Catálogo de publicaciones - libros

The paper presents an attempt at the construction of a wide scale parser for English based on Inductive Learning and limited resources. The parser loosely preserves the scheme, enriched with powerful actions, and a compound instead of the decision table. The attempt originates directly from Hermjakob’s ideas [3], but an important goal was to analyse possible extensions to a wide scale solution. Several supporting heuristics, as well as the overview of the development process and experiments, are described in the paper.

Initial experiments in learning valence (subcategorisation) frames of Polish verbs from a morphosyntactically annotated corpus are reported here. The learning algorithm consists of a linguistic module, responsible for very simple shallow parsing of the input text (nominal and prepositional phrase recognition) and for the identification of valence frame cues (hypotheses), and a statistical module which implements three well-known inferential statistics (likelihood ratio, test, binomial miscue probability test). The results of the three statistics are evaluated and compared with a baseline approach of selecting frames on the basis of the relative frequencies of frame/verb co-occurrences. The results, while clearly reflecting the many deficiencies of the linguistic analysis and the inadequacy of the statistical measures employed here for a free word order language rich in ellipsis and morphosyntactic syncretisms, are nevertheless promising.

We propose a new method to deal with missing values in the gene expression data. It is applied to improve the quality of clustering genes with respect to their functionality. Calculations are run against real-life data, within the framework of self-organizing maps. The applied gene distances correspond to the rank-based Spearman correlation and entropy-based information measure.

In this paper experiments done towards improving the performance of systems retrieving musical rhythm are described. Authors briefly review machine learning models used to estimate tendency of sounds to be located in accented positions. This is done on the basis of their physical attributes such as duration, frequency and amplitude. For this purpose Data Mining association rule model and neural networks with discrete output - LVQ networks are used. By means of evaluation method introduced by the authors it is possible to compare the results returned by both models. This work aims at retrieving multi-level rhythmic structure of a musical piece on the basis of its melody, which may result in systems retrieval systems for automatic music identification.

In this paper, we use the concept of rough sets to define equivalence classes encoding input data, and to eliminate redundant or insignificant attributes in data sets which leads to reduction of the complexity of designed systems. In order to deal with ill-defined or real experimental data, we represent input object as fuzzy variables by fuzzy membership function. Furthermore we incorporate the significance factor of the input feature, corresponding to output pattern classification, in order to constitute a fuzzy inference which enhances classification considered as a nonlinear mapping. A new kind of rough fuzzy neural classifier and a learning algorithm with LSE are proposed in this paper. The neuro-fuzzy classifier proposed here can realize a nonlinear mapping from the input feature vector space (that may have the overlapping characteristic) into the output classification vector space.

In this paper the conception of evolutionary multi-agent model for knowledge acquisition has been introduced. The basic idea of the proposed solution is to use the multi-agent paradigm in order to enable the integration and co-operation of different knowledge acquisition and representation methods. At the single-agent level the reinforcement learning process is realized, while the obtained knowledge is represented as the set of simple decision rules. One of the conditions of effective agent learning is the optimization of the set of it’s features (parameters) that are represented by the genotype’s vector. The evolutionary optimization runs at the level of population of agents.

In the paper a class of infinite information systems is described. For decision tables over each such information system there exist low upper bounds on minimal complexity of decision trees and polynomial algorithms of decision tree optimization for various complexity measures.

This paper presents a conception of fast and useful inference process in knowledge based systems. The main known weakness is long and not smart process of looking for rules during the inference process. Basic inference algorithm, which is used by the rule interpreter, tries to fit the facts to rules in knowledge base. So it takes each rule and tries to execute it. As a result we receive the set of new facts, but it often contains redundant information unexpected for user. The main goal of our works is to discover the methods of inference process controlling, which allow us to obtain only necessary decision information. The main idea of them is to create rules partitions, which can drive inference process. That is why we try to use the hierarchical clustering to agglomerate the rules.

The problem of verification of rule knowledge base covers the verification of dynamic properties, which reflect the processes occurring during inference. Process of detection of these anomalies requires modelling of dynamics of these processes. Suggested in previous papers the decision unit conception did not guarantee such a possibility. This paper gives attention to the analysis of possible use of Petri nets and incidence matrix as the way of representation of knowledge base. The paper presents the relation between Petri nets and decision units’ nets and simultaneously points at the possible use of Petri nets to develop the properties of decision units.

This paper presents some results obtained in experiments with artificial neural networks trained with different learning algorithms in case of lack of some data in training and testing sets.