Catálogo de publicaciones - libros

This report describes a technology of coding of the text documents used within the framework of a working breadboard model of meta-search system developed in ”Bineuro” with context-dependent processing and with classification of search results. The technology of automatic analysis and coding of text documents is based on text corpus representation in the form of an associative semantic network. Code vectors are generated as the equilibrium points of neural network with feedback and with parallel dynamics, such as Hopfild network with an asymmetrical matrix of feedback. These code vectors can be used for tasks of ranging, automatic cluster analysis of documents and in many other tasks connected to automatic post processing of search results by the search engines on the Internet. As an example of application of code vectors we will consider below a task of classification of text documents.

Many statistics-based approaches have been proposed to label semantic roles automatically. For language lack of large-scale semantically annotated corpora, e.g. Chinese, these approaches do not work well. In this paper we proposed an unsupervised syntax and semantic linking algorithm, and make full use of current Chinese language resources to equip it with detailed linking knowledge. Therefore, the semantic role labeling task can be attributed to a data-driven application of the linking algorithm. Some preliminary experiment results demonstrate the ability of the linking algorithm for the automatic semantic role labeling on current Chinese treebank.

To retrieve relevant information, indexing should be achieved using the concepts of the document that a writer intends to highlight. Moreover, the user involvement is increasingly required to extract relevant information from information sources. Therefore, in the present work we propose a fuzzy retrieval model indexed by concept identification: (1) a concept identification based indexing and (2) a novel fuzzy ranking model. The concept based indexing identifies index terms by considering the concepts of a document, and a novel fuzzy ranking model based on the user preference is presented, which is able to calculates the relevance ranking based on the user preference.

Since sentences are the basic propositional units of text, knowing their themes should help various tasks requiring the knowledge about the semantic content of text. In this paper, we examine the notion of sentence theme and propose an automatic scheme where head-driven patterns are used for theme assignment. We tested our scheme with sentences in encyclopedia articles and obtained a promising result of 98.96% in F-score for training data and 88.57% for testing data, which outperform the baseline.

In language engineering, language models are employed in order to improve system performance. These language models are usually -gram models which are estimated from large text databases using the occurrence frequencies of these -grams. An alternative to conventional frequency-based estimation of -gram probabilities consists on using neural networks to this end. In this paper, an approach to language modeling with a hybrid language model is presented as a linear combination of a connectionist -gram model, which is used to represent the global relations between certain linguistic categories, and a stochastic model of word distribution into such categories. The hybrid language model is tested on the corpus of the Wall Street journal processed in the Penn Treebank project.

This paper deals with the problem of efficient speaker adaptation in large vocabulary continuous speech recognition (LVCSR) systems. The main goal is to adapt acoustic models of speech and to increase the recognition accuracy of these systems in tasks, where only one user is expected (e.g. voice dictation) or where the speaking person can be identified automatically (e.g. broadcast news transcription). For this purpose, we propose several modifications of the well known MLLR (Maximum Likelihood Linear Regression) method and we combine them with the MAP (Maximum A Posteriori) method. The results from a series of experiments show that the error rate of our 300K-word Czech recogniser can be reduced by about 9.9 % when only 30 seconds of supervised data are used for adaptation or by about 9.6 % when unsupervised adaptation on the same data is performed.

Human listeners use lexical stress for word segmentation and disambiguation. We look into using lexical stress for speech recognition by examining a Dutch-language corpus. We propose that different spectral features are needed for different phonemes and that, besides vowels, consonants should be taken into account.

The goal of this paper is to describe the development of the sound database for the allophone-based model for English concatenative speech synthesis. The procedure of the sound unit inventory construction is described and its main results are presented. At present moment the optimized sound units inventory of the allophonic database for English concatenative speech synthesis contains 1200 elements (1000 vowel allophones and 200 consonant allophones). The smoothness of junctions between the allophones shows high quality of the segmentation made. The decrease in the number of the database components in the result of optimization does not affect the quality of the resulting synthesized speech. At the level of segments it can be evaluated as fairly high in terms of both naturalness and intelligibility.

Automatic Speech Recognition (ASR) in reverberant rooms can be improved by choosing training data from the same acoustical environment as the test data. In a real-world application this is often not possible. A solution for this problem is to use speech signals from a close-talking microphone and reverberate them artificially with multiple room impulse responses. This paper shows results on recognizers whose training data differ in size and percentage of reverberated signals in order to find the best combination for data sets with different degrees of reverberation. The average error rate on a close-talking and a distant-talking test set could thus be reduced by 29% relative.

Multilingual access to information and services is a key requirement in any pervasive or ubiquitous computing environment. In this paper we describe our efforts towards multilingual speech recognition with a focus on applications that are designed to run on embedded devices, like e.g. a commercially available PDA. We give an overview on speech recognition techniques suited for the special requirements of the expected phonetic and acoustic environments and explore the ability to create multilingual acoustic models and applications that are able to run on embedded devices in real-time.