Abstract
dc:descriptionThe subject of this PhD thesis is the study of multimodal language processing. The research is two-fold; on one hand it deals with statistical speech analysis and processing, in particular speech segmentation, while on the other hand it deals with text processing, Web information retrieval, and text language identification of multilingual documents. The research in speech processing is focused on the identification of phone boundaries in continuous speech using Bayesian analysis for the statistical hypothesis testing. Emphasis is given to the multivariate statistical analysis of the interrelationships between the Mel frequency cepstral coefficients and the independency of the observations for small sample sizes is tested, in order to estimate the parameters of the exponential distributions and to assess the multivariate likelihood ratio tests. Model selection techniques based on Akaike and Bayesian information criteria are used, which are derived from entropic and Bayes theories respectively. In specific, the hybrid method DISTBIC of Delacourt and Wellekens, which has been successfully applied to speech and speaker segmentation, is studied. The potential of its application to segmentation problems, where the analysis windows are small and the decision making distances are in word or phone level, is investigated. The method is appropriately adjusted so that the heuristic estimation of its parameters is limited, while its performance is adequate in speech with low signalto- noise ratios. Specifically, we propose the replacement of Bayesian information criterion with others criteria, such as the Bayesian Information Criterion-Corrected and the Approximate Bayes Factors criterion, which demonstrate better efficiency in situations with small sample sizes and, in parallel, we model speech features with the generalised Gamma distribution assuming independence between the Mel frequency cepstral coefficients. We evaluate these modifications in additive noisy environments using two speech datasets, NTIMIT and M2VTS, where we yield significant improvement in false alarm and missed detection rates relative to the baseline algorithm DISTBIC. In a different study, we take into account the role of parameter interrelationships in the distribution, by regarding the Information Complexity criterion. We examine the application of Bayesian Information and Information Complexity when we use statistically robust estimators for the covariance matrices while avoiding the demand of statistical independence between the variables. Evaluating these adjustments for the problem of phone segmentation in speech under additive noise, we yield total performance improvement in boundary detection over 5% in 13 minutes of speech signal. Regarding text processing, research is done on focused crawling of World Wide Web pages, in order to create vertical search engines with low computational cost. The proposed algorithm, Hyper-Content Latent Analysis (HCLA), uses the method of latent semantic indexing in order to incorporate text features (text content of the web pages) to link analysis and follows best first search visiting strategy. Significant performance improvement is attained, in both precision and recall rates, with respect to well known information retrieval techniques, such as PageRank and Shark-Search. In particular, in a corpus of 1638 web pages the improvement in precision is 24% and 40% respectively, when the recall approaches 100%, In a corpus of 582 hypertexts, the precision of the proposed algorithm is significantly better than the Probabilistic Lantent Semantic Indexing algorithm for high recall rates, while the two algorithms demonstrate comparative precision rates in low and medium recall rates.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2008
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Αλμπανίδης, Γεώργιος
- Almpanidis, Georgios
Subjects
dc:subject × 20- Κατάτμηση φθόγγων
- Πληροφοριακά κριτήρια
- Γενικευμένη Γάμμα κατανομή
- Εύρωστη στατιστική επιλογή μοντέλου
- Αυτόματη προσκομιδή ιστοσελίδων
- Λανθάνουσα σημασιολογική δεικτοδότηση
- Ταυτοποίηση γλώσσας κειμένων
- Κρυμμένα μοντέλα Μαρκώφ
- Phonemic segmentation
- Information criteria
- Generalised Gamma distribution
- Robust statistical model selection
- Focused web crawling
- Latent semantic indexing
- Text language identification
- Hidden Markov models
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Natural Sciences
- Computer and Information Sciences
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/18866