Text in Afrikaans / Die doel van hierdie studie was om te bepaal tot watter mate ’n suiwer patroongebaseerde benadering tot woordafbreking bevredigende resultate lewer. Die masjienleertegnieke kunsmatige neurale netwerke, beslissingsbome en die TEX-algoritme is ondersoek aangesien dit met letterpatrone uit woordelyste afgerig kan word om lettergreep- en saamgesteldewoordverdeling te doen. ’n Leksikon van Afrikaanse woorde is uit ’n korpus van elektroniese teks genereer. Om lyste vir lettergreep- en saamgesteldewoordverdeling te kry, is woorde in die leksikon in lettergrepe verdeel en saamgestelde woorde is in hul samestellende dele verdeel. Uit elkeen van hierdie lyste van ±183 000 woorde is ±10 000 woorde as toetsdata gereserveer terwyl die res as afrigtingsdata gebruik is. ’n Rekursiewe algoritme is vir saamgesteldewoordverdeling ontwikkel. In hierdie algoritme word alle ooreenstemmende woorde uit ’n verwysingslys (die leksikon) onttrek deur stringpassing van die begin en einde van woorde af. Verdelingspunte word dan op grond van woordlengte uit die samestelling van begin- en eindwoorde bepaal. Die algoritme is uitgebrei deur die tekortkominge van hierdie basiese prosedure aan te spreek. Neurale netwerke en beslissingsbome is afgerig en variasies van beide tegnieke is ondersoek om die optimale modelle te kry. Patrone vir die TEX-algoritme is met die OPatGen-program gegenereer. Tydens toetsing het die TEX-algoritme die beste op beide lettergreep- en saamgesteldewoordverdeling presteer met 99,56% en 99,12% akkuraatheid, respektiewelik. Dit kan dus vir woordafbreking gebruik word met min risiko vir afbrekingsfoute in gedrukte teks. Die neurale netwerk met 98,82% en 98,42% akkuraatheid op lettergreep- en saamgesteldewoordverdeling, respektiewelik, is ook bruikbaar vir lettergreepverdeling, maar dis meer riskant. Ons het bevind dat beslissingsbome te riskant is om vir lettergreepverdeling en veral vir woordverdeling te gebruik, met 97,91% en 90,71% akkuraatheid, respektiewelik. ’n Gekombineerde algoritme is ontwerp waarin saamgesteldewoordverdeling eers met die TEXalgoritme gedoen word, waarna die resultate van lettergreepverdeling deur beide die TEXalgoritme en die neurale netwerk gekombineer word. Die algoritme het 1,3% minder foute as die TEX-algoritme gemaak. ’n Toets op gepubliseerde Afrikaanse teks het getoon dat die risiko vir woordafbrekingsfoute in teks met gemiddeld tien woorde per re¨el ±0,02% is. / The aim of this study was to determine the level of success achievable with a purely pattern based approach to hyphenation in Afrikaans. The machine learning techniques artificial neural networks, decision trees and the TEX algorithm were investigated since they can be trained with patterns of letters from word lists for syllabification and decompounding. A lexicon of Afrikaans words was extracted from a corpus of electronic text. To obtain lists for syllabification and decompounding, words in the lexicon were respectively syllabified and compound words were decomposed. From each list of ±183 000 words, ±10 000 words were reserved as testing data and the rest was used as training data. A recursive algorithm for decompounding was developed. In this algorithm all words corresponding with a reference list (the lexicon) are extracted by string fitting from beginning and end of words. Splitting points are then determined based on the length of reassembled words. The algorithm was expanded by addressing shortcomings of this basic procedure. Artificial neural networks and decision trees were trained and variations of both were examined to find optimal syllabification and decompounding models. Patterns for the TEX algorithm were generated by using the program OPatGen. Testing showed that the TEX algorithm performed best on both syllabification and decompounding tasks with 99,56% and 99,12% accuracy, respectively. It can therefore be used for hyphenation in Afrikaans with little risk of hyphenation errors in printed text. The performance of the artificial neural network was lower, but still acceptable, with 98,82% and 98,42% accuracy for syllabification and decompounding, respectively. The decision tree with accuracy of 97,91% on syllabification and 90,71% on decompounding was found to be too risky to use for either of the tasks A combined algorithm was developed where words are first decompounded by using the TEX algorithm before syllabifying them with both the TEX algoritm and the neural network and combining the results. This algoritm reduced the number of errors made by the TEX algorithm by 1,3% but missed more hyphens. Testing the algorithm on Afrikaans publications showed the risk for hyphenation errors to be ±0,02% for text assumed to have an average of ten words per line. / Decision Sciences / D. Phil. (Operational Research)

