Modularized User Modeling in Conversational Recommender Systems Pontus warnestal Department of Computer Science, Linkoping University, Sweden o@ida.liu. se 1 Introduction ly research interest lies in investigating user-adaptive interaction in a conver sational setting for recommender systems, with particular focus on modularized user model components and the use of a dialogue partner(DP) in such systems Research on recommender systems have focused mostly on the back-end al- gorithms they employ, whereas front-end interaction issues have not been as horoughly studied [2, and most recommender systems rely on the graphical point-and-click metaphor. The area of conversational recommender systems ex- plores another interaction metaphor; that of natural language(NL) interaction in a dialogue (e. g. with spoken interaction [8, 5). Other recommender system approaches acknowledge the benefits of NL and / or dialogue-based interaction in their systems, but employ graphical user interfaces since their primary focus lies elsewhere(e. g. [1, 6 ). Underlying motivations for the conversational approach are that the initializing and updating of the user preference model can be carried out efficiently and enjoyably using natural language, that an on-going dialogue supports evolving user queries and continuous updating of user preference data, and that NL facilitates rich feedback in a way that is natural to end-users. A related metaphor--however often lacking proper conversational interac tion--is the virtual assistant approach, which takes the form of an animated character(e.g. the Microsoft Office assistant, and the SitePal guides)that pro- vides help with a software application or web site. Such virtual partners display a number of different interaction strategies and varying adaptive functionality User-adaptive functionality based on user models has long been proposed in order to meet individual needs and make interaction more usable and efficient 4 In order to make full use of and to better understand conversational dp interac- tion with user-adaptive recommenders, research addressing the following issues seems to be needed: (a)how to design and implement conversational virtual assis- tants,(b)how to endow such assistants with appropriate Nl dialogue capabilities to effectively render them as DPs, and(c)how to adapt DPs to individual users 2 Research problems Implementation details may vary, but essentially a generic NL dialogue system ar- chitecture can be described as consisting of the following phases(in order):(a)in- terpretation, (b)dialogue management, (c) domain reasoning, and(d)generation rdissono, P. Brna, and A. Mitrovic(Eds ) UM 2005, LNAI 3538, pp. 527-529, 2005
Modularized User Modeling in Conversational Recommender Systems Pontus W¨arnest˚al Department of Computer Science, Link¨oping University, Sweden ponjo@ida.liu.se 1 Introduction My research interest lies in investigating user-adaptive interaction in a conversational setting for recommender systems, with particular focus on modularized user model components and the use of a dialogue partner (dp) in such systems. Research on recommender systems have focused mostly on the back-end algorithms they employ, whereas front-end interaction issues have not been as thoroughly studied [2], and most recommender systems rely on the graphical point-and-click metaphor. The area of conversational recommender systems explores another interaction metaphor; that of natural language (nl) interaction in a dialogue (e.g. with spoken interaction [8, 5]). Other recommender system approaches acknowledge the benefits of nl and/or dialogue-based interaction in their systems, but employ graphical user interfaces since their primary focus lies elsewhere (e.g. [1, 6]). Underlying motivations for the conversational approach are that the initializing and updating of the user preference model can be carried out efficiently and enjoyably using natural language, that an on-going dialogue supports evolving user queries and continuous updating of user preference data, and that nl facilitates rich feedback in a way that is natural to end-users. A related metaphor—however often lacking proper conversational interac tion— is the virtual assistant approach, which takes the form of an animated character (e.g. the Microsoft Office assistant, and the SitePal guides) that provides help with a software application or web site. Such virtual partners display a number of different interaction strategies and varying adaptive functionality. User-adaptive functionality based on user models has long been proposed in order to meet individual needs and make interaction more usable and efficient [4]. In order to make full use of and to better understand conversational dp interaction with user-adaptive recommenders, research addressing the following issues seems to be needed: (a) how to design and implement conversational virtual assistants, (b) how to endow such assistants with appropriate nl dialogue capabilities to effectively render them as dps, and (c) how to adapt dps to individual users. 2 Research Problems Implementation details may vary, but essentially a generic nl dialogue system architecture can be described as consisting of the following phases (in order): (a) interpretation, (b) dialogue management, (c) domain reasoning, and (d) generation. L. Ardissono, P. Brna, and A. Mitrovic (Eds.): UM 2005, LNAI 3538, pp. 527–529, 2005. c Springer-Verlag Berlin Heidelberg 2005
