Retrieval Scenarios

Abstract

In the previous chapters we tried to keep our definitions very abstract &$x2014; talking about conceptual “representation” spaces, “transform” functions and unspecified “parameter” vectors. Our motivation was to keep the model as general as possible, so it would be applicable to a wide range of retrieval problems. Now is the time to bring our discussion down to earth and provide specific definitions for a number of popular retrieval scenarios. We will discuss the following retrieval scenarios:

• 1. Ad-hoc retrieval: we have a collection of English documents, and a short English query. The goal is to retrieve documents relevant to the query.

• 2. Relevance feedback: in addition to the query, the user provides us with a few examples of relevant documents. The goal is to retrieve more relevant documents.

• 3. Cross-language retrieval: we have a collection of Chinese documents and an English query. The goal is to find Chinese relevant documents.

• 4. Handwriting retrieval: we have a set of historical manuscripts, represented as bitmap images. The goal is to search the collection using text queries.

• 5. Image retrieval: we have a collection of un-labeled photographs. The goal is to identify photographs relevant to a given text query (e.g., find “tiger in the grass”).

• 6. Video retrieval: we have a collection of un-annotated video footage. The goal is to find video shots containing objects of interest (e.g., “forest fire”).

• 7. Structured search with missing data: we have a database with missing field values in many records. The goal is to satisfy structured queries in the face of incomplete data.

• 8. Topic detection and tracking: we have a live stream of news reports. The goal is to organize the reports according to the events discussed in them.