Impact of Domain-Adapted Multilingual Neural Machine Translation in the Medical Domain
Authors:
Miguel Rios,
Raluca-Maria Chereji,
Alina Secara,
Dragos Ciobanu
Abstract:
Multilingual Neural Machine Translation (MNMT) models leverage many language pairs during training to improve translation quality for low-resource languages by transferring knowledge from high-resource languages. We study the quality of a domain-adapted MNMT model in the medical domain for English-Romanian with automatic metrics and a human error typology annotation which includes terminology-spec…
▽ More
Multilingual Neural Machine Translation (MNMT) models leverage many language pairs during training to improve translation quality for low-resource languages by transferring knowledge from high-resource languages. We study the quality of a domain-adapted MNMT model in the medical domain for English-Romanian with automatic metrics and a human error typology annotation which includes terminology-specific error categories. We compare the out-of-domain MNMT with the in-domain adapted MNMT. The in-domain MNMT model outperforms the out-of-domain MNMT in all measured automatic metrics and produces fewer terminology errors.
△ Less
Submitted 5 December, 2022;
originally announced December 2022.
Dozens of virtual impactor orbits eliminated by the EURONEAR VIMP DECam data mining project
Authors:
O. Vaduvescu,
L. Curelaru,
M. Popescu,
B. Danila,
D. Ciobanu
Abstract:
Massive data mining of image archives observed with large etendue facilities represents a great opportunity for orbital amelioration of poorly known virtual impactor asteroids (VIs). There are more than 1000 VIs known today; most of them have very short observed arcs and many are considered lost as they became extremely faint soon after discovery. We aim to improve the orbits of VIs and eliminate…
▽ More
Massive data mining of image archives observed with large etendue facilities represents a great opportunity for orbital amelioration of poorly known virtual impactor asteroids (VIs). There are more than 1000 VIs known today; most of them have very short observed arcs and many are considered lost as they became extremely faint soon after discovery. We aim to improve the orbits of VIs and eliminate their status by data mining the existing image archives. Within the European Near Earth Asteroids Research (EURONEAR) project, we developed the Virtual Impactor search using Mega-Precovery (VIMP) software endowed with a very effective (fast and accurate) algorithm to predict apparitions of candidate pairs for subsequent guided human search. Considering a simple geometric model, the VIMP algorithm searches for any possible intersection in space and time between the positional uncertainty of any VI and the bounding sky projection of any image archive. We applied VIMP to mine the data of 451914 Blanco/DECam images observed between 12 Sep 2012 and 11 Jul 2019, identifying 212 VIs that possibly fall into 1286 candidate images leading to either precovery or recovery events. Following a careful search of candidate images, we recovered and measured 54 VIs in 183 DECam images. About 4000 impact orbits were eliminated from both lists, 27 VIs were removed from at least one list, while 14 objects were eliminated from both lists. The faintest detections were around V~24.0, while the majority fall between 21<V<23. The minimal orbital intersection distances remains constant for 67% detections, increasing for eight objects and decreasing for 10 objects. Most eliminated VIs had short initial arcs of less than 5 days. Some unexpected photometric discovery has emerged regarding the rotation period of 2018 DB, based on the close inspection of longer trailed VIs and the measurement of their fluxes along the trails.
△ Less
Submitted 1 September, 2020;
originally announced September 2020.
