TIKHONOV, M. Bilingual Dictionary Based Neural Machine Translation [online]. Brno: Vysoké učení technické v Brně. Fakulta informačních technologií. 2023.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Informace k zadání | - The thesis work meets the requirements of the specification, and even partly exceeds, as the experiments were conducted on 4 language pairs from various families. | ||
| Práce s literaturou | - The student's activity and knowledge regarding the current literature is good. | ||
| Aktivita během řešení, konzultace, komunikace | - The student consultation during the first semester was sparse as agreed apriori, but the student consulted regularly during the second semester with updates regarding the work. | ||
| Aktivita při dokončování | - The specified work was completed in time. The student worked on some extensions towards the end. | ||
| Publikační činnost, ocenění | - The student also pariticipated in "International Workshop in Spoken Language Translation (IWSLT'23)" low-resource speech translation from Marathi - Hindi, during March-April. - This has lead to acollaborative workshop publication that will appear in the proceedings of IWSLT'23 ACL. |
This is a well executed bachelor project in a very challenging setup. The student has demonstrated solid understanding of very recent literature in the field and was able to apply the knowledge in a novel way.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Náročnost zadání | The assigment asks the student to work at the bleeding edge of research of neural machine translation, on the rather challenging task of dealing with little parallel data. This would be a difficult assignment even at the Masters level. | ||
| Rozsah splnění požadavků zadání | |||
| Rozsah technické zprávy | |||
| Prezentační úroveň technické zprávy | 89 | The technical report imprints haste on the reader. While it may be acceptable for reviewing the wider context of NMT (Chapter 2), it hurts the quality badly in description of the core techniques used in this bachelor project (Chapter 3). The fast pace is kept in the experiements chapter, but it is does not prevent understanding there. On the contrary, the careful selection of baselines (and, where appropriate, even toplines) for individual experiments allows the reader to asses the achieved results and clearly demonstrates student's firm grasp of the ideas under test. A final comparison of different techniques, preferably on a held-out test set, is missing, but it is comforting to have all the intermediate results presented on the same validation data. | |
| Formální úprava technické zprávy | 85 | The technical report is written in plain English and it can be read front-to-back easily. On the other hand, the text could use more polishing, it is sometimes over-fragmented (esp. the early parts of Chapter 5), and weird verb forms are occasionally chosen ("it was decided", "I want to present", "we (!) did XY"). The uniform visual style of graphs is really helpful in following the technical results. | |
| Práce s literaturou | 90 | Student's own ideas are clearly distinguished from previous work. It is clear that the student has got very familiar with recent advances of a very dynamic field. The only room for improvement I have identified is the lack of references when claims are made relating the performance of NMT to that of (professional) human translators. | |
| Realizační výstup | 85 | The work correctly uses publicly available toolkits to do the heavy-lifting of the actual training of NMT systems. Student's own scripts are straightforward enough for other people to follow, replicate and possibly build upon the experiments conducted in the bachelor project. | |
| Využitelnost výsledků | The findings of this bachelor project do move the boundaries of the known, albeit not by much. I believe the thesis could be the core of a solid conference paper, esp. if improvements can be found for pretrained models. |
eVSKP id 146753