PaperlayerView exact revision on arXiv

Source-linked AI summary

A Personalized System for Conversational Recommendations

M. H. Goker, P. Langley, C. A. Thompson

arXiv:1107.0029v1cs.IRcs.AI

TL;DR

Large information spaces make item selection difficult, while recommendation systems and dialogue systems each have usability limitations. The Adaptive Place Advisor combines them by learning long-term preferences unobtrusively and using a personalized user model to guide conversational search. The experiment found improved conversation efficiency over time, with differences in interactions and time attributed to updating user models rather than user learning.

  • Problem

    Recommendation and dialogue systems have limited support for personalized, efficient item-selection conversations, especially across multiple conversations.

  • Method

    The Adaptive Place Advisor learns long-term preferences unobtrusively during recommendation dialogues and uses a probabilistic user model to guide item search and conversational questions.

  • Results

    The experiment found significantly different trends between modeling and control groups for both interactions and conversation time, indicating improved efficiency as the system updated user models.

  • Takeaways & Limitations

    Personalized conversational recommendation can make interactions more efficient over time while collecting preferences as part of ordinary item search.

  • Takeaways & Limitations

    Effectiveness results were ambiguous, with no significant modeling-control differences, and the study did not control for the difficulty of finding particular items.

Abstract

from arXiv · show

Searching for and making decisions about information is becoming increasingly difficult as the amount of information and number of choices increases. Recommendation systems help users find items of interest of a particular type, such as movies or restaurants, but are still somewhat awkward to use. Our solution is to take advantage of the complementary strengths of personalized recommendation systems and dialogue systems, creating personalized aides. We present a system -- the Adaptive Place Advisor -- that treats item selection as an interactive, conversational process, with the program inquiring about item attributes and the user responding. Individual, long-term user preferences are unobtrusively obtained in the course of normal recommendation dialogues and used to direct future conversations with the same user. We present a novel user model that influences both item search and the questions asked during a conversation. We demonstrate the effectiveness of our system in significantly reducing the time and number of interactions required to find a satisfactory item, as compared to a control group of users interacting with a non-adaptive version of the system.

1. Introduction and Motivation

The paper combines personalized recommendation with conversational interaction to help users select restaurants while learning long-term preferences during ordinary dialogues. The Adaptive Place Advisor uses these models to make future conversations more efficient and reports fewer interactions and less time to find satisfactory items.

  • The Adaptive Place Advisor helps users select restaurants matching their preferences through personalized conversational recommendation.
  • The paper reports significantly fewer system-user interactions and less conversation time needed to find a satisfactory item.
  • The system combines dialogue systems with personalized recommendation to address repetitive interactions and burdensome explicit preference collection.
  • Long-term user preferences are acquired unobtrusively during normal recommendation conversations and stored for future interactions with the same user.
  • The user model represents preferences for both items and particular values of item characteristics, influencing acquisition and use during search as well as item presentation.

2. Personalized Conversational Recommendation Systems

The paper frames conversational recommendation as a personalized alternative to forms and ranked lists, using dialogue to progressively narrow choices. Its approach maintains user models across conversations and unobtrusively infers preferences to guide both search and interaction.

  • The paper targets adaptive recommendation through conversation and personalization across multiple interactions.
  • The system assumes a large item database and enough attributes to describe items; simpler techniques may suffice for smaller or easier-to-describe domains.
  • The Adaptive Place Advisor infers user preferences unobtrusively from normal online behavior rather than soliciting them directly through forms.
  • 2.2 Conversational Recommendation: Conversational recommendation uses sequences of questions to eliminate items from consideration instead of presenting a ranked list.
  • 2.2 Conversational Recommendation: The interaction can help users clarify their desires while letting them control item and attribute choices during each conversation.
  • Personalized conversational search aims to become more efficient and effective over time as the advisor uses information from previous conversations.

