Search

Item does not contain fulltext

Since long, computer science has distinguished between information retrieval and data retrieval, where information retrieval entails the problem of ranking textual documents on their content (with the goal to identify documents relevant for satisfying a user's information need) while data retrieval involves exact match, that is, checking a data collection for presence or absence of (precisely specified) items. But, now that XML has become a standard document model that allows structure and text content to be represented in a combined way, new generations of information retrieval systems are expected to handle semi-structured documents instead of plain text, with usage scenarios that require the combination of `conventional' ranking with other query constraints; based on the structure of text documents, on the information extracted from various media (or various media representations), or through additional information induced during the query process.

The challenge of coping with the overload of multimedia data has been growing since the advent of digital cameras and broadband connections to the home. Not only the techniques for storing, annotating and search are essential for helping the user to face this challenge, a well-considered user interaction design and an intuitive user interface are equally important in helping the user to find interesting content.

In this chapter, we provide the tools and methodology for comparing the effectiveness of two or more multimedia retrieval systems in a meaningful way. Several aspects of multimedia retrieval systems can be evaluated without consulting the potential users or customers of the system, such as the query processing time (measured for instance in milliseconds per query) or the query throughput (measured for instance as the number of queries per second). In this chapter, however, we will focus on aspects of the system that influence the effectiveness of the retrieved results. In order to measure the effectiveness of search results, one must at some point consult the potential user of the system. For, what are the correct results for the query “black jaguar”? Cars, or cats? Ultimately, the user has to decide.

Digital Rights Management, or DRM for short, is a much-discussed topic nowadays. The main reason for this is that DRM technology is often mentioned in the context of protection of digital audio and video content, for example to avoid large scale copying of CDs and DVDs via peer-to-peer networks in the Internet. However, DRM technology is much more than a simple copy protection technology. It is one of the enabling technologies that open the way to secure distribution and exchange of digital content over open digital infrastructures such as the Internet.

Many documents contain, besides text, also images, tables, and so on. This chapter concentrates on the text part only. Traditionally, systems handling text documents are called information storage and retrieval systems. Before the World-Wide Web emerged, such systems were almost exclusively used by professional users, so-called indexers and searchers, e.g., for medical research, in libraries, by governmental organizations and archives. Typically, professional users act as “search intermediaries” for end users. They try to fig out in an interactive dialogue with the system and the end user what it is the end user needs, and how this information should be used in a successful search. Professionals know the collection, they know how documents in the collection are represented in the system, and they know how to use Boolean search operators to control the number of retrieved documents.

The term meta origins from the Greek word µ∈τα, meaning after. The word Metaphysics is the title of Aristotle’s book coming after his book on nature called Physics. This has given meta the modern connotation of a nature of a higher order or of a more fundamental kind [1]. Literally, metadata is “data about data‿. It can be any descriptive information about other data sources that is used to aid the organization, identification, representation, localization, interoperability, management, and use of the data [1, 14, 17].

People interact with multimedia every day: reading books, watching television, listening to music, etc. For quite some time we have faced astonishing technological developments causing an explosion of digital multimedia information. Large amounts of text, images, speech, and video are converted to digital form. Think of catalog information of libraries, information about museums with nice pictures of paintings or famous speeches that are available on DVD. Moreover, much information is produced directly in digital form: TV programs, audio-visual data from surveillance cameras, photos. Major advantages of digitized data over analog data are easy storage, processing and sharing of data. Multimedia applications influence our daily life. Consider for example the following scenarios.

This chapter will focus on the automatic extraction of information from the speech in multimedia documents. This approach is often referred to as speech indexing and it can be regarded as a subfield of audio indexing that also incorporates for example the analysis of music and sounds. If the objective of the recognition of the words spoken is to support retrieval, one commonly speaks of spoken document retrieval (SDR). If the objective is on the coupling of various media types the term media mining or even cross-media mining is used. Most attention in this chapter will go to SDR. The focus is less on searching (an index of ) a multimedia database, but on enabling multiple views on the data by cross-linking all the available multifaceted information sources in a multimedia database. In section 1.6 cross-media mining will be discussed in more detail.

In this chapter we address two approaches to extract high-level concepts from video footage and show the integrated use of both. We also describe an experiment used for validation.

This chapter is a case study showing how important events (highlights) can be automatically detected in video recordings of Formula 1 car racing. Numerous approaches presented in literature have shown that it is becoming possible to extract interesting events from video. However, the majority of the approaches uses individual visual or audio cues. According to the current understanding of human perception it is expected that using evidence obtained from different modalities should result in a more robust and accurate perception of video. On the other hand, fusion of multimodal evidence is quite challenging, since it has to deal with indications which may contradict each other. In this chapter we deal with three topics, one being fusion of evidence from different modalities.

