This study aims to build a new language model that can be used in a Turkish large vocabulary continuous speech recognition system. Turkish is a very productive language in terms of word forms because of its agglutinative nature. For such languages like Turkish, the vocabulary size is far from being acceptable. From only one simple stem, thousands of new word forms can be generated using inflectional or derivational suffixes. In this thesis, words are parsed into their stems and endings. One ending includes the suffixes attached to the associated root. Then the search network based on bigrams is constructed. Bigrams are obtained either using stem and endings, or using only stems. The language model proposed is based on bigrams obtained using only stems. All work is done in HTK (Hidden Markov Model Toolkit) environment, except parsing and network transforming. Besides of offering a new language model for Turkish, this study involves a comprehensive work about speech recognition inspecting into concepts in the state of the art speech recognition systems. To acquire good command of these concepts and processes in speech recognition isolated word, connected word and continuous speech recognition tasks are performed. The experimental results associated with these tasks are also given.

In the modern world, Internet has become a popular place, people with speech hearing disabilities and search engines can't take part of speech content in podcast les. In order to solve the problem partially, the Sphinx decoders such as Sphinx-3, Sphinx-4 can be used to implement a Auto Transcript Generator application, by coupling already existing large acoustic model, language model and a existing dictionary, or by training your own large acoustic model, language model and creating your own dictionary to support continuous speaker independent speech recognition system.

Les différents dialectes de la langue arabe (DA) présentent de grandes variations phonologiques, morphologiques, lexicales et syntaxiques par rapport à la langue Arabe Standard Moderne (MSA). Jusqu’à récemment, ces dialectes n’étaient présents que sous leurs formes orales et la plupart des ressources existantes pour la langue arabe se limite à l’Arabe Standard (MSA), conduisant à une abondance d’outils pour le traitement automatique de cette variété. Étant donné les différences significatives entre le MSA et les DA, les performances de ces outils s’écroulent lors du traitement des DA. Cette situation conduit à une augmentation notable de l’ambiguïté dans les approches computationnelles des DA. Les travaux décrits dans cette thèse s’inscrivent dans ce cadre à travers la modélisation de l’oral parlé dans les médias tunisiens. Cette source de données contient une quantité importante d’Alternance Codique (AC) entre la langue normative MSA et le dialecte parlé en Tunisie (DT). La présence de ce dernier d’une manière désordonnée dans le discours pose une sérieuse problématique pour le Traitement Automatique de Langue et fait de cet oral une langue peu dotée. Toutefois, les ressources nécessaires pour modéliser cet oral sont quasiment inexistantes. Ainsi, l’objectif de cette thèse consiste à pallier ce manque afin de construire un modèle de langage dédié à un système de reconnaissance automatique pour l’oral parlé dans les médias tunisiens. Pour ce fait, nous décrivons dans cette thèse une méthodologie de création de ressources et nous l’évaluons par rapport à une tâche de modélisation de langage. Les résultats obtenu sont encourageants. / The different dialects of the arabic language have a large phonological, morphological, lexical and syntactic variations when compared to the standard written arabic language called MSA (Modern Standard Arabic). Until recently, these dialects were presented only in their oral form and most of the existing resources for the Arabic language is limited to the Standard Arabic (MSA), leading to an abundance of tools for the automatic processing of this variety. Given the significant differences between the MSA and DA, the performance of these tools fall down when processing AD. This situation leads to a significant increase of the ambiguity in computational approaches of AD.This thesis is part of this framework by modeling the oral spoken in the Tunisian media. This data source contains a significant amount of Code Switching (CS) between the normative language MSA and the Dialect spoken in Tunisia (DT). The presence of the latter in a disorderly manner in the discourse poses a serious problem for NLP (Natural Language Processing) and makes this oral a less resourced language. However, the resources required to model this oral are almost nonexistent. Thus, the objective of this thesis is to fill this gap in order to build a language model dedicated to an automatic recognition system for the oral spoken in the Tunisian media. For this reason, we describe in this thesis a resource generation methodologyand we evaluate it relative to a language modeling task. The results obtained are encouraging.

京都大学 / 0048 / 新制・課程博士 / 博士(医学) / 甲第22693号 / 医博第4637号 / 新制||医||1045(附属図書館) / 京都大学大学院医学研究科医学専攻 / (主査)教授 高橋 淳, 教授 林 康紀, 教授 伊佐 正 / 学位規則第4条第1項該当 / Doctor of Medical Science / Kyoto University / DFAM

The goal of this project was to create a chatbot capable of debating a user using limited resources including a discussion thread from the online debate forum Kialo. A retrieval based bot was designed and the discussion thread was converted into a database which the bot could interpret and choose an appropriate answer from. Which answer is appropriate is decided by the bot using a few key features in a given input sentence. The main features are word similarity, sentiment distance and BERT-encoding (a model for vector representation of text created by Google). The similarity of these features where then used to score claims from the dataset. Combining and weighting the scores was then used to find the correct response to a given input sentence. The most successful of the features was BERT-encoding. Once the bot had been refined it was brought online and tested using the communication platform Discord.

