Conversational Recommendation

The goal of recommendation systems is to help users find items of interest, and thus they must eliminate some choices from consideration while keeping others. To achieve this, they must either know or determine the interests of the current user, then present items that satisfy those interests. One of the most widely used approaches for information acquisition in recommendation systems is the use of a simple form (usually a single query box) that the user fills in. Coupled with this, a common presentation format is the ranked list approach. In this scheme, the system presents a usually long list of results, ordered by their predicted relevance to the user. The combined technique of information acquisition through a form and presentation in a list originated in information retrieval, where it was used to help find documents or reports during library research. The basic method, which we refer to as the ranked list approach, is more general, however, and it underlies the simpler search engines for the World Wide Web, which are used regularly for many different selection tasks. In contrast, the ADAPTIVE PLACE ADVISOR supports conversations as its mode of interaction. In this approach, one conversational participant, which we call the inquirer, asks a second participant, which we call the advisor, for assistance in selecting an item. For example, if the inquirer is deciding on a restaurant at which to have dinner, the conversation may proceed as follows:
1. Inquirer: Where do you think I should eat tonight?
2. Advisor: What type of food would you like?
3. Inquirer: What types are there?
4. Advisor: You can say things like Chinese, Indian, and Mediterranean.
5. Inquirer: Oh, maybe a cheap Indian place.
6. Advisor: What kind of parking would you like?
7. Inquirer: I don't care, as long as it's in Palo Alto.
8. Advisor: I'm sorry, I don't know of any restaurants like that, would you like to search for any price?
9. Inquirer: No, I think I'd like Chinese instead.
10. Advisor: How does Mandarin Gourmet on 420 Ramona sound?¹
11. Inquirer: No, what else do you have?
12. Advisor: How about Jing-Jing Szechwan Hunan Gourmet on 443 Emerson Street?
13. Inquirer: Sure, that sounds fine.
We will refer to systems that mimic this approach to recommendation as conversational recommendation systems. The interaction supported by conversational recommendation systems is quite different from that found in the ranked list approach. The most important distinction is that the interaction takes the form of a sequence of questions, most designed to eliminate some items from consideration. Answering these questions plays a similar role to giving keywords in a form, but the aim is to remove alternatives rather than to rank results in a list. This conversational process can also help the inquirer better understand his own desires, since thinking about possible questions and answers may clarify goals in ways that keyword entry does not encourage. In addition, a conversation can support user variability by letting the user control item and attribute choices in each conversation. Finally, the inquirer never hears about a complete item until at most a few choices remain. This avoids presenting a long list of items and lets the user narrow down the choices in an iterative, manageable fashion. Such dialogues seem better for recommendations that must be delivered by speech rather than visually, for example those engaged in while the inquirer is driving. They also seem ideal, independent of modality, for tasks like destination selection or help-desk support [36,4], in which the user needs to converge on at most a few items. On the other hand, keyword entry and ranked list methods seem more appropriate in situations where the user prefers to provide all requirements at once, in situations where information can be presented visually, and in situations where the user may want to examine many options. By eliminating options, conversational recommendation systems ultimately direct their users to a suitable solution. However, such a conversation can become tiring and the quality of the first result returned may not be acceptable for each user. Just as interactions with a friend who knows your concerns can be more directed and produce better results than those with a stranger, dialogues with a conversational advisor should become more efficient and effective over time. Our goals for user modeling include improvement of the subjective quality and effectiveness of both the results (found items) and the conversation that leads to these results. For example, after several conversations with the inquirer above, a new interaction may proceed as follows, where the question about parking has been eliminated and the item presentation order has changed:
1. Inquirer: Where do you think I should eat tonight?
2. Advisor: What type of food would you like?
3. Inquirer: Oh, maybe a Chinese place.²
4. Advisor: What city do you prefer?
5. Inquirer: Do you have something in Palo Alto?
6. Advisor: How does Jing-Jing Szechuan Gourmet on 443 Emerson sound?
7. Inquirer: Sure, that sounds fine.
We turn next to our design choices concerning the management of such conversations.