Back to search

Aristotle University Of Thessaloniki (AUTH)

Γλωσσική τεχνολογία στατιστικά μοντέλα για γλώσσα, ομιλία και εικόνα

Abstract

dc:description

The present thesis studies the statistical modeling of information inherent in text, speech, and images in combination with clustering based on various similarity criteria, so that multimedia documents are efficiently organized and handled even in a constantly changing environment. In more detail, the research lies in two main fields: language modeling and clustering. In the field of language modeling, the classic bigram model, the long distance bigram models, with and without interpolation, and the trigger-pairs together with smoothing were studied for word clustering in groups of conceptual synonyms. For this purpose, the language models were incorporated in three different clustering algorithms. The clustering results were evaluated with relative and external validity indices. In the special case of estimating the external indices where a ground truth is needed this was extracted systematically for the first time by means of Wordnet. The language models were also studied in image processing by proposing a new multi-channel color histogram equalization method that entails the smoothed bigram model. The new method was additionally extended with a non-linear transformation in order to deal with the potential problem of out of gamut colors. In the research field of clustering, the research was also extended in speech by proposing a new speaker diarization method that exploits the eigengap criterion for estimating the number of speakers and the cluster ensembles, which are produced by various clustering algorithms. Moreover, the problem of outliers (that is, singleton clusters) was studied by presenting two techniques for dealing with them. Finally, the updating of the Probabilistic Latent Semantic Analysis (PLSA) model was studied that was estimated from an initial document collection, when new documents are added in the initial document collection. Two new updating methods were proposed for updating the parameters of the statistical asymmetric and symmetric aspect model. The first method (Recursive PLSA) incorporates documents that are added one by one in an existing document collection. The second method (extended RPLSA) consists an extension of the first method, since it adds new documents, which may contain words from an open vocabulary, in batches. In addition, the initialization of the conditional probabilities for the incoming documents and of the initial PLSA model parameter was studied. The proposed updating methods were proved superior to corresponding updating methods with respect to correctly estimating the PLSA model parameters, to their speed and their effective application in document clustering.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2012

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Bassiou, Nikoletta
  • Μπάσιου, Νικολέττα

Subjects

dc:subject × 13

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/30145

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Bassiou, Nikoletta; Μπάσιου, Νικολέττα. Γλωσσική τεχνολογία στατιστικά μοντέλα για γλώσσα, ομιλία και εικόνα. Aristotle University Of Thessaloniki (AUTH), 2012. http://hdl.handle.net/10442/hedi/30145