Thesis (PhD)--Macquarie University, Division of Information and Communication Sciences, Department of Computing, 2008. / Thesis submitted for the joint institutional requirements for the double-badged degree of Doctor of Philosophy and Docteur de l'Université de Provence, Spécialité : Informatique. Includes bibliography (p. 229-240) and index.

Thesis (M.Comp.Sc.-Res.)--University of Wollongong, 2008. / Typescript. Includes bibliographical references: leaf 84-95.

Thesis (MEng)--University of Stellenbosch, 2004. / ENGLISH ABSTRACT: Automatic speech recognition applications have been developed for many languages in other countries but not much research has been conducted on developing Human Language Technology (HLT) for S.A. languages. Research has been performed on informally gathered speech data but until now a speech corpus that could be used to develop HLT for S.A. languages did not exist. With the development of the African Speech Technology Speech Corpora, it has now become possible to develop commercial applications of HLT. The two main objectives of this work are the accurate modelling of phonemes, suitable for the purposes of LVCSR, and the evaluation of the untried S.A. English speech corpus. Three different aspects of phoneme modelling was investigated by performing isolated phoneme recognition on the NTIMIT speech corpus. The three aspects were signal processing, statistical modelling of HMM state distributions and context-dependent phoneme modelling. Research has shown that the use of phonetic context when modelling phonemes forms an integral part of most modern LVCSR systems. To facilitate the context-dependent phoneme modelling, a method of constructing robust and accurate models using decision tree-based state clustering techniques is described. The strength of this method is the ability to construct accurate models of contexts that did not occur in the training data. The method incorporates linguistic knowledge about the phonetic context, in conjunction with the training data, to decide which phoneme contexts are similar and should share model parameters. As LVCSR typically consists of continuous recognition of spoken words, the contextdependent and context-independent phoneme models that were created for the isolated recognition experiments are evaluated by performing continuous phoneme recognition. The phoneme recognition experiments are performed, without the aid of a grammar or language model, on the S.A. English corpus. As the S.A. English corpus is newly created, no previous research exist to which the continuous recognition results can be compared to. Therefore, it was necessary to create comparable baseline results, by performing continuous phoneme recognition on the NTIMIT corpus. It was found that acceptable recognition accuracy was obtained on both the NTIMIT and S.A. English corpora. Furthermore, the results on S.A. English was 2 - 6% better than the results on NTIMIT, indicating that the S.A. English corpus is of a high enough quality that it can be used for the development of HLT. / AFRIKAANSE OPSOMMING: Automatiese spraak-herkenning is al ontwikkel vir ander tale in ander lande maar, daar nog nie baie navorsing gedoen om menslike taal-tegnologie (HLT) te ontwikkel vir Suid- Afrikaanse tale. Daar is al navorsing gedoen op spraak wat informeel versamel is, maar tot nou toe was daar nie 'n spraak databasis wat vir die ontwikkeling van HLT vir S.A. tale. Met die ontwikkeling van die African Speech Technology Speech Corpora, het dit moontlik geword om HLT te ontwikkel vir wat geskik is vir kornmersiele doeleindes. Die twee hoofdoele van hierdie tesis is die akkurate modellering van foneme, geskik vir groot-woordeskat kontinue spraak-herkenning (LVCSR), asook die evaluasie van die S.A. Engels spraak-databasis. Drie aspekte van foneem-modellering word ondersoek deur isoleerde foneem-herkenning te doen op die NTIMIT spraak-databasis. Die drie aspekte wat ondersoek word is sein prosessering, statistiese modellering van die HMM toestands distribusies, en konteksafhanklike foneem-modellering. Navorsing het getoon dat die gebruik van fonetiese konteks 'n integrale deel vorm van meeste moderne LVCSR stelsels. Dit is dus nodig om robuuste en akkurate konteks-afhanklike modelle te kan bou. Hiervoor word 'n besluitnemingsboom- gebaseerde trosvormings tegniek beskryf. Die tegniek is ook in staat is om akkurate modelle te bou van kontekste van nie voorgekom het in die afrigdata nie. Om te besluit watter fonetiese kontekste is soortgelyk en dus model parameters moet deel, maak die tegniek gebruik van die afrigdata en inkorporeer taalkundige kennis oor die fonetiese kontekste. Omdat LVCSR tipies is oor die kontinue herkenning van woorde, word die konteksafhanklike en konteks-onafhanklike modelle, wat gebou is vir die isoleerde foneem-herkenningseksperimente, evalueer d.m.v. kontinue foneem-herkening. Die kontinue foneemherkenningseksperimente word gedoen op die S.A. Engels databasis, sonder die hulp van 'n taalmodel of grammatika. Omdat die S.A. Engels databasis nuut is, is daar nog geen ander navorsing waarteen die result ate vergelyk kan word nie. Dit is dus nodig om kontinue foneem-herkennings result ate op die NTIMIT databasis te genereer, waarteen die S.A. Engels resulte vergelyk kan word. Die resulate dui op aanvaarbare foneem her kenning op beide die NTIMIT en S.A. Engels databassise. Die resultate op S.A. Engels is selfs 2 - 6% beter as die resultate op NTIMIT, wat daarop dui dat die S.A. Engels spraak-databasis geskik is vir die ontwikkeling van HLT.