This thesis represents the Model-View-Controller pattern. It is focused especially at the description of the particular architecture layers principle and its functionality. It deals with the reasons of the three-layer architecture invention and it also deals with the advantages and the disadvantages provided by this pattern. The most frequent implementation of the MVC – the access to the data stored in the database through the web user interface is also described in this chapter. The next part of the thesis is concentrated on .NET Framework platform created from very voluminous, language-neutral library that is basically huge collection of the source codes providing the solution for common programmer’s problems and from the interface used for running up the applications created in .NET environment. The goal of the chapter concering with .NET Framework is to describe its architecture. The thesis also describes the platform invention, various versions of the .NET, the data access technology ADO.NET and the ASP.NET member ObjectDataSource. The chapter describing languages supported by the .NET Framework is focused on the C# language and its versions. The application “Multilayer applications in .NET environment” is the practical implementation of the mentioned technologies and it is described in the last chapter. The application’s architecture with the concentration on the particular Model-View-Controller layers implementation in the form of Microsoft Visual Studio 2005 projects is also described in the thesis. Special attention is paid to each operation over the data stored in the database tables that the application enables the user to execute, as for example data inserting, updating, selecting or deleting. The common business object’s child generation process is also described deep into the details.

The goal of this thesis is to design and test techniques for unsupervised adaptation of speech recognizers on some audio data without any textual transcripts. A training set is prepared at first, and a baseline speech recognition system is trained. This sistem is used to transcribe some unseen data. We will experiment with an adaptation data selection process based on some speech transcript quality measurement. The system is re-trained on this new set than, and the accuracy is evaluated. Then we experiment with the amount of adaptation data.

The way people acquire knowledge has largely shifted from print to web resources. Meanwhile, search has become the main medium to access information. Amongst various search behaviors, exploratory search represents a learning process that involves complex cognitive activities and knowledge acquisition. Research on exploratory search studies on how to make search systems help people seek information and develop intellectual skills. This research focuses on information retrieval and aims to build an exploratory search system that shows higher clustering performance and diversified search results. In this study, a new language model that integrates the state-of-the-art vector language model (i.e., BERT) with human knowledge is built to better understand and organize search results. The clustering performance of the new model (i.e., RDF+BERT) was similar to the original model but slight improvement was observed with conversational texts compared to the pre-trained language model and an exploratory search baseline. With the addition of the enrichment phase of expanding search results to related documents, the novel system also can display more diverse search results.

Social media and platforms utilise moderation to removeunwanted content such as cyberbullying, an aggressive acttowards an individual or group that occurs over any type ofdigital technology, e.g. social platforms. However,moderating platforms manually is nearly impossible, and thedemand for automatic moderation is rising. Research ontechnical solutions for cyberbullying detection on socialplatforms is scarce and is mostly focused on MachineLearning models to detect cyberbullying without theconnection to platform moderation. This study aims toenhance the research on cyberbullying detection models byusing a GPT-3 Large Language model and reduce the gap toplatform moderation. The model is tweaked and tested todetect cyberbullying using popular cyberbullying datasetsand compared to previous Machine Learning- and LargeLanguage models using common performance metrics.Furthermore, the latency of the model is measured to test if itcan be used as an auto-moderation tool to detectcyberbullying on social platforms. The results show that themodel is on par with the previous models and that finetuning a Large Language model is the preferred way totweak the model in cyberbullying detection. Further, theresults show that Large Language models have higherlatency than Machine Learning models but can be improvedby using multiple threads and can be used as a platformmoderation tool to detect cyberbullying.

Conventional language models estimate the probability that a word sequence within a chosen language will occur. By contrast, the purpose of our work is to estimate the probability that the word sequence belongs to the chosen language. The language of interest in our research is comprehensible well-formed English. We explain how conventional language models assume what we refer to as a degree of generalization, the extent to which a model generalizes from a given sequence. We explain why such an assumption may hinder estimation of the probability that a sequence belongs. We show that the probability that a word sequence belongs to a chosen language (represented by a given sequence) can be estimated by avoiding an assumed degree of generalization, and we introduce two methods for doing so: Minimal Number of Segments (MINS) and Segment Selection. We demonstrate that in some cases both MINS and Segment Selection perform better at distinguishing sequences that belong from those that do not than any other method we tested, including Good-Turing, interpolated modified Kneser-Ney, and the Sequence Memoizer.