Optimized Long Context Translation Models for English-Icelandic translations (22.09)

  Authors

Jónsson, Haukur Páll ; Símonarson, Haukur Barri ; Ragnarsson, Pétur Orri ; Ingólfsdóttir, Svanhvít Lilja and Þorsteinsson, Vilhjálmur

  Item identifier

http://hdl.handle.net/20.500.12537/283

 Project URL

https://velthyding.is

 Demo URL

https://velthyding.is

 Date issued

2022-09-27

 Type

toolService

 Language(s)

English , Icelandic

 Description

ENGLISH: These models are optimized versions of the translation models released in http://hdl.handle.net/20.500.12537/278. Instead of the 24 layers used in the full model, they have been shrunk down to 7 layers. The computational resources required to run inference on the models is thus significantly less than using the original models. Performance is comparable to the original models when evaluated on general topics such as news, but for expert knowledge from the training data (e.g. EEA regulations) the original models are more capable. The models are capable of translating between English and Icelandic, in both directions. They are capable of translating several sentences at once and are robust to some input errors such as spelling errors. The models are based on the pretrained mBART25 model (http://hdl.handle.net/20.500.12537/125, https://arxiv.org/abs/2001.08210) and finetuned on bilingual EN-IS data and backtranslated data (including http://hdl.handle.net/20.500.12537/260). The full backtranslation data used includes texts from the following sources: The Icelandic Gigaword Corpus (Without sport) (IGC), The Icelandic Common Crawl Corpus (IC3), Student theses (skemman.is), Greynir News, Wikipedia, Icelandic sagas, Icelandic e-books, Books3, NewsCrawl, Wikipedia, EuroPARL, Reykjavik Grapevine, Iceland Review. The true parallel long context data used is from European Economic Area (EEA) regulations, document-level Icelandic Student Theses Abstracts corpus (IPAC), Stúdentablaðið (university student magazine), The report of the Special Investigation Commision (Rannsóknarnefnd Alþingis), The Bible and Jehovah’s witnesses corpus (JW300). Provided here are model files, a SentencePiece subword-tokenizing model and dictionary files for running the model locally along with scripts for translating sentences on the command line. We refer to the included README for instructions on running inference. ÍSLENSKA: Þessi líkön eru smækkaðar útgáfur af líkönunum sem má finna á http://hdl.handle.net/20.500.12537/278 . Upphaflegu líkönin eru með 24 lög en þessar útgáfur eru með 7 lög og eru skilvirkari í keyrslu. Frammistaða líkananna er á pari við þau upphaflegu fyrir almennan texta, svo sem í fréttum. Á sérhæfðari texta sem er að finna í þjálfunargögnunum standa þau sig verr, t.d. á evrópureglugerðum. Þessi líkön geta þýtt á milli ensku og íslensku. Líkönin geta þýtt margar málsgreinar í einu og eru þolin gagnvart villum og smávægilegu fráviki í inntaki. Líkönin eru áframþjálfuð þýðingarlíkön sem voru þjálfuð frá mBART25 líkaninu (http://hdl.handle.net/20.500.12537/125, https://arxiv.org/abs/2001.08210). Þjálfunargögin eru samhliða ensk-íslensk gögn ásamt bakþýðingum (