Thesis (MScEng (Electrical and Electronic Engineering))--University of Stellenbosch, 2011. / ENGLISH ABSTRACT: This thesis presents a system that is designed to replace the manual process of generating a pronunciation dictionary for use in automatic speech recognition. The proposed system has several stages. The first stage segments the audio into what will be known as the subword units, using a frequency domain method. In the second stage, dynamic time warping is used to determine the similarity between the segments of each possible pair of these acoustic segments. These similarities are used to cluster similar acoustic segments into acoustic clusters. The final stage derives a pronunciation dictionary from the orthography of the training data and corresponding sequence of acoustic clusters. This process begins with an initial mapping between words and their sequence of clusters, established by Viterbi alignment with the orthographic transcription. The dictionary is refined iteratively by pruning redundant mappings, hidden Markov model estimation and Viterbi re-alignment in each iteration. This approach is evaluated experimentally by applying it to two subsets of the TIMIT corpus. It is found that, when test words are repeated often in the training material, the approach leads to a system whose accuracy is almost as good as one trained using the phonetic transcriptions. When test words are not repeated often in the training set, the proposed approach leads to better results than those achieved using the phonetic transcriptions, although the recognition is poor overall in this case. / AFRIKAANSE OPSOMMING: Die doelwit van die tesis is om ’n stelsel te beskryf wat ontwerp is om die handgedrewe proses in die samestelling van ’n woordeboek, vir die gebruik in outomatiese spraakherkenningsstelsels, te vervang. Die voorgestelde stelsel bestaan uit ’n aantal stappe. Die eerste stap is die segmentering van die oudio in sogenaamde sub-woord eenhede deur gebruik te maak van ’n frekwensie gebied tegniek. Met die tweede stap word die dinamiese tydverplasingsalgoritme ingespan om die ooreenkoms tussen die segmente van elkeen van die moontlike pare van die akoestiese segmente bepaal. Die ooreenkomste word dan gebruik om die akoestiese segmente te groepeer in akoestiese groepe. Die laaste stap stel die woordeboek saam deur gebruik te maak van die ortografiese transkripsie van afrigtingsdata en die ooreenstemmende reeks akoestiese groepe. Die finale stap begin met ’n aanvanklike afbeelding vanaf woorde tot hul reeks groep identifiseerders, bewerkstellig deur Viterbi belyning en die ortografiese transkripsie. Die woordeboek word iteratief verfyn deur oortollige afbeeldings te snoei, verskuilde Markov modelle af te rig en deur Viterbi belyning te gebruik in elke iterasie. Die benadering is getoets deur dit eksperimenteel te evalueer op twee subversamelings data vanuit die TIMIT korpus. Daar is bevind dat, wanneer woorde herhaal word in die afrigtingsdata, die stelsel se benadering die akkuraatheid ewenaar van ’n stelsel wat met die fonetiese transkripsie afgerig is. As die woorde nie herhaal word in die afrigtingsdata nie, is die akkuraatheid van die stelsel se benadering beter as wanneer die stelsel afgerig word met die fonetiese transkripsie, alhoewel die akkuraatheid in die algemeen swak is.

