Catálogo de publicaciones - libros

This paper presents how machine learning can be used to automatically obtain rules to divide words in Spanish into syllables. Machine learning is used in this case not only as a classifier to decide when a rule is used but to generate meaningful rules which then can be used to syllabify new words. Syllabification is an important task in speech recognition and synthesis since every syllable represents the sound in a single effort of articulation. Experiments were carried out using an Inductive Logic Programming (ILP) tool. The experiments were made on different sets of words to ascertain the importance of the number of examples in obtaining useful rules. The results show that it is possible to automatically obtain rules for syllabifying.

Question Classification (QC) is usually the first stage in a Question Answering system. This paper presents a multilingual SVM-based question classification system aiming to be language and domain independent. For this purpose, we use only surface text features. The system has been tested on the TREC QA track questions set obtaining encouraging results.

Passage Retrieval (PR) is typically used as the first step in current Question Answering (QA) systems. Most methods are based on the vector space model allowing the finding of relevant passages for general user needs, but failing on selecting pertinent passages for specific user questions. This paper describes a simple PR method specially suited for the QA task. This method considers the structure of the question, favoring the passages that contain the longer -gram structures from the question. Experimental results of this method on Spanish, French and Italian show that this approach can be useful for multilingual question answering systems.

This paper studies the problem of building text classifiers using positive and unlabeled examples. The primary challenge of this problem as compared with classical text classification problem is that no labeled negative documents are available in the training example set. We call this problem PU-Oriented text Classification. Our text classifier adopts traditional two-step approach by making use of both positive and unlabeled examples. In the first step, we improved the 1-DNF algorithm by identifying much more reliable negative documents with very low error rate. In the second step, we build a set of classifiers by iteratively applying SVM algorithm on training data set, which is augmented during iteration. Different from previous PU-oriented text classification works, we adopt the weighted vote of all classifiers generated in the iteration steps to construct the final classifier instead of choosing one of the classifiers as the final classifier. Experimental results on the Reuter data set show that our method increases the performance (F1-measure) of classifier by 1.734 percent compared with PEBL.

We present a method for creating natural language interfaces to databases (NLIDB) that allow for translating natural language queries into SQL. The method is domain independent, i.e., it avoids the tedious process of configuring the NLIDB for a given domain. We automatically generate the domain dictionary for query translation using semantic metadata of the database. Our semantic representation of a query is a graph including information from database metadata. The query is translated taking into account the parts of speech of its words (obtained with some linguistic processing). Specifically, unlike most existing NLIDBs, we take seriously auxiliary words (prepositions and conjunctions) as set theory operators, which allows for processing more complex queries. Experimental results (conducted on two Spanish databases from different domains) show that treatment of auxiliary words improves correctness of translation by 12.1%. With the developed NLIDB 82of queries were correctly translated (and thus answered). Reconfiguring the NLIDB from one domain to the other took only ten minutes.

This paper presents an innovative hybrid approach for online recognition of handwritten symbols. The approach is composed of two main techniques. Firstly, fuzzy rules are used to extract a meaningful set of features from a handwritten symbol, and secondly a recurrent neural network uses the feature set as input to recognise the symbol. The extracted feature set is a set of basic shapes capturing what is distinctive about each symbol, thus making the network’s classification task easier. We propose a new recurrent neural network architecture, associated with an efficient learning algorithm derived from the gradient descent method. We describe the network and explain the relationship between the network and the Markov chains. The approach has achieved high recognition rates using benchmark datasets from the Unipen database.

In this paper, we describe environment compensation approach based on MAP (maximum a posteriori) estimation assuming that the noise can be modeled as a single Gaussian distribution. It employs the prior information of the noise to deal with environmental variabilities. The acoustic-distorted environment model in the cepstral domain is approximated by the truncated first-order vector Taylor series(VTS) expansion and the clean speech is trained by using Self-Organizing Map (SOM) neural network with the assumption that the speech can be well represented as the multivariate diagonal Gaussian mixtures model (GMM). With the reasonable environment model approximation and effective clustering for the clean model, the noise is well refined using batch-EM algorithm under MAP criterion. Experiment with large vocabulary speaker-independent continuous speech recognition shows that this approach achieves considerable improvement on recognition performance.

In this work, we propose to use as source of speech information the Short-MelfrequencyCepstra Time Transform (SMCTT), (). The SMCTT studies the time properties at quefrency . Since the SMCTT signal, (), comes from a nonlinear transformation of the speech signal, (), it makes the STMCTT a potential signal with new properties in time, frequency, quefrency, etc. The goal of this work is to present the performance of the SMCTT signal when the SMCTT is applied to an Automatic Speech Recognition (ASR) task. Our experiment results show that important information is given by this SMCTT waveform, ().

In this paper we describe the work done to build a new voice based on diphone concatenation in the Spanish spoken in Mexico. This voice is compatible with the Text to Speech Synthesis System Festival. In the development of each module of the system the own features of Spanish were taken into account. In this work we hope to enhance the naturalness of the synthesized voice by including a prosodic model. The prosodic factors taken into consideration by the model are: phrasing, accentuation, duration and F0 contour. Duration and F0 prediction models were trained from natural speech corpora. We found the best prediction models by testing several machine learning methods and two different corpora. The paper describes the building, and training process as well as the results and their respective interpretation.

A method to represent arbitrary sequences (strings) is discussed. We emphasize the application of the method to the analysis of the similarity of sets of proteins expressed as sequences of amino acids. We define a pattern of arbitrary structure called a . An implementation of a detailed representation is discussed. We show that a protein may be expressed as a collection of metasymbols in a way such that the underlying structural similarities are easier to identify.