This document discusses developing conversational agents for selling situations. It proposes a formal framework for dialogue management by proactive agents in different selling scenarios like before, during, and after sales. It describes CSO LP, an artificial intelligence system for natural language conversations. CSO LP can manage dialog sessions, understand language, interact with databases, and respond to users. The document also discusses using argumentation-based approaches for negotiation and decision making in conversational agents.
WordPress Websites for Engineers: Elevate Your Brand
Conversational agent in selling situations
1. Context State-of-the-art Proposal Conclusion Biblio
Conversational agent in selling situations
GT ACA
Sameh ABDELNABY, Bruno BEAUFILS,
Maxime MORGE, Yann SECQ
Équipe SMAC, LIFL
Paris, November 2009
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 1
2. Context State-of-the-art Proposal Conclusion Biblio
Ubiquitous Virtual Seller
Objectives :
◮ Improve the transformation ratio (sales/visitors).
2% of 3suisses.fr = 500 k euros
◮ Integration of conversational agents in e-commerce websites
◮ Development, implementation, evaluation and validation of selling
solutions
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 2
3. Context State-of-the-art Proposal Conclusion Biblio
Scenarios
◮ Before-sale : facilitate information retrieval about the
products/services
◮ Sale : simplify the decision making and the purchasing when the
needs have been identified
◮ Cross-sale : promote the selling of an additional product/service to
an existing customer
◮ After-sale : assist the customers in using the product/service
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 3
4. Context State-of-the-art Proposal Conclusion Biblio
CSO LP Artificial Solutions’ Language Processor
Features :
◮ Manage sessions
◮ Handles misspellings, language dependent preprocessing
◮ Dialogue Context
◮ Select and carry out best system action according to
interaction rules in knowledge base
◮ Interact with back end (e.g. databases)
◮ Hand out answer document to requesting application/front end
◮ Write log files (for analysis)
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 4
5. Context State-of-the-art Proposal Conclusion Biblio
CSO LP Artificial Solutions’ Language Processor
Knowledge :
◮ Dialogue store = variable, value
◮ Abbreviation list
◮ Auto-correction dictionary
◮ Knowledge base ⊇ interaction rules with priority , i.e.
priority : if user input + dialogical context
then action
Research challenge :
Formal framework for dialogue management by
proactive agents in different selling situations
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 5
6. Context State-of-the-art Proposal Conclusion Biblio
Mixed-initiative dialogue systems for collaborative problem solving
◮ Existing dialogue systems :
◮ TRAINS-93 [Ferguson 07]
◮ Collagen [Rich et al. 01]
◮ Artemis Agent Technology [Sadek 05]
◮ “The LCD projector is no longer working.” means
◮ My need is a LCD projector (before-sale) ?
◮ Our joint goal is the purchasing of a LCD projector (sale) ?
◮ Your goal is to assist me (after-sale) ?
◮ Complexity and heuristics of goal/plan recognition
◮ But small is beautiful . . .
◮ ⇒ Dialectical approach
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 6
7. Context State-of-the-art Proposal Conclusion Biblio
Argumentation-based negotiation with MARGO.sf.net [Morge 09]
The goal consists of replying
with the optimal utterance
optimal ← locution(request), product(bike, price, deliveryTime, quality ),
lastlocution(none) Begin with the optimal product
optimal ← locution(concede), product(bike, price, deliveryTime, quality ),
lastlocution(reply), Reply with a proposal not yet rejected
notrejected(bike, price, deliveryTime, quality ),
notlastoffer(buyer, bike, price, deliveryTime, quality )
respond ← locution(standstill), product(bike, price, deliveryTime, quality ),
lastlocution(reply), Repeat the previous proposal
lastoffer(seller, bike, price, deliveryTime, quality )
optimal ← locution(concede), product(bike, price, deliveryTime, quality ),
lastlocution(concede), Concede with a less optimal proposal
notrejected(bike, price, deliveryTime, quality ),
notlastoffer(seller, bike, price, deliveryTime, quality )
respond ← locution(standstill), product(bike, price, deliveryTime, quality ),
lastlocution(concede), Repeat the previous proposal
lastoffer(buyer, bike, price, deliveryTime, quality )
Otherwise, the goal consists of replying
with a legal utterance
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 7
8. Context State-of-the-art Proposal Conclusion Biblio
Argumentation-based negotiation with MARGO.sf.net [Morge 09] (cont.)