This thesis investigates the role of linguistically-motivated generative models of syntax and semantic structure in natural language processing (NLP). Syntactic well-formedness is crucial in language generation, but most statistical models do not account for the hierarchical structure of sentences. Many applications exhibiting natural language understanding rely on structured semantic representations to enable querying, inference and reasoning. Yet most semantic parsers produce domain-specific or inadequately expressive representations. We propose a series of generative transition-based models for dependency syntax which can be applied as both parsers and language models while being amenable to supervised or unsupervised learning. Two models are based on Markov assumptions commonly made in NLP: The first is a Bayesian model with hierarchical smoothing, the second is parameterised by feed-forward neural networks. The Bayesian model enables careful analysis of the structure of the conditioning contexts required for generative parsers, but the neural network is more accurate. As a language model the syntactic neural model outperforms both the Bayesian model and n-gram neural networks, pointing to the complementary nature of distributed and structured representations for syntactic prediction. We propose approximate inference methods based on particle filtering. The third model is parameterised by recurrent neural networks (RNNs), dropping the Markov assumptions. Exact inference with dynamic programming is made tractable here by simplifying the structure of the conditioning contexts. We then shift the focus to semantics and propose models for parsing sentences to labelled semantic graphs. We introduce a transition-based parser which incrementally predicts graph nodes (predicates) and edges (arguments). This approach is contrasted against predicting top-down graph traversals. RNNs and pointer networks are key components in approaching graph parsing as an incremental prediction problem. The RNN architecture is augmented to condition the model explicitly on the transition system configuration. We develop a robust parser for Minimal Recursion Semantics, a linguistically-expressive framework for compositional semantics which has previously been parsed only with grammar-based approaches. Our parser is much faster than the grammar-based model, while the same approach improves the accuracy of neural Abstract Meaning Representation parsing.

We present a multifaceted investigation into the relevance of word order in machine translation. We introduce two tools, DTED and DERP, each using dependency structure to detect differences between the structures of machine-produced translations and human-produced references. DTED applies the principle of Tree Edit Distance to calculate edit operations required to convert one structure into another. Four variants of DTED have been produced, differing in the importance they place on words which match between the two sentences. DERP represents a more detailed procedure, making use of the dependency relations between words when evaluating the disparities between paths connecting matching nodes. In order to empirically evaluate DTED and DERP, and as a standalone contribution, we have produced WOJ-DB, a database of human judgments. Containing scores relating to translation adequacy and more specifically to word order quality, this is intended to support investigations into a wide range of translation phenomena. We report an internal evaluation of the information in WOJ-DB, then use it to evaluate variants of DTED and DERP, both to determine their relative merit and their strength relative to third-party baselines. We present our conclusions about the importance of structure to the tools and their relevance to word order specifically, then propose further related avenues of research suggested or enabled by our work.

