|
|
Silence in dialogue
Laichter, Štěpán ; Fernandez Rovira, Raquel (advisor) ; Dušek, Ondřej (referee)
Silence is an indispensable aspect of dialogue. The following thesis examines the silence in dialogue from a variety of perspectives. First, I provide a background on the historical development of theories of dialogue and the place of silence within them. Second, I conduct a study of the capacity of one of the most prominent contemporary language models, called the GPT-3, to model silence in dialogue. I fine-tune the model on a dataset based on movie subtitle data. I evaluate its performance on its capacity to infer the length of silence between subtitle pairs. The experiment proposes a method of fine-tuning the language model via silence encoded as character strings. The results show that GPT-3 fine-tuning can indeed improve the model's performance by inferring silence gaps between subtitle turns. Keywords: dialogue, silence, GPT-3, fine-tuning, language models I declare that I have written my diploma thesis independently and that I have properly cited all the sources and literature used, and that the work has not been used in the context of another university study or to obtain another or the same degree. vii
|
|
|
Ethnographic study of Role-playing in World of Warcraft
Dušek, Ondřej ; Švelch, Jaroslav (advisor) ; Reifová, Irena (referee)
This thesis explores the almost unexplored phenomenon of role-playing in the MMORPG video game genre, focusing specifically on role-playing in World of Warcraft (WoW). Its assumption is that although players of this genre overwhelmingly focus on the gameplay aspects of the game, role-playing in MMORPGs exists and takes place as an activity of the fringe culture of players. The aim of this thesis is to investigate this phenomenon in WoW, to find out how role- playing in the game takes place and what artificial reality role-playing participants create in the virtual world of the game, and also to define this phenomenon as a form of role-playing The theoretical part of the thesis is divided into three segments. The first segment deals with role-playing itself and the role-playing game (RPG) genre, which is superior to the MMORPG subgenre. The second segment delves into the known knowledge about role-playing in MMORPGs, and the third describes the examined game WoW. In order to fulfill the goals of this thesis and answer the research questions, ethnographic procedures are used in the research part, by which the author of this thesis infiltrates the virtual role-playing environment in WoW and actively participates in the events, which he further describes and analyzes. Based on the analysis of field...
|
|
|
Tackling Hallucinations in Chart Summarization
Obaid ul Islam, Saad ; Dušek, Ondřej (advisor) ; Rosa, Rudolf (referee)
Thesis Abstract Saad Obaid ul Islam Charles University, Saarland University Title Tackling Hallucinations in Chart Summarization Abstract Information visualizations like bar charts, line charts, and pie charts are a common way of communicating quantitative data. They are used to get important insights and make well informed decisions. Automatic Chart Summarization is the task to explain and summarize the key takeaways from the chart. Like other natural language generation (NLG) systems, chart summarization systems suffer from a phenomenon called halluci- nations. Hallucinations occur when the system generates text that is not grounded in the input. In this research work, we try to tackle the problem of hallucinations in chart summarization. Our analysis shows that a lot of additional information is present in the training data that leads to hallucinations during inference. We also found out that reducing long distance dependencies and addition of chart related information like title and legends improve the overall performance of the system. Furthermore, we propose a natural language inference (NLI) based method to clean the training data and show that our method produces faithful summaries. 1
|
|
|
Restoring and improving the technical quality of audio recordings using machine learning methods
Lechovský, Adam ; Peterek, Nino (advisor) ; Dušek, Ondřej (referee)
The goal of this thesis is to train an artificial neural network which will be able to improve the technical quality of audio recordings. To achieve this, three artificial audio distortions are used to train seven different deep neural networks on pairs of distorted and undistorted audio. The resulting 21 networks are then evaluated using a number of objective and subjective measures. In the end, the networks learned to remove artificial distortions very well, but they did not learn to improve the technical quality of undistorted inputs. 1
|
|
|
An Assistant for Conversational Search
Motlíček, Ondřej ; Dušek, Ondřej (advisor) ; Galuščáková, Petra (referee)
This thesis aims to improve the user experience for an existing information retrieval system in the domain of educational content by implementing a conversational search feature. Conversational search resembles a dialogue with an assistant - a librarian. The conversational search process involves an initial query gradually specified by clarifying questions. In addition to the conversational search feature, we propose further extensions: our own query expansion method based on an ontology and a facet recommendation approach. The conversational search feature is evaluated with real teachers, i.e. the prospective users of the system, showing that the new feature is a useful alternative to the standard search. 1
|
|
|
Theatre play generation
Schmidtová, Patrícia ; Dušek, Ondřej (advisor) ; Flek, Lucie (referee)
