that
guides the natural language dialog. We are investigating
automated tools for helping developers and maintainers extract
relevant concepts and terms on the basis of user descriptions and
queries about products.
3. EVALUATION
We conducted several user studies to evaluate the usability of
NLA (Chai et al. 2000). In one study, seventeen test subjects
preferred the dialog-driven navigation of NLA two to one over
menu-driven navigation. Moreover, with NLA, the average
number of clicks was reduced by 63.2% and the average time
was reduced by 33.3%. Analysis of the user queries (average
length = 5.31 words long; standard deviation = 2.62; 85% of
inputs are noun phrases) revealed the brevity and relative
linguistic simplicity of user input. Hence, shallow parsing
techniques were adequate for processing user input. In general,
sophisticated dialog management appears to be more important
than the ability to handle complex natural language sentences.
The user studies also highlighted the need to combine multiple
modalities and styles of interaction.
4. REFERENCES
[1] Androutsopoulos, Ion & Ritchie, Graeme. Natural
Language Interfaces to Databases – An Introduction,
Natural Language Engineering 1.1:29-81, 1995.
[2] Carpenter, Bob & Chu-Carroll, Jeniffer. Natural Language
Call Routing: A Robust, Self-organizing Approach,
Proceedings of the 5th International Conference on Spoken
Language Processing, 1998.
[3] Chai, J., Lin, J., Zadrozny, W., Ye, Y., Budzikowska, M.,
Horvath,V.,Kambhatla,N.&Wolf,C.Comparative
Evaluation of a Natural Language Dialog Based System
and a Menu-Driven System for Information Access: A Case
Study, Proceedings of RIAO 2000, Paris, 2000.
[4] Saito, M. & Ohmura, K. A Cognitive Model for Searching
for Ill-defined Targets on the Web – The Relationship
between Search Strategies and User Satisfaction. 21st Int.
Conf. on Research and Development in Information
Retrieval, Australia, 1998.
[5] Walker, M., Fromer, J. & Narayanan, S. Learning Optimal
Dialogue Strategies: A Case Study of a Spoken Dialogue
Agent for Email, 36th Annual Meeting of the ACL,
Montreal, Canada, 1998.
[6] Zadrozny, W., Wolf, C., Kambhatla, N. & Ye, Y.
Conversation Machines for Transaction Processing,
Proceedings of AAAI / IAAI - 1998, Madison, Wisconsin,
U.S.A., 1998.
HTML
Application
Server
Client
HTTP
Server
HTML
Servlet
Web Server
Network
(HTTP)
Presentation
Manager
Dialog
Manager
Acti on
Manager
Quick
Parser
Response
Generator
Vector Space Engine
Product
Database
Business Rules
Concepts
Data Management
(Off line)
User Interface
Concept
Interpreter
Explanation
Model
Presentation
Strategies
Dialog
Strategies
Action
Strategies
input
output
Communication
Acts
Communication
Acts
Action Specs
Online
Interaction
Discourse
Analyzer
Extended
PD
Database
Query
Discourse
History
ActionResults
State
Interprete r
