| Abstract [eng] |
Rapid technological progress is leading to growing amounts of digital information and the scale of its distribution. At the same time, at the level of state organisations, awareness of the importance of effective communication with the public is spreading worldwide, therefore, the need to present public information to the masses in plain language is growing. Advancing technologies also mean new opportunities for translators. Progress in the field of machine translation (MT) systems enables translators to cope with increasing translation volumes by performing post-editing of draft translations automatically generated by the MT system. As an alternative to popular MT systems (DeepL, Google Translate, etc.), relatively new large language models have begun to be studied, including the ChatGPT robot, which emerged only in 2022. This robot is characterised not only by the ability to translate, but also to summarise text and carry out instructions, therefore, scientists are studying its capabilities to perform not only machine translation, but also text simplification. The relevance of the analysis is based on the fact that, as the ideas of plain language continue to become more popular in Lithuania, the aim is to find out whether translators could perform automated text translation and simplification using the ChatGPT robot and obtain a text of sufficiently good quality suitable for post-editing. The novelty of the thesis is revealed in the chosen topic, as there are no scientific works investigating the use of major language models for translating texts into Lithuanian and at the same time simplifying those texts. The object of the thesis is the text translation and simplification capabilities of the ChatGPT robot. The aim of the thesis is to test several methods for translating and simplifying texts using the ChatGPT robot and to analyse the results obtained. To achieve the aim, the following three objectives are set: 1. to review the scientific literature on the possibilities of using large language models for text translation and simplification, plain language concept, methods for assessing the quality of machine translation and text readability; 2. to translate and simplify texts with the ChatGPT robot using 3 different coined prompts; 3. to assess the quality and readability of the resulting translations and compare the results. Main results: 1. After analysing the Automated Readability Index (ARI) evaluation results, which showed how a widely used AI application ChatGPT manages to simplify texts when presented with different prompts, it was found that a more detailed prompt which provides simplification rules worked best. ChatGPT can follow such simplification instructions as splitting sentences, using simpler vocabulary, omitting unnecessary words, and using simple tenses. However, even after simplification, the texts did not reach the 7th grade reading level indicated in the detailed prompt, although the reading level on average was lowered by 4.22 grades. This could also be related to the choice of the ARI score. Its limitation is the fact that the formula is created for English language, thus because of linguistic differences, the formula should be changed to fit the Lithuanian language better. 2. After the human-performed translation quality evaluation according to the MQM taxonomy adapted for this research, it was found that ChatGPT-generated translations have a lot of errors. The majority of mistakes was related to accuracy (mistranslations, omissions and additions). However, because the chatbot was instructed not only to translate, but also to simplify the texts, it is possible that a part of the identified errors could have appeared because of the task of text simplification. Admittedly, the double task of text translation and simplification provided to ChatGPT might have negatively influenced the quality of translations. However, considering that any machine translation output has to be post-edited by a translator, it would not be any different. And the translator would not have to manually simplify the translation. Looking holistically, it could be imagined that large language models, especially GPTs, are becoming one of the everyday tools for translators, especially when the simpler communication will become more widely expected in our societies. The thesis has the following parts: introduction, theoretical part, methodology and empirical part, conclusions, lists of references and information sources, and appendices. |