In this paper we present a systematic analysis of document retrieval using unstructured and structured queries within the score region algebra (SRA) structured retrieval framework. The behavior of di®erent retrieval models, namely Boolean, tf.idf, GPX, language models, and Okapi, is tested using the transparent SRA framework in our three-level structured retrieval system called TIJAH. The retrieval models are implemented along four elementary retrieval aspects: element and term selection, element score computation, score combination, and score propagation. The analysis is performed on a numerous experiments evaluated on TREC and CLEF collections, using manually generated unstructured and structured queries. Unstructured queries range from the short title queries to long title + description + narrative queries. For generating structured queries we exploit the knowledge of the document structure and the content used to semantically describe or classify documents. We show that such structured information can be utilized in retrieval engines to give more precise answers to user queries then when using unstructured queries.

In many applications, uncertainty and ignorance go hand in hand. Therefore, to deliver database support for effective decision making, an integrated view of uncertainty and ignorance should be taken. So far, most of the efforts attempted to capture uncertainty and ignorance with probability theory. In this paper, we discuss the weakness to capture ignorance with probability theory, and propose an approach inspired by the Dempster-Shafer theory to capture uncertainty and ignorance. Then, we present a rule to combine dependent data that are represented in different relations. Such a rule is required to perform joins in a consistent way. We illustrate that our rule is able to solve the so-called problem of information loss, which was considered as an open problem so far.

In this paper we present the extension of our prototype three-level database system (TIJAH) developed for structured information retrieval. The extension is aimed at modeling vague search on XML elements. All three levels (conceptual, logical, and physical) of the TIJAH system are enhanced to support vague search concepts. The vague search is implemented as vague selection of XML elements using XML element name expansion lists and rewriting techniques. We test the performance of retrieval models using automatically generated expansion lists and compared them with models that use manual ones. The goal is to find the best approach for structured information retrieval with vague structural constraints on element names expressed in the query.

A unified database framework that will enable better comprehension of ranked XML retrieval is still a challenge in the XML database field. We propose a logical algebra, named score region algebra, that enables transparent specification of information retrieval (IR) models for XML databases. The transparency is achieved by a possibility to instantiate various retrieval models, using abstract score functions within algebra operators, while logical query plan and operator definitions remain unchanged. Our algebra operators model three important aspects of XML retrieval: element relevance score computation, element score propagation, and element score combination. To illustrate the usefulness of our algebra we instantiate four different, well known IR scoring models, and combine them with different score propagation and combination functions. We implemented the algebra operators in a prototype system on top of a low-level database kernel. The evaluation of the system is performed on a collection of IEEE articles in XML format provided by INEX. We argue that state of the art XML IR models can be transparently implemented using our score region algebra framework on top of any low-level physical database engine or existing RDBMS, allowing a more systematic investigation of retrieval model behavior.

In this paper we address the problem of immediate translation of eXtensible Mark-up Language (XML) information retrieval (IR) queries to relational database expressions and stress the benefits of using an intermediate XML-specific algebra over relational algebra. We show how adding an XML-specific algebra at the logical level of a DBMS enables a level of abstraction from both query languages for information retrieval in XML and the underlying physical storage and manipulation. We picked a region algebra as a basis for defining the structure aware (SA) view on XML in which we can distinguish among different XML entities, such as element nodes, text nodes, words, and determine their containment relation. Region algebras are already well established in semi-structured document processing as shown in an extensive overview of region algebra approaches in this paper. Furthermore, we propose a variant of region algebra that can support ranking operators in an elegant way while staying algebraic. As relevance scores are computed for regions in our region algebra we named it score region algebra (SRA). The benefits of introducing score region algebra are explained on a set of query examples. Besides abstracting from the query language used and the physical implementation, SRA enables a certain degree of abstraction from the retrieval model used and the opportunity to use the query optimization at the logical level of a database. Various retrieval models can be instantiated at the physical level based on the abstract specification of SRA operators. We also discuss numerous region algebra operator properties that provide a firm ground for query rewriting and optimization at the SA level, which is an important premise for the existence of such a logical view on XML.

It has been widely recognized that in many real-life database applications there is growing demand to model uncertainty and ignorance. However the relational model does not provide this possibility. Through the years a number of efforts has been devoted to the capture of uncertainty and ignorance in databases. Most of these efforts attempted to capture uncertainty using the classic probability theory. As a consequence, the limitations of probability theory are inherited by these approaches, such as the problem of information loss. In this paper, we extend the relational model with uncertainty and ignorance without these limitations posed by the other approaches. Our approach is based on the so-called theory of belief functions, which may be considered as a generalization of probability theory. Belief functions have an attractive mathematical underpinning and many intuitively appealing properties.

Advanced non-traditional application domains such as geographic information systems and digital library systems demand advanced data management support. In an effort to cope with this demand, we present the concept of a novel multi-model DBMS architecture which provides evaluation of queries on complexly structured data without sacrificing efficiency. A vital role in this architecture is played by the Moa language featuring a nested relational data model based on XNF2, in which we placed renewed interest. Furthermore, extensibility in Moa avoids optimization obstacles due to black-box treatment of ADTs. The combination of a mapping of queries on complexly structured data to an efficient physical algebra expression via a nested relational algebra, extensibility open to optimization, and the consequently better integration of domain-specific algorithms, makes that the Moa system can efficiently and effectively handle complex queries from non-traditional application domains.