Popular fiction books describe rich visual environments that contain characters, objects, and behaviour. This research develops automated processes for converting text sourced from fiction books into animated virtual environments and multi-modal films. This involves the analysis of unrestricted natural language fiction to identify appropriate visual descriptions, and the interpretation of the identified descriptions for constructing animated 3D virtual environments. The goal of the text analysis stage is the creation of annotated fiction text, which identifies visual descriptions in a structured manner. A hierarchical rule-based learning system is created that induces patterns from example annotations provided by a human, and uses these for the creation of additional annotations. Patterns are expressed as tree structures that abstract the input text on different levels according to structural (token, sentence) and syntactic (parts-of-speech, syntactic function) categories. Patterns are generalized using pair-wise merging, where dissimilar sub-trees are replaced with wild-cards. The result is a small set of generalized patterns that are able to create correct annotations. A set of generalized patterns represents a model of an annotator's mental process regarding a particular annotation category. Annotated text is interpreted automatically for constructing detailed scene descriptions. This includes identifying which scenes to visualize, and identifying the contents and behaviour in each scene. Entity behaviour in a 3D virtual environment is formulated using time-based constraints that are automatically derived from annotations. Constraints are expressed as non-linear symbolic functions that restrict the trajectories of a pair of entities over a continuous interval of time. Solutions to these constraints specify precise behaviour. We create an innovative quantified constraint optimizer for locating sound solutions, which uses interval arithmetic for treating time and space as contiguous quantities. This optimization method uses a technique of constraint relaxation and tightening that allows solution approximations to be located where constraint systems are inconsistent (an ability not previously explored in interval-based quantified constraint solving). 3D virtual environments are populated by automatically selecting geometric models or procedural geometry-creation methods from a library. 3D models are animated according to trajectories derived from constraint solutions. The final animated film is sequenced using a range of modalities including animated 3D graphics, textual subtitles, audio narrations, and foleys. Hierarchical rule-based learning is evaluated over a range of annotation categories. Models are induced for different categories of annotation without modifying the core learning algorithms, and these models are shown to be applicable to different types of books. Models are induced automatically with accuracies ranging between 51.4% and 90.4%, depending on the category. We show that models are refined if further examples are provided, and this supports a boot-strapping process for training the learning mechanism. The task of interpreting annotated fiction text and populating 3D virtual environments is successfully automated using our described techniques. Detailed scene descriptions are created accurately, where between 83% and 96% of the automatically generated descriptions require no manual modification (depending on the type of description). The interval-based quantified constraint optimizer fully automates the behaviour specification process. Sample animated multi-modal 3D films are created using extracts from fiction books that are unrestricted in terms of complexity or subject matter (unlike existing text-to-graphics systems). These examples demonstrate that: behaviour is visualized that corresponds to the descriptions in the original text; appropriate geometry is selected (or created) for visualizing entities in each scene; sequences of scenes are created for a film-like presentation of the story; and that multiple modalities are combined to create a coherent multi-modal representation of the fiction text. This research demonstrates that visual descriptions in fiction text can be automatically identified, and that these descriptions can be converted into corresponding animated virtual environments. Unlike existing text-to-graphics systems, we describe techniques that function over unrestricted natural language text and perform the conversion process without the need for manually constructed repositories of world knowledge. This enables the rapid production of animated 3D virtual environments, allowing the human designer to focus on creative aspects.

The need for interacting with machines using spoken natural language is growing, along with the expectation that synthetic speech in this context sound natural. Such interaction includes answering questions, where prosody plays an important role in producing natural English synthetic speech by communicating the information structure of utterances. CCG is a theoretical framework that exploits the notion that, in English, information structure, prosodic structure and syntactic structure are isomorphic. This provides a way to convert a semantic representation of an utterance into a prosodically natural spoken utterance. GF is a framework for writing grammars, where abstract tree structures capture the semantic structure and concrete grammars render these structures in linearised strings. This research combines these frameworks to develop a system that converts semantic representations of utterances into linearised strings of natural language that are marked up to inform the prosody-generating component of a speech synthesis system. / Computing / M. Sc. (Computing)

