
		AT&T Laboratories Cambridge

DART: Information Retrieval

Information Retrieval

Representing information

Information retrieval is the task of finding information items (documents, pictures, videos etc.) which are relevant to an information need (a query or request from a user). There are many types of information, and many ways of representing the content of information. Textual information may be represented by words, word stems (e.g. fish for fishing, fishes etc.), or even letter sequences (e.g. fish as fis + ish), and this makes it easy to characterise what the information is about. However, for visual information, there are many more ways to characterise the picture. We may use such features as colour, texture and shape. Exactly which aspects of an image are important depends on the application, and therefore no single method of describing the content of a picture will suffice for all situations. This means that we need great flexibility in a database used to store and retrieve such information.

Traditional databases enable exact-match querying of information. For example, we might wish to find all documents containing the word `integration'. The aim of information retrieval is to find documents about a topic. This gives rise to the need for inexact, or best-match queries. Each candidate is assessed to determine the extent to which it satisfies the query. Some number of these results which best match the query can be returned. Often we do not require an exhaustive list--perhaps the best ten will do.

Relevance can be very subjective, and so it is a good idea to give the user a chance to indicate what is relevant and what is not. Relevance feedback is the name given to this process, and is a technique which has been around for many years in the context of text retrieval. Similar ideas can be used in retrieval of other media types. After an initial set of results is returned to the user, he or she can mark good matches as relevant, and ask the system to try again.

The DART OODB

We have developed an object-oriented multimedia information retrieval database which exploits and extends the ODMG Object Database Standard. Its powerful data model makes it easy to describe application-specific content information. In addition to the attributes and relationships which are normally used to describe the state of a database object, our OODB allows the database designer to describe how the content of a document is written.

Prototyped as part of a PhD project at the University of Cambridge Computer Laboratory under the name Cobra, our OODB is being further developed: we are extending and optimising its database functionality, while exploring its suitability for complex content representations. It provides best match querying and has facilities to support relevance feedback.

The OODB acts as a backbone to the DART project. It brings together aspects of information processing, such as image analysis, video parsing and speech recognition under a single framework. The object-oriented approach makes for rapid construction of applications, and the data model naturally supports complex document objects, media types and relationships between documents. We aim to provide a full range of database facilities, including transactions, concurrency control, recovery from failure and declarative query formulation.