3. The Adaptive Place Advisor

The Adaptive Place Advisor combines conversational search with a personalized user model to retrieve, rank, and present items while adapting future questions and recommendations. Its dialogue uses constraints, similarity, and learned preferences to narrow searches and update personalization over time.

  • System architecture: The architecture distributes personalized interaction across a Dialogue Manager, Retrieval Engine, and User Modeling System, with case-based retrieval and adaptive similarity weights.The system directs conversation with frame-based methods while retrieving and ranking items through case-based reasoning.
  • Personalized search: The system initializes an expanded query from long-term and short-term preferences, then refines it as users confirm or disconfirm assumptions.The query combines implicit constraints from the user model with explicit conversational requirements.
  • Personalized search: The Retrieval Engine retrieves items matching explicit constraints, computes their similarity to user preferences, and selects attributes to constrain or relax.Its decisions depend on the number of highly similar results.
  • User model: The user model represents preferences about attributes, values, and items, including attribute weights, value distributions, and accepted-to-presented item ratios.Combination and diversity preferences are not currently modeled; value preferences can begin with a uniform distribution, while item ratios are initialized at 90%.
  • Conversational search: When many items match, the system asks about the highest-ranked unconstrained attribute; when none match, it suggests relaxing a ranked fixed constraint.The Number-of-Items dialogue state variable determines the system move during interactive constraint satisfaction.
  • User model: The system learns preferences unobtrusively from recommendation interactions, modifying feature and value weights and updating item-preference counts without unnecessary feedback requests.Updates occur after specific Accept and Reject speech acts associated with Suggest-Relax and Recommend-Item moves.

4. System Evaluation

The Adaptive Place Advisor was evaluated with Bay Area users to test whether persistent user modeling improves conversational recommendation efficiency and effectiveness over time. Modeling users showed significantly greater reductions in interactions and conversation time than controls, while effectiveness and satisfaction differences were not significant.

  • Experimental Design: The experiment compared 13 modeling-group subjects using updated user models with 11 control subjects using a non-updating system.Users interacted with restaurant recommendations in the San Francisco Bay Area across successive conversations.
  • Measures: The evaluation measured interactions needed to find an accepted restaurant, conversation time, system rejections, rejection rate, hit rate, and questionnaire responses.An interaction was defined as a prompt-to-recognition cycle, while conversation time ran from transaction start until completion.
  • Efficiency Results: 8.7 to 5.5 interactions were required by the modeling group, while the control group increased from 7.6 to 10.3; the regression slopes differed significantly (p = 0.017).The modeling trend was smaller than the control trend.
  • Efficiency Results: 181 to 96 seconds was the modeling group’s change in conversation time, compared with 132 to 152 seconds for controls; regression slopes differed significantly (p = 0.011).The reported times include system initialization, which could comprise up to 10% of dialogue time.
  • Recognition Results: 28 percent of interactions were rejected by the speech recognizer, while misrecognition occurred in seven percent; expanding supported utterances might reduce rejection but increase misrecognition.Rejections slowed conversations but did not introduce errors.
  • Effectiveness and Satisfaction: Effectiveness and satisfaction differences were not significant: rejection-rate slopes had p = 0.515, hit-rate slopes had p = 0.1354, and questionnaire checks showed no group differences.The authors describe effectiveness results as ambiguous and suggest a larger study may be needed.
  • Discussion: The authors conclude that user-model updates, rather than subjects learning the interface, produced the efficiency improvement over time.This conclusion was based on significant differences between modeling and control groups for interactions and conversation time despite large error bars and a small sample.

5. Related Research

The paper situates the Adaptive Place Advisor among personalized recommendation, conversational, adaptive dialogue, and mixed-initiative systems. Its distinguishing focus is persistent personalization that influences item search and conversation, rather than only adapting within a single interaction.