In this dissertation a method for the classification of emotion in speech recordings made in a customer service call centre of a large business is presented. The problem addressed here is that customer service analysts at large businesses have to listen to large numbers of call centre recordings in order to discover customer service-related issues. Since recordings where the customer exhibits emotion are more likely to contain useful information for service improvement than “neutral” ones, being able to identify those recordings should save a lot of time for the customer service analyst. MTN South Africa agreed to provide assistance for this project. The system that has been developed for this project can interface with MTN’s call centre database, download recordings, classify them according to their emotional content, and provide feedback to the user. The system faces the additional challenge that it is required to classify emotion notwith- standing the fact that the caller may have one of several South African accents. It should also be able to function with recordings made at telephone quality sample rates. The project identifies several speech features that can be used to classify a speech recording according to its emotional content. The project uses these features to research the general methods by which the problem of emotion classification in speech can be approached. The project examines both a K-Nearest Neighbours Approach and an Artificial Neural Network- Based Approach to classify the emotion of the speaker. Research is also done with regard to classifying a recording according to the gender of the speaker using a neural network approach. The reason for this classification is that the gender of a speaker may be useful input into an emotional classifier. The project furthermore examines the problem of identifying smaller segments of speech in a recording. In the typical call centre conversation, a recording may start with the agent greeting the customer, the customer stating his or her problem, the agent performing an action, during which time no speech occurs, the agent reporting back to the user and the call being terminated. The approach taken by this project allows the program to isolate these different segments of speech in a recording and discard segments of the recording where no speech occurs. This project suggests and implements a practical approach to the creation of a classifier in a commercial environment through its use of a scripting language interpreter that can train a classifier in one script and use the trained classifier in another script to classify unknown recordings. The project also examines the practical issues involved in implementing an emotional clas- sifier. It addresses the downloading of recordings from the call centre, classifying the recording and presenting the results to the customer service analyst. AFRIKAANS : n Metode vir die klassifisering van emosie in spraakopnames in die oproepsentrum van ’n groot sake-onderneming word in hierdie verhandeling aangebied. Die probleem wat hierdeur aangespreek word, is dat kli¨entediens ontleders in ondernemings na groot hoeveelhede oproepsentrum opnames moet luister ten einde kli¨entediens aangeleenthede te identifiseer. Aangesien opnames waarin die kli¨ent emosie toon, heel waarskynlik nuttige inligting bevat oor diensverbetering, behoort die vermo¨e om daardie opnames te identifiseer vir die analis baie tyd te spaar. MTN Suid-Afrika het ingestem om bystand vir die projek te verleen. Die stelsel wat ontwikkel is kan opnames vanuit MTN se oproepsentrum databasis verkry, klassifiseer volgens emosionele inhoud en terugvoering aan die gebruiker verskaf. Die stelsel moet die verdere uitdaging kan oorkom om emosie te kan klassifiseer nieteenstaande die feit dat die spreker een van verskeie Suid-Afrikaanse aksente het. Dit moet ook in staat wees om opnames wat gemaak is teen telefoon gehalte tempos te analiseer. Die projek identifiseer verskeie spraak eienskappe wat gebruik kan word om ’n opname volgens emosionele inhoud te klassifiseer. Die projek gebruik hierdie eienskappe om die algemene metodes waarmee die probleem van emosie klassifisering in spraak benader kan word, na te vors. Die projek gebruik ’n K-Naaste Bure en ’n Neurale Netwerk benadering om die emosie van die spreker te klassifiseer. Navorsing is voorts gedoen met betrekking tot die klassifisering van die geslag van die spreker deur ’n neurale netwerk. Die rede vir hierdie klassifisering is dat die geslag van die spreker ’n nuttige inset vir ’n emosie klassifiseerder mag wees. Die projek ondersoek ook die probleem van identifisering van spraakgedeeltes in ’n opname. In ’n tipiese oproepsentrum gesprek mag die opname begin met die agent wat die kli¨ent groet, die kli¨ent wat sy of haar probleem stel, die agent wat ’n aksie uitvoer sonder spraak, die agent wat terugrapporteer aan die gebruiker en die oproep wat be¨eindig word. Die benadering van hierdie projek laat die program toe om hierdie verskillende gedeeltes te isoleer uit die opname en om gedeeltes waar daar geen spraak plaasvind nie, uit te sny. Die projek stel ’n praktiese benadering vir die ontwikkeling van ’n klassifiseerder in ’n kommersi¨ele omgewing voor en implementeer dit deur gebruik te maak van ’n programeer taal interpreteerder wat ’n klassifiseerder kan oplei in een program en die opgeleide klassifiseerder gebruik om ’n onbekende opname te klassifiseer met behulp van ’n ander program. Die projek ondersoek ook die praktiese aspekte van die implementering van ’n emosionele klassifiseerder. Dit spreek die aflaai van opnames uit die oproep sentrum, die klassifisering daarvan, en die aanbieding van die resultate aan die kli¨entediens analis, aan. Copyright / Dissertation (MEng)--University of Pretoria, 2010. / Electrical, Electronic and Computer Engineering / unrestricted