528 P. Warnestal Different forms of adaptivity are prominent in these phases. In each phase there is thus a potential for modeling user attributes and or preferences. For example, individual users have their own way of expressing queries and their vocabulary preferences. This is clearly an adaptive functionality of the interpre. ation phase, and is facilitated by a personalized lexicon and grammar. Moving to the dialogue management phase, it is considered necessary to include adap- tation if the dialogue is to be viewed as cooperative 7. One important example of adaptation of the dialogue is the DP's strategy and initiative, such as whether the DP should be more pro-active or more reactive. When considering domain reasoning-and connection to the system back-end--there are yet other types of user data that require modeling(e.g. item ratings for collaborative filtering systems, or user needs and interests for knowledge-based recommender engines etc. ) Finally, Zukerman and Litman [10 conclude that user models are required in order to enable systems to generate appropriate and relevant responses in logue systems. The generation phase is concerned with generating responses that fit a specific user(e.g. content planning, surface generation, and modalit considerations and feedback) User-adaptive system performance depends on how the user model is (1) initialized,(2)updated, and ( 3) put to use in order to achieve adaptive func- tionality. These three aspects need to be addressed for each phase's user model. The aspects and phases define a two-dimensional problem space of modularized user modeling components that frames this research My work is aimed at investigating what kind of modeling is carried out at the different points of this problem space; and finding out how different phenomena and problems are handled in each phase Contributions for developers include a theoretical framework and corresponding tool linking what needs to be modeled and how, with the desired adaptive functionality of the system 3 Previous and Current Work So far, I have developed a conversational movie recommender system 5, which implements an empirically based recommendation dialogue strategy described in 9. The system supports initialization and continuous updating of a user' movie preferences, and gives personalized recommendations and explanations through nl dialogue. In terms of the architecture outlined above this work thus focuses on the domain reasoning phase, and forms the base on which I will continue to develop and investigate user modeling for the remaining phases Using an existing phase-based dialogue system architecture 3. I have started to work on a user modeling component framework that functions as pluggable modularized intercepting filters for each of the standard dialogue system phases as outlined above. The filters are configurable and will contain formalisms and mechanisms for initializing, updating, and putting the different models to use
528 P. W¨arnest˚al Different forms of adaptivity are prominent in these phases. In each phase there is thus a potential for modeling user attributes and/or preferences. For example, individual users have their own way of expressing queries and their vocabulary preferences. This is clearly an adaptive functionality of the interpretation phase, and is facilitated by a personalized lexicon and grammar. Moving to the dialogue management phase, it is considered necessary to include adaptation if the dialogue is to be viewed as cooperative [7]. One important example of adaptation of the dialogue is the dp’s strategy and initiative, such as whether the dp should be more pro-active or more reactive. When considering domain reasoning—and connection to the system back-end—there are yet other types of user data that require modeling (e.g. item ratings for collaborative filtering systems, or user needs and interests for knowledge-based recommender engines, etc.). Finally, Zukerman and Litman [10] conclude that user models are required in order to enable systems to generate appropriate and relevant responses in dialogue systems. The generation phase is concerned with generating responses that fit a specific user (e.g. content planning, surface generation, and modality considerations and feedback). User-adaptive system performance depends on how the user model is (1) initialized, (2) updated, and (3) put to use in order to achieve adaptive functionality. These three aspects need to be addressed for each phase’s user model. The aspects and phases define a two-dimensional problem space of modularized user modeling components that frames this research. My work is aimed at investigating what kind of modeling is carried out at the different points of this problem space; and finding out how different phenomena and problems are handled in each phase. Contributions for developers include a theoretical framework and corresponding tool linking what needs to be modeled, and how, with the desired adaptive functionality of the system. 3 Previous and Current Work So far, I have developed a conversational movie recommender system [5], which implements an empirically based recommendation dialogue strategy described in [9]. The system supports initialization and continuous updating of a user’s movie preferences, and gives personalized recommendations and explanations through nl dialogue. In terms of the architecture outlined above this work thus focuses on the domain reasoning phase, and forms the base on which I will continue to develop and investigate user modeling for the remaining phases. Using an existing phase-based dialogue system architecture [3], I have started to work on a user modeling component framework that functions as pluggable modularized intercepting filters for each of the standard dialogue system phases as outlined above. The filters are configurable and will contain formalisms and mechanisms for initializing, updating, and putting the different models to use
