1. Doctorado

Now showing 1 - 2 of 2

Towards automatic detection of lexical borrowings in wordlists - with application to Latin American languages
(Pontificia Universidad Católica del Perú, 2024-11-18) Miller, John Edward; Beltrán Castañón, César Armando; Zariquiey Biondi, Roberto Daniel; List, Johann-Mattis
Knowing what words of a language are inherited from the ancestor language, which are borrowed from contact languages, which are recently created, and the timing of critical events in the culture, enables modeling of language history including language phylogeny, language contact, and other novel influences on the culture. However, determining which words or forms are borrowed and from whom is a difficult, time consuming, and often fascinating task, usually performed by historical linguists, which is limited by the time and expertise available. While there are semi-automated methods available to identify borrowed words and their word donors, there is still substantial opportunity for improvement. We construct a new language model based monolingual method, competing cross-entropies, based on word source groupings within monolingual wordlists; improve existing multilingual sequence comparison methods, closest match on language pairs and cognate-based on multiple languages; and construct a classifier based meta-method, combining closest match and cross-entropy functions. We also define an alternative goal of borrowing detection for dominant donor languages, which allows determination of both borrowing and source. We apply monolingual methods to a global dataset of 41 languages, and multilingual and meta methods to a newly constituted dataset of seven Latin American languages. We also initiate work on a dataset of 21 Pano-Tacanan and regional languages with added Spanish, Portuguese, and Quechua donor languages for subsequent application of borrowing detection methods. The competing cross-entropies method establishes a benchmark for automatic borrowing detection for the world online loan database, the dominant donor multiple sequence comparison method improves over the competing cross-entropies method, and the classifier meta-method with sequence comparison and crossentropy functions performs substantially better overall.
Classifier based on straight line segments: an overview and theoretical improvements
(Pontificia Universidad Católica del Perú, 2022-09-09) Medina Rodríguez, Rosario Alejandra; Beltrán Castañón, César Armando
Literature offers several supervised machine learning algorithms focused on binary classification for solving daily problems. Compared to well-known conventional classifiers, the Straight-line Segment Classifier (SLS Classifier) stands out for its low complexity and competitiveness. It takes advantage of some good characteristics of Learning Vector Quantization and Nearest Feature Line. In addition, it has lower computational complexity than Support Vector Machines. The SLS binary classifier is based on distances between a set of points and two sets of straight line segments. Therefore, it involves finding the optimal placement of straight line segment extremities to achieve the minimum mean square error. In previous works, we explored three different evolutive algorithms as optimization methods to increase the possibilities of finding a global optimum generating different solutions as the initial population. Additionally, we proposed a new way of estimating the number of straight line segments by applying an unsupervised clustering method. However, some interesting questions remained to be further analyzed, such as a detailed analysis of the parameters and base definitions of the optimization algorithm. Furthermore, it was straightforward that the straight-line segment lengths can grow significantly during the training phase, negatively impacting the classification rate. Therefore, the main goal of this thesis is to outline the SLS Classifier baseline and propose some theoretical improvements, such as (i) Formulating an optimization approach to provide optimal final positions for the straight line segments; (ii) Proposing a model selection approach for the SLS Classifier; and, (iii) Determining the SLS Classifier performance when applied on real problems (10 artificial and 8 UCI public datasets). The proposed methodology showed promising results compared to the original SLS Classifier version and other classifiers. Moreover, this classifier can be used in research and industry for decisionmaking problems due to the straightforward interpretation and classification rates.

1. Doctorado

Browse

Filters

Settings

Sort By

Results per page

Search Results