Orientador: Jacques Wainer / Tese (doutorado) - Universidade Estadual de Campinas, Instituto de Computação / Made available in DSpace on 2018-08-24T00:42:33Z (GMT). No. of bitstreams: 1 Passos_AlexandreTachard_D.pdf: 2615030 bytes, checksum: 93841a46120b968f6da6c9aea28953b7 (MD5) Previous issue date: 2013 / Resumo: Em processamento de linguagem natural, e em aprendizado de máquina em geral, é comum o uso de modelos gráficos probabilísticos (probabilistic graphical models). Embora estes modelos sejam muito convenientes, possibilitando a expressão de relações complexas entre várias variáveis que se deseja prever dado uma sentença ou um documento, algoritmos comuns de aprendizado e de previsão utilizando estes modelos são frequentemente ineficientes. Por isso têm-se explorado recentemente o uso de relaxações usando programação linear deste problema de inferência. Esta tese apresenta duas contribuições para a teoria e prática de relaxações de programação linear para inferência em modelos probabilísticos gráficos. Primeiro, apresentamos um novo algoritmo, baseado na técnica de geração de colunas (dual à técnica dos planos de corte) que acelera a execução do algoritmo de Viterbi, a técnica mais utilizada para inferência em modelos lineares. O algoritmo apresentado também se aplica em modelos que são árvores e em hipergrafos. Em segundo mostramos uma nova relaxação linear para o problema de inferência conjunta, quando se quer acoplar vários modelos, em cada qual inferência é eficiente, mas em cuja junção inferência é NP-completa. Esta tese propõe uma extensão à técnica de decomposição dual (dual decomposition) que permite além de juntar vários modelos a adição de fatores que tocam mais de um submodelo eficientemente / Abstract: In natural language processing, and in general machine learning, probabilistic graphical models (and more generally structured linear models) are commonly used. Although these models are convenient, allowing the expression of complex relationships between many random variables one wants to predict given a document or sentence, most learning and prediction algorithms for general models are inefficient. Hence there has recently been interest in using linear programming relaxations for the inference tasks necessary when learning or applying these models. This thesis presents two contributions to the theory and practice of linear programming relaxations for inference in structured linear models. First we present a new algorithm, based on column generation (a technique which is dual to the cutting planes method) to accelerate the Viterbi algorithm, the most popular exact inference technique for linear-chain graphical models. The method is also applicable to tree graphical models and hypergraph models. Then we present a new linear programming relaxation for the problem of joint inference, when one has many submodels and wants to predict using all of them at once. In general joint inference is NP-complete, but algorithms based on dual decomposition have proven to be efficiently applicable for the case when the joint model can be expressed as many separate models plus linear equality constraints. This thesis proposes an extension to dual decomposition which allows also the presence of factors which score parts that belong in different submodels, improving the expressivity of dual decomposition at no extra computational cost / Doutorado / Ciência da Computação / Doutor em Ciência da Computação

Twitter has quickly become one of the most popular social media sites. It has 313 million monthly active users, and 500 million tweets are published daily. With the massive number of tweets, Twitter users share information about a location along with the temporal awareness. In this work, I focus on tweets where author of the tweets exclusively mentions a location in the tweet. Natural language processing systems can leverage wide range of information from the tweets to build applications like recommender systems that predict the location of the author. This kind of system can be used to increase the visibility of the targeted audience and can also provide recommendations interesting places to visit, hotels to stay, restaurants to eat, targeted on-line advertising, and co-traveler matching based on the temporal information extracted from a tweet. In this work I determine if the author of the tweet is present in the mentioned location of the tweet. I also determine if the author is present in the location before tweeting, while tweeting, or after tweeting. I introduce 5 temporal tags (before the tweet but > 24 hours; before the tweet but < 24 hours; during the tweet is posted; after the tweet is posted but < 24 hours; and after the tweet is posted but > 24 hours). The major contributions of this paper are: (1) creation of a corpus of 1062 tweets containing 1200 location named entities, containing annotations whether author of a tweet is or is not located in the location he tweets about with respect to 5 temporal tags; (2) detailed corpus analysis including real annotation examples and label distributions per temporal tag; (3) detailed inter-annotator agreements, including Cohen's kappa, Krippendorff's alpha and confusion matrices per temporal tag; (4) label distributions and analysis; and (5) supervised learning experiments, along with the results.