doktorska disertacija
Miha Pavlinek (Author), Vili Podgorelec (Mentor)

Abstract

V svetu vseprisotnega računalništva se s kopičenjem naprav ter množično uporabo družbenih omrežij, elektronske komunikacije in drugih oblik IKT storitev naglo povečuje tudi količina nestrukturiranih vsebin. To nas sili k uporabi inteligentnih rešitev, ki za nas te vsebine organizirajo, se namesto nas odločajo o njihovi pomembnosti in nam posredujejo zgolj najbolj relevantne med njimi. Osnovna zmožnost takšnih rešitev je klasifikacija vsebin, zato so v njih avtomatski klasifikatorji nepogrešljiv člen. Zanje je tipično, da za učenje potrebujejo številne označene primerke z ustrezno predstavitvijo, v praksi pa označeni primerki niso vedno na voljo, zato je potrebno avtomatske klasifikatorje prilagoditi tako, da so sposobni pri učenju uporabljati tudi druge, neoznačene vsebine. V disertaciji smo predstavili metodo ST LDA (ang. Self-Training with LDA) za klasifikacijo besedil, ki za učenje klasifikatorja potrebuje le minimalno množico označenih in veliko večjo množico neoznačenih primerkov. Predlagali smo algoritem, ki temelji na metodi samoučenja ter predstavitvi besedil na osnovi tematskega modela, kar prinaša dodatne faktorje, od katerih je odvisna njegova uspešnost. Za vsak faktor smo, na podlagi številnih eksperimentov nad sedmimi besedilnimi podatkovnimi zbirkami, ocenili vpliv na uspešnost klasifikacije ter definirali model za določanje vrednosti parametrov, s čimer se izognemo dodatnim nastavitvam. Uspešnost metode smo primerjali z uspešnostjo drugih uveljavljenih metod in predstavitev, pri čemer predlagana metoda ST LDA dosega nadpovprečne rezultate, kar smo navsezadnje potrdili z neparametričnimi statističnimi testi.

Keywords

obdelava naravnega jezika;tekstovno rudarjenje;klasifikacija;tematsko modeliranje;delno nadzorovano učenje;samoučenje;doktorske disertacije;

Data

Language: Slovenian
Year of publishing:
Typology: 2.08 - Doctoral Dissertation
Organization: UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher: [M. Pavlinek]
UDC: 004.89:316.472.4(043.3)
COBISS: 19820310 Link will open in a new window
Views: 1400
Downloads: 162
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Development of an intelligent decision support model based on unstructured content analysis
Secondary abstract: In a world of ubiquitous computing, with an explosion in the availability of devices and the widespread use of social networks, electronic communications and other forms of ICT services, unstructured content has been rapidly increasing. This forces us to use intelligent systems to organize our content, make decisions about their importance and recommend only the most relevant among them. The core characteristic of such systems is the ability to classify content. Therefore, automatic classifiers play an indispensable role. They are usually limited to supervised learning, where all the data is labeled. In practice, however, labeled examples are not always available, so it is necessary to adapt classifiers to learn from both labeled and unlabeled data. In this dissertation, we propose a Self-Training with LDA (ST LDA) method for text classification in the presence of a minimal amount of labeled data and a much larger set of unlabeled data. It is based on a self-training method and text representation based on topic models, which brings additional factors that affect method’s performance. The influence of each factor was estimated via several variations of the experiment over seven datasets and according to achievements, we developed a general model with which every parameter can be precisely defined for any given collection. The performance of the method is compared with other state-of-the-art methods and methods based on typical representations. It turned out that ST LDA outperforms other compared methods and variations. This was additionally confirmed with nonparametric statistical tests.
Secondary keywords: natural language processing;text mining;classification;topic modeling;semi-supervised learning;self-training;Odločanje;Disertacije;Modeli;Razvoj;
URN: URN:SI:UM:
Type (COBISS): Dissertation
Thesis comment: Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko
Pages: VIII, 133 str.
ID: 9161059