Text in Afrikaans / Die doel van hierdie studie was om te bepaal tot watter mate ’n suiwer patroongebaseerde benadering tot woordafbreking bevredigende resultate lewer. Die masjienleertegnieke kunsmatige neurale netwerke, beslissingsbome en die TEX-algoritme is ondersoek aangesien dit met letterpatrone uit woordelyste afgerig kan word om lettergreep- en saamgesteldewoordverdeling te doen. ’n Leksikon van Afrikaanse woorde is uit ’n korpus van elektroniese teks genereer. Om lyste vir lettergreep- en saamgesteldewoordverdeling te kry, is woorde in die leksikon in lettergrepe verdeel en saamgestelde woorde is in hul samestellende dele verdeel. Uit elkeen van hierdie lyste van ±183 000 woorde is ±10 000 woorde as toetsdata gereserveer terwyl die res as afrigtingsdata gebruik is. ’n Rekursiewe algoritme is vir saamgesteldewoordverdeling ontwikkel. In hierdie algoritme word alle ooreenstemmende woorde uit ’n verwysingslys (die leksikon) onttrek deur stringpassing van die begin en einde van woorde af. Verdelingspunte word dan op grond van woordlengte uit die samestelling van begin- en eindwoorde bepaal. Die algoritme is uitgebrei deur die tekortkominge van hierdie basiese prosedure aan te spreek. Neurale netwerke en beslissingsbome is afgerig en variasies van beide tegnieke is ondersoek om die optimale modelle te kry. Patrone vir die TEX-algoritme is met die OPatGen-program gegenereer. Tydens toetsing het die TEX-algoritme die beste op beide lettergreep- en saamgesteldewoordverdeling presteer met 99,56% en 99,12% akkuraatheid, respektiewelik. Dit kan dus vir woordafbreking gebruik word met min risiko vir afbrekingsfoute in gedrukte teks. Die neurale netwerk met 98,82% en 98,42% akkuraatheid op lettergreep- en saamgesteldewoordverdeling, respektiewelik, is ook bruikbaar vir lettergreepverdeling, maar dis meer riskant. Ons het bevind dat beslissingsbome te riskant is om vir lettergreepverdeling en veral vir woordverdeling te gebruik, met 97,91% en 90,71% akkuraatheid, respektiewelik. ’n Gekombineerde algoritme is ontwerp waarin saamgesteldewoordverdeling eers met die TEXalgoritme gedoen word, waarna die resultate van lettergreepverdeling deur beide die TEXalgoritme en die neurale netwerk gekombineer word. Die algoritme het 1,3% minder foute as die TEX-algoritme gemaak. ’n Toets op gepubliseerde Afrikaanse teks het getoon dat die risiko vir woordafbrekingsfoute in teks met gemiddeld tien woorde per re¨el ±0,02% is. / The aim of this study was to determine the level of success achievable with a purely pattern based approach to hyphenation in Afrikaans. The machine learning techniques artificial neural networks, decision trees and the TEX algorithm were investigated since they can be trained with patterns of letters from word lists for syllabification and decompounding. A lexicon of Afrikaans words was extracted from a corpus of electronic text. To obtain lists for syllabification and decompounding, words in the lexicon were respectively syllabified and compound words were decomposed. From each list of ±183 000 words, ±10 000 words were reserved as testing data and the rest was used as training data. A recursive algorithm for decompounding was developed. In this algorithm all words corresponding with a reference list (the lexicon) are extracted by string fitting from beginning and end of words. Splitting points are then determined based on the length of reassembled words. The algorithm was expanded by addressing shortcomings of this basic procedure. Artificial neural networks and decision trees were trained and variations of both were examined to find optimal syllabification and decompounding models. Patterns for the TEX algorithm were generated by using the program OPatGen. Testing showed that the TEX algorithm performed best on both syllabification and decompounding tasks with 99,56% and 99,12% accuracy, respectively. It can therefore be used for hyphenation in Afrikaans with little risk of hyphenation errors in printed text. The performance of the artificial neural network was lower, but still acceptable, with 98,82% and 98,42% accuracy for syllabification and decompounding, respectively. The decision tree with accuracy of 97,91% on syllabification and 90,71% on decompounding was found to be too risky to use for either of the tasks A combined algorithm was developed where words are first decompounded by using the TEX algorithm before syllabifying them with both the TEX algoritm and the neural network and combining the results. This algoritm reduced the number of errors made by the TEX algorithm by 1,3% but missed more hyphens. Testing the algorithm on Afrikaans publications showed the risk for hyphenation errors to be ±0,02% for text assumed to have an average of ten words per line. / Decision Sciences / D. Phil. (Operational Research)