This thesis explores different ways of improving theatre play script gen- eration. To generate a theatre play script, we fine-tune the GPT-2 medium language model on a mixture of theatre plays, movies, and TV show scripts, training it to continue a human-written script start. As plain language-model generation is not sufficient for generating a coherent full-length theatre play, we propose a two step hierarchical generation approach: first generating a plot summary and then conditioning the script generation on the plot summary. We train the hierarchical model on a dataset which we created by aligning plot sum- maries to script scenes using dynamic programming. We also propose additional approaches to make the generated text more consistent (with respect to char- acters and contradictions). We compare our models to a strong vanilla GPT-2 XL baseline and achieve comparable performance with our models being smaller and faster. Moreover, our hierarchical generation approach allows generating a coherent full-length theatre play script under limited human supervision, which has also been verified in practice by theatre professionals within the THEaiTRE project. 1
|
|
|
Natural Language Correction With Focus on Czech
Náplava, Jakub ; Straka, Milan (advisor) ; Grundkiewicz, Roman (referee) ; Dušek, Ondřej (referee)
Natural language correction, a subfield of natural language processing (NLP), is the task of automatically correcting user errors in written texts. It includes, but is not lim- ited to, grammatical error correction, spelling error correction and diacritics restoration. During the course of the work on this thesis, we witnessed a great advance in this field, with the emergence of new approaches to correct user errors, new datasets and also new evaluation metrics. This thesis presents, in the form of a dissertation by publication, our contributions to this field. As Czech is the primary language of the thesis author, special focus was devoted to improving natural language correction in Czech. The main con- tributions are (1) the creation of the Grammar Error Correction Corpus for Czech that comprises multiple sources of noisy texts such as essays or online discussion posts, eval- uation of strong neural models on this dataset, and meta-evaluation of existing metrics, (2) the development of grammar error correction systems suited to scenarios in which only low amount of annotated data is available, and (3) the development of two state-of- the-art models and the creation of the new multilingual dataset comprising 12 languages for diacritics restoration. 1
|
|
|
English poetry generation with pretrained language models
John, Vojtěch ; Dušek, Ondřej (advisor) ; Musil, Tomáš (referee)
Title: English poetry generation with pretrained language models Author: Vojtěch John Department: Institute of Formal and Applied Linguistics Supervisor: Mgr. et Mgr. Ondřej Dušek, Ph.D., Institute of Formal and Applied Linguistics Abstract: In this thesis we deal with possible ways of English poetry generation independent on a concrete form of poetic form. We briefly discuss selected known attempts to English poetry generation with respect to their adaptability. Based on the already existing limerick generator LimGen we created a new English poetry generator called PGen, flexible with respect to rhyme and metric schemes and allowing usage of finetuned language models. We present PGen together with subjective evaluation of poems it generated (in 4 different poetic forms) by several human raters. Keywords: poetry generation, natural language generation, natural language processing, pretrained language models
|
|
|
Hloubková automatická analýza angličtiny
Dušek, Ondřej ; Hajič, Jan (advisor) ; Vidová Hladká, Barbora (referee)
This thesis contains an account of our studies of deep or semantic analysis of English, particularly as described using predicate-argument structure description. Our main goal is to create a system for automatic inference of semantic relations between predicates and arguments - semantic role labeling. We developed a framework for parallel processing of our experiments, integrating third-party machine learning tools and implementing well-known as well as novel procedures. We investigated the current approaches to the problem and proposed several improvements, such as new classi cation features, separate handling of adverbial modi ers or special treatment for rare predicates. Based on our research, we designed and implemented our own semantic analysis system, consisting of predicate disambiguation and argument classi cation subtasks. We evaluated our solution using the CoNLL 2009 Shared Task English corpus.
|
|
|
Protection of domain names
Dušek, Ondřej ; Horáček, Vít (advisor) ; Horáček, Tomáš (referee)
In the diploma work, there are examined domain names, which have become, due to a mass development of the Internet, an economically very valuable estate. Domain names lack, although being a generally known notion, any legal anchoring in the Czech system of law. The absence of legal regulation of domain names is not an anomaly of the Czech legal environment, but it can also be seen in the states with advanced legal culture. Domain names fulfill, in addition to being a very valuable commercial article, identification and/or promotional functions. Just for those reasons, the domain names need to be provided a corresponding legal protection. The introductory part of the diploma work, containing an introduction into the issues of domain names, addresses technical aspects of domain names and the method of their hierarchical structuring. Further, the diploma work contains a list of institutions which play, in the area of domain names, an important role for various reasons. The historical development of the institutions, their organization, and activity contents are briefly described. Then the legal nature of domain names is examined and a question whether a domain name can be considered to be a thing/item, right and/or any other asset value is answered. Furthermore, the diploma work contains a description...
|
guest ::
RSS feed.