Modularized User Modeling in Conversational Recommender Systems 529 4 Contributions Concrete contributions of this work will include a theory for explaining and designing DP recommender systems with modu larized and transparent adaptivity according to the problem space described in section 2 a tool that implements the theory, and that provides developers with con- figurable and transparent user modeling components that generate desired adaptive functionality for their end-users, an an application built with the tool for end-user evaluation of the approach. Acknowledgments. This work is supervised by arne Jonsson and Lars Degerst edt, and supported by GSLT(Sweden)and Santa Anna IT Research References 1. Robin D. Burke, Kristian J. Hammond, and Benjamin C. Young. The FindMe Approach to Assisted Browsing. IEEE Expert, 12(4): 32-40, 1997 2. Giuseppe Carenini, Jocelyin Smith, and David Poole. Towards more conversa- tional and collaborative recommender systems. In Proceedings of the international Conference of Intelligent User Interfaces, pages 12-18, Miami, Florida, USA, 2003 3. Lars Degerstedt and Pontus Johansson. Evolutionary Development of Phase-Based Dialogue Systems. In Proceedings of the 8th Scandinavian Conference on Artificial Intelligence, pages 59-67, Bergen, Norway, 2003 4. Gerhard Fischer. User modeling in human-computer interaction. User Modeling nd User-Adapted Interaction, 11: 65-86, 2001 5. Pontus Johansson. Design and development of recommender dialogue systems. Licentiate Thesis 1079, Linkoping Studies in Science and Technology, Linkoping University, April 2004 6. Lorraine McGinty and Barry Smyth. Deep dialogue vs casual conversation in rec- mender systems. In F. Ricci and B Smyth, editors, Personalization in e corm- merce at the Second International Conference on Adaptive Hypermedia and Wet Based Systems(AH-02), pages 80-89, Malaga, Spain, 2002. 7. Katarina Morik. Discourse models, dialog memories, and user models. Computa- tional Linguistics, 14: 95-97, 1988 8. Cynthia Thompson, Mehmet Goker, and Pat Langley. A personalized system for conversational recommendations. Journal of Artificial Intelligence Research 21:393-428,2004. 9. Pontus Warnestal. Modeling a dialogue strategy for personalized movie recom- mendations. In Beyond Personalization 05 Workshop, pages 77-82, San Diego, CA, USA, January 2005 10. Ingrid Zukerman and Diane Litman. Natural language processing and user mod eling: Synergies and limitations. User Modeling and User-Adapted Interaction 11:129-158,2001
Modularized User Modeling in Conversational Recommender Systems 529 4 Contributions Concrete contributions of this work will include: – a theory for explaining and designing dp recommender systems with modularized and transparent adaptivity according to the problem space described in section 2, – a tool that implements the theory, and that provides developers with con- figurable and transparent user modeling components that generate desired adaptive functionality for their end-users, and – an application built with the tool for end-user evaluation of the approach. Acknowledgments. This work is supervised by Arne J¨onsson and Lars Degerstedt, and supported by GSLT (Sweden) and Santa Anna IT Research. References 1. Robin D. Burke, Kristian J. Hammond, and Benjamin C. Young. The FindMe Approach to Assisted Browsing. IEEE Expert, 12(4):32–40, 1997. 2. Giuseppe Carenini, Jocelyin Smith, and David Poole. Towards more conversational and collaborative recommender systems. In Proceedings of the International Conference of Intelligent User Interfaces, pages 12–18, Miami, Florida, USA, 2003. 3. Lars Degerstedt and Pontus Johansson. Evolutionary Development of Phase-Based Dialogue Systems. In Proceedings of the 8th Scandinavian Conference on Artificial Intelligence, pages 59–67, Bergen, Norway, 2003. 4. Gerhard Fischer. User modeling in human-computer interaction. User Modeling and User-Adapted Interaction, 11:65–86, 2001. 5. Pontus Johansson. Design and development of recommender dialogue systems. Licentiate Thesis 1079, Link¨oping Studies in Science and Technology, Link¨oping University, April 2004. 6. Lorraine McGinty and Barry Smyth. Deep dialogue vs casual conversation in recommender systems. In F. Ricci and B Smyth, editors, Personalization in eCommerce at the Second International Conference on Adaptive Hypermedia and WebBased Systems (AH-02), pages 80–89, Malaga, Spain, 2002. 7. Katarina Morik. Discourse models, dialog memories, and user models. Computational Linguistics, 14:95–97, 1988. 8. Cynthia Thompson, Mehmet G¨oker, and Pat Langley. A personalized system for conversational recommendations. Journal of Artificial Intelligence Research, 21:393–428, 2004. 9. Pontus W¨arnest˚al. Modeling a dialogue strategy for personalized movie recommendations. In Beyond Personalization’05 Workshop, pages 77–82, San Diego, CA, USA, January 2005. 10. Ingrid Zukerman and Diane Litman. Natural language processing and user modeling: Synergies and limitations. User Modeling and User-Adapted Interaction, 11:129–158, 2001