◮ The arguments
◮ argument 1 : repeat the previous proposal
◮ argument 2 : concede with a “less optimal” proposal
◮ argument 3 : the previous proposal has been rejected
◮ The relations
◮ argument 1 and 2 attack one another
◮ argument 3 attacks argument 1
◮ The decision
◮ Since arguments 2 and 3 together “win”, choose to
concede ! ! !
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 8
9. Context State-of-the-art Proposal Conclusion Biblio
To take away
◮ A dialectical framework for dialogue formalization
◮ Proactive agents in different selling situations
◮ MARGO.sf.net, argumentation over motivation for
conceding
◮ Agent behaviours s.t. the minimal concession strategy
◮ Dialogue-game protocols for information-seeking,
enquiry, deliberation and negotiation
◮ Toward development, implementation, evaluation
and validation of prototypes
◮ http://www.lifl.fr/SMAC/projects/vvu/
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 9
10. Context State-of-the-art Proposal Conclusion Biblio
References
J. A. George Ferguson. Mixed-initiative systems for collaborative problem solving.
AI Magazine, 28(2) :23–32, 2007.
C. Rich, C. L. Sidner, and N. Lesh. COLLAGEN applying collaborative discourse theory
to human-computer interaction.
AI Magazine, 22(4) :15–25, 2001.
D. Sadek. Multi-Agent Programming, chapter Artimis Rational Dialogue Agent
Technology : An Overview, pages 217–225.
Springer-Verlag, 2005.
C. L. Hamblin. Fallacies.
Methuen, 1970.
D. Walton and E. Krabbe. Commitment in Dialogue.
SUNY Press, 1995.
M. Morge and P. Mancarella. The hedgehog and the fox. An argumentation-based
decision support system.
In Proc. of the 4th International Workshop on Argumentation in Multi-Agent Systems
(ArgMAS), pages 1–18, Honolulu, Hawai, 2007.
M. Morge and P. Mancarella. Assumption-based argumentation for the minimal
concession strategy.
In Proc. of the 6th International Workshop on Argumentation in Multi-Agent Systems
(ArgMAS), pages 1–18, Budapest, Hungary, 2009. logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 10
11. Context State-of-the-art Proposal Conclusion Biblio
CSO LP Artificial Solutions’ Language Processor
Process :
1. Input = inquiry (user id/input)
2. Dialogue management = identification of the session
3. Recognition = division in sentences, words, spelling correction, . . .
4. Interpretation = answer retrieval for single sentences of user input
5. Selection of the final answer = preparation of answer data
6. Generation of answer= replacement of template variables
7. Output = HTTP response
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 11
12. Context State-of-the-art Proposal Conclusion Biblio
Argumentation-based decision making in MARGO.sf.net [Morge 07]
The goal is the selection of
the optimal product wrt
the user’s preferences
The most important feature
is the price
cheap ← product(bike, price, deliveryTime, quality ), price < 140euros
good ← product(bike, price, deliveryTime, quality ), quality > high
fast ← product(bike, price, deliveryTime, quality ), deliveryTime < 48h
The least important feature
is the delivery time
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 12
13. Context State-of-the-art Proposal Conclusion Biblio
Argumentation-based decision making in MARGO.sf.net [Morge 07] (cont.)
◮ The arguments
◮ argument 1 : bike1 is a possible choice
◮ argument 2 : bike2 is a possible alternative choice
◮ argument 3 : bike2 is preferred because it is cheap
◮ The relations
◮ argument 1 and 2 attack one another
◮ argument 3 attacks argument 1
◮ The decision
◮ Since arguments 2 and 3 together “win”, choose
bike2 ! ! !
logo/lille1.
Maxime Morge VVU/PICOM - GT ACA - Page 13