Catálogo de publicaciones - libros

Action rules introduced in [12] and extended further to e-action rules [21 have been investigated in [22], [13], [20]. They assume that attributes in a database are divided into two groups: stable and flexible. In general, an action rule can be constructed from two rules extracted earlier from the same database. Furthermore, we assume that these two rules describe two different decision classes and our goal is to re-classify objects from one of these classes into the other one. Flexible attributes are essential in achieving that goal since they provide a tool for making hints to a user what changes within some values of flexible attributes are needed for a given set of objects to re-classify them into a new decision class. There are two aspects of interestingness of rules that have been studied in data mining literature, objective and subjective measures [8], [1], [14], [15], [23]. In this paper we focus on a cost of an action rule which was introduced in [22] as an objective measure. An action rule was called interesting if its cost is below and support higher than some user-defined threshold values. We assume that our attributes are hierarchical and we focus on solving the failing problem of interesting action rules discovery. Our process is cooperative and it has some similarities with cooperative answering of queries presented in [3], [5], [6].

In this paper, we propose an image classification system employing multiple modules. The proposed system hierarchically categorizes given sports images into one of the predefined sports classes, eight in this experiment. The image first categorized into one of the two classes in the global module. The corresponding local module is selected accordingly, and then used in the local classification step. By employing multiple modules, the system can specialize each local module properly for the given class feature. The simulation results show that the proposed system successfully classifies images with the correct rate of over 70%.

This paper describes a new approach for describing contents through the use of interlinguas in order to facilitate the extraction of specific pieces of information. The authors highlight the different dimensions of a document and how these dimensions define the capacities of their respective contents to be found in the scalable process of finding information. A specific interlingua, UNL, will be described. This approach is illustrated both with rich examples of the followed model and with actual applications, that includes the description of some running projects based on the interlingual representation of contents.

We use the hypergeometric distribution to extract relevant information from documents. The hypergeometric distribution gives the probability estimate of observing a given term frequency with respect to a prior. The lower the probability the higher the amount of information is carried by the term. Given a subset of documents, the information items are weighted by using the inversely related function of of the hypergeometric distribution. We here provide an exemplifying introduction to a topic-driven information extraction from a document collection based on the hypergeometric distribution.

In this paper, a novel approach for building synopses is proposed by using a service and message-oriented architecture. The summarization system initially designed for very large stored databases, by its intrinsic features, is capable of dealing with the requirements inherent to the data stream environment. Its incremental maintenance of the output summaries and its scalability allows it to be a serious challenger to existing techniques. The resulting summaries present on the one hand the incoming data in a less precise form but is still on the other hand very informative on the actual content. We expose a novel way of exploiting this semantically rich information for query answering with an approach mid-way between blunt query answering and mid-way between data mining.

In this paper, we propose an algorithm for detecting a face in a target image using sketch operators and vertical facial symmetry (VFS). The former are operators which effectively reflect perceptual characteristics of human visual system to compute sketchiness of pixels and the latter means the bilateral symmetry which a face shows about its central longitudinal axis. In the proposed algorithm, horizontal and vertical sketch images are first obtained from a target image by using a directional BDIP (block difference inverse probabilities) operator which is modified from the BDIP operator. The pair of sketch images is next transformed into a generalized symmetry magnitude (GSM) image by the generalized symmetry transform (GST). From the GSM image, face candidates are then extracted which are quadrangular regions enclosing the triangles that satisfy eyes-mouth triangle (EMT) conditions and VFS. The sketch image for each candidate is obtained by the BDIP operator and classified into a face or nonface by the Bayesian classifier. Among the face candidates classified into faces, one with the largest VFS becomes the output where the EMT gives the location of two eyes and a mouth of a target face. If the procedure detects no face, then it is executed again after illumination compensation on the target image. Experimental results for 1,000 320x240 target images of various backgrounds and circumstances show that the proposed method yields about 97% detection rate and takes a time less than 0.25 second per target image.

Evaluating the performance of classifiers is a difficult task in machine learning. Many criteria have been proposed and used in such a process. Each criterion measures some facets of classifiers. However, none is good enough for all cases. In this communication, we justify the use of discrimination measures for evaluating classifiers. The justification is mainly based on a hierarchical model for discrimination measures, which was introduced and used in the induction of decision trees.

We present an approach for relation extraction from texts aimed to enrich the semantic annotations produced by a semantic web portal. The approach exploits linguistic and empirical strategies, by means of a pipeline method involving processes such as a parser, part-of-speech tagger, named entity recognition system, pattern-based classification and word sense disambiguation models, and resources such as an ontology, knowledge base and lexical databases. With the use of knowledge intensive strategies to process the input data and corpus-based techniques to deal both with unpredicted cases and ambiguity problems, we expect to accurately discover most of the relevant relations for known and new entities, in an automated way.

This paper presents a general platform for a Basque monolingual question answering (QA) system. It focuses on the architecture of the platform, paying special attention to: 1) the integration of the development and evaluation environments, and 2) the systematic use of XML declarative files to control the execution of the modules and the communication between them. Moreover, a first pilot experiment is discussed.

This paper is about using ontologies to help the execution of imperative requests expressed in natural language. In order to achieve this goal, we developed the prototype of an Ontology-Based Application Server to the execution of Natural Language requests (NL-OBAS). The NL-OBAS provides services to allow users to describe requests in several natural languages and uses software components to execute them. One of the advantages of our approach is that natural language is first converted to an interlingua, UNL (Universal Networking Language). The interlingua allows the use of different human languages to express the requests (other systems are restricted to English). The semantics of the interlingua, enhanced by ontologies, is used to retrieve the appropriated software components to compose a dynamic service to execute the requests expressed in natural language.