People seek information online daily, for example, exploring vacation packages or browsing house listings. However, information seeking can often be very difficult and time consuming for two main reasons. First, users often cannot directly specify their information desires using conventional WIMP-based interfaces. This becomes more evident when users do not know exactly what they are looking for at the beginning of a search (e.g., finding a dream home). In such cases, users must examine multiple data aspects and explore different but related data entities to determine the targets. To express a query like U1 in Figure 1, users may only be able to approximate the query by filling multiple forms plus manually stitching together information gathered at different steps (e.g., finding all desired cities first and then using the results to find houses). Moreover, most existing systems are context insensitive. Continuing the above example, most systems cannot directly respond to follow-on queries like U2/U3. Without being able to easily revise their queries in context, users may start over again. Second, users cannot easily distill information from often scattered, one-size-fits-all presentations of retrieved data. As a result, users have to manually integrate relevant information gathered at multiple places (e.g., one display on houses and another on schools). A context-insensitive output may also prevent users from easily identifying important information or comprehending the information as a whole.

U1: 4+ br, 2+ bath, colonials under $500k in cities in the north along Hudson R1: I found 15 houses

U2: Tell me more about the cities R2: Here are the 4 cities you asked

U3: Just those in cities with more than 55000 people R3: I found 9 houses

Figure 1. A fragment of user-system conversation.