Neural Machine Translation - How machines learn to translate patent language: An overview, evaluation and tutorial
| Autor: | Lang, Christian |
|---|---|
| Jazyk: | angličtina |
| Rok vydání: | 2023 |
| Předmět: |
NMÜ
neuronale maschinelle Übersetzung MÜ maschinelle Übersetzung TQA translation quality assessment Patentübersetzung Englisch Japanisch statistische maschinelle Übersetzung SMÜ regelbasierte maschinelle Übersetzung RBMÜ AutorInnen-Schlagwörter (eng.): NMT neural machine translation MT machine translation patent translation English Japanese statistical machine translation SMT rule-based machine translation RBMT |
| Popis: | This work strives to be an easy to understand overview of how the current state-of-the-art in machine translation, neural machine translation (NMT), works. Using the example of patent translation, the thesis aims to both demystify the terms “AI” and “deep-learning”, that are often associated with NMT and aims to provide an accessible guide for translators and Translation Studies scholars to work with, create and understand their own NMT models. A theoretical foundation to MT is provided on which the work presents the creation and evaluation of five NMT models to determine the impact of data selection before model training. For this purpose, the five models were trained on five different datasets sorted by domain: A mixed dataset, an optics dataset, a dataset containing all domains but optics and two smaller versions of the mixed and optics-free dataset for parity in data quantity with the optics dataset. It was found that the network’s performance varied noticeably depending on how much and which data was used for training. While the common conception, that more data equals better results, held true in the automatic evaluation, it was shown that the domain specific training can help with improving results in the human evaluation. In fact, a large discrepancy between the automatic evaluation (BLEU score) and the human evaluation (based on SAE J2450) could be ovserved, with the worst performing model in the automatic metric having the best results in the human evaluation. Many of the problems can be attributed to the lack of extra-sentential context consideration in both the translation and evaluation. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|