  • Research areas: Related research spans personalized recommendation systems, conversational interfaces, and adaptive dialogue systems.The paper restricts its discussion to the most strongly related work.
  • Conversational recommendation: Unlike many recommendation systems that use keywords and ranked lists, the Adaptive Place Advisor progressively narrows item choices through conversation.Its constraint-based interaction treats item search as an interactive process rather than a one-shot query.
  • Comparison with prior systems: Earlier conversational recommendation and travel systems commonly used fixed question sequences, candidate critique, or profiles that were not carried into future conversations.The cited systems differ in how they acquire preferences, represent them, and use them during search.
  • Comparison with prior systems: The Adaptive Place Advisor combines task and discourse search while using a fine-grained, personalized user model during item search.This contrasts with approaches that omit personalization, separate task and discourse spaces, or use preferences only after search.
  • Adaptive dialogue: Other adaptive dialogue systems optimize initiative, confirmation, or task completion across users rather than maintaining an individual model for future conversations.This paper instead emphasizes personalization for the same user across recommendation dialogues.

6. Directions for Future Work

Future work targets more flexible search, richer dialogue, expanded preference modeling, and larger evaluations. The authors especially plan to address uncertainty about the observed group differences and their generality across tasks and users.

  • Search mechanism: Future search work will replace or augment the current similarity threshold with learned ranking and stronger preference-based similarity methods.Planned alternatives include detecting steep similarity drops and combining preference-based and domain-specific metrics.
  • Search mechanism: The system will explore conditional entropy for selecting attributes to constrain and additional dialogue acts for confirmation and clarification.The conditional-entropy ranking had been implemented but not yet evaluated.
  • User model: Future user models will add combination and diversity preferences to capture more complex behavior and repeated-item or repeated-value intervals.The proposed diversity model uses elapsed time since a prior selection or rejection and a user-desired interval.
  • User model: The authors also plan faster model acquisition through additional learning situations and collaborative models that initialize individual profiles.These directions are intended to complement the current individual user-modeling approach.
  • Evaluation: A larger study is needed because the current evaluation did not control for task difficulty, which can change how quickly matching items appear.The authors implemented a movie-recommendation version to broaden the user base and assess satisfaction more easily.

7. Conclusions

The paper presents the Adaptive Place Advisor as an adaptive conversational assistant that learns long-term user models unobtrusively during recommendation dialogues. Its experiments show reduced interaction counts and conversation time versus a control group, while broader generality and model improvements remain open questions.

  • Contributions: The Adaptive Place Advisor acquires an individual, long-term user model unobtrusively during recommendation conversations.The work extends prior adaptive recommendation systems that were not conversational and dialogue systems that were not user adaptive.
  • Contributions: The system addresses adaptive recommendation in conversational interaction and personalization in dialogue systems through a novel approach to user-model acquisition, use, and representation.Its model includes information beyond complete item preferences to support personalization during conversations.
  • Results: The evaluation demonstrated reductions in the number of interactions and conversation time for users of the adaptive system compared with a control group.The conclusion describes these findings as showing the promise of the technique.
  • Open questions: Open questions include improving the user, conversational, and search models and testing the approach with books, movies, and other assistants.These extensions are intended to test the generality of adaptive recommendation beyond destination selection.

Appendix A. Questionnaire

The questionnaire asks users to assess their interaction with the system, the ease and usefulness of restaurant selection, the quality of delivered restaurants, conversational naturalness, and distraction.

  • Interaction and search: Users rated their overall interaction with the system and how easy it was to find a restaurant they liked.These questions address both interaction experience and search difficulty.
  • Recommendation assessment: Users reported whether the system delivered restaurants they liked and whether they considered the Adaptive Place Advisor useful.These items assess perceived recommendation quality and practical usefulness.
  • Conversational experience: Users rated the interaction between standard human-computer interaction and a person-to-person telephone conversation.The scale captures perceived conversational naturalness.
  • Conversational experience: Users judged whether the system conversation was more distracting than a similar conversation with a real person.This question evaluates a potential interaction burden.
Loading 1107.0029v1…