The CIRQUID project plans to design and build a DBMS that seemlessly integrates relevance-oriented querying of semi-structured data (XML) with traditional querying of this data. The project is funded by the Netherlands Organisation of Scientific Research (NWO project number 612.061.210).

Advanced non-traditional application domains such as geographic information systems and digital library systems demand advanced data management support. In an effort to cope with this demand, we present a novel multi-model DBMS architecture which provides efficient evaluation of queries on complexly structured data. A vital role in this architecture is played by the Moa language featuring a nested relational data model based on XNF2, in which we placed renewed interest. Furthermore, the architecture allows extensibility on all of its levels providing the means to better integrate domain-specific algorithms into the system. In addition to this, the extensibility of the Moa language is designed in a way that optimization obstacles due to blackbox treatment of ADTs is avoided. This combination of well-integrated domainspecific algorithms, extensibility open to optimization, and a mapping of queries on complexly structured data to an efficient physical algebra expression via a nested relational algebra, makes that the Moa system can efficiently handle complex queries from non-traditional application domains.

There is a growing need for systems that can process queries, combining both structured data and text. One way to provide such functionality is to integrate information retrieval (IR) techniques in a database management system (DBMS). However, both IR and database research have been separate research fields for decades, resulting in different - even conflicting - approaches to data management. Each DBMS has a component called a "query optimizer", which plays a crucial role in the efficiency and flexibility of the system. So, for successful integration the IR techniques and data structures, as well as the DBMS query optimizer, should be adapted to enable mutual cooperation. The author concentrates on top-N queries - a common class of IR queries. An IR top-N query asks for the N best documents given a set of keywords. The author proposes processing the data in batches as a compromise between IR and DBMS query processing. Experiments with this technique show that porting IR optimization techniques is (still) not a promising option due to the additional administrative overhead. Two new mathematical models are introduced to eliminate this overhead: a model that predicts selectivity, which is a crucial factor in the execution costs, and a model that predicts the quality of the top-N.

Recent advances in computing, communications and data storage have led to an increasing number of large digital libraries, which are nowadays publicly available on the Internet. However, to find required information in that enormous mass of data becomes very difficult. In this demo we present a digital library search engine that combines efforts of the AMIS and DMW research projects, each covering significant parts of this problem. The most important contributions of our work are the following: 1. We demonstrate a flexible solution for extraction and querying of meta-data from multimedia documents in general 2. Scalability and efficiency support are illustrated for full text indexing and retrieval. 3. We show how for a more limited domain, like an Intranet, conceptual modeling can offer additional and more powerful query facilities. 4. In the limited domain case, we demonstrate how domain knowledge can be used to interpret low-level features into semantic content. In this short description, we will focus on the first and the fourth item. For the other two items, as well as for a more detailed description of our architecture of the search engine we refer to the demo web site: http://www.cs.utwente.nl/~dmw/ICDE2002/

Efficient, exible, and scalable integration of full text information retrieval (IR) in a DBMS is not a trivial case. This holds in particular for query optimization in such a context. To facilitate the bulk-oriented behavior of database query processing, a priori knowledge of how to limit the data efficiently prior to query evaluation is very valuable at optimization time. The usually imprecise nature of IR querying provides an extra opportunity to limit the data by a trade-off with the quality of the answer. In this paper we present a mathematically derived model to predict the quality implications of neglecting information before query execution. In particular we investigate the possibility to predict the retrieval quality for a document collection for which no training information is available, which is usually the case in practice. Instead, we construct a model that can be trained on other document collections for which the necessary quality information is available, or can be obtained quite easily. We validate our model for several document collections and present the experimental results. These results show that our model performs quite well, even for the case were we did not train it on the test collection itself.

Data abstraction and query processing techniques are usually studied in the domain of administrative applications. We present a case-study in the non-standard domain of (multimedia) information retrieval, mainly intended as a feasibility study in favor of the `database approach' to data management. Top-N queries form a natural query class when dealing with content retrieval. In the IR field, a lot of research has been done on processing top-N queries efficiently. Unfortunately, these results cannot directly be ported to the database environment, because their tuple-oriented nature would seriously limit the freedom of the query optimizer to select appropriate query plans. By horizontally fragmenting our database containing document statistics, we are able to combine some of the best of the IR and database optimization principles, providing good retrieval quality as well as database `goodies' like flexibility, scalability, efficiency, and generality. Key issues we address in this paper concern the effects of our fragmentation approach on speed and quality of the answers, opportunities for scalability, supported by experimental results.

New application domains cause todays database sizes to grow rapidly, posing great demands on technology. Data fragmentation facilitates techniques (like distribution, parallelization, and main-memory computing) meeting these demands. Also, fragmentation might help improving effcient processing of query types such as top N. Database design and query optimization require a good notion of the costs resulting from a certain fragmentation. Our mathematically derived selectivity model facilitates this. Once its two parameters have been computed based on the fragmentation, after each (though usually infrequent) update, our model can forget the data distribution, resulting in fast and quite good selectivity estimation. We show experimental verification for Zipfian distributed IR databases.