Utilizing interaction to improve information retrieval

点击翻译按钮获取中文摘要

Information retrieval (IR) has become an important part of life with people regularly using IR systems to search the web, their email, and other collections of documents. Even with the success of commercial web search companies, people still fail to find the information that they need to answer their questions. We propose to better utilize the power of user interaction to improve the quality of IR systems. It is well-known that user interaction has the potential to significantly improve retrieval quality, but many user interaction mechanisms fail to be adopted by either commercial search providers or by the users of search systems. To create interaction mechanisms that are both powerful and adopted, we propose to create a collection of simple mechanisms that can be easily added to existing search systems. These simple interaction mechanisms, when taken as a whole, should speed the rate at which users can find answers to their questions. The proposed research program will build on our past work where we have shown that a widely adopted, simple interaction mechanism, "find-similar", has the potential to boost retrieval quality by over 20% compared to a state-of-the-art non-interactive retrieval system. We first plan to validate via user studies our earlier simulation based results for the find-similar interaction mechanism. Next we intend to develop new interaction mechanisms that will speed the rate that users can answer complex questions with an IR system. Finally, we will create new interactive IR evaluation techniques by further merging automated usability testing methods (e.g. user simulations) with traditional IR evaluation methods. The research program will follow a cycle of formative design of interaction mechanisms followed by summative user testing. Each part of the cycle will feedback into the other. Applying this process to the problem of complex question answering, we expect to achieve a 50% or greater improvement over non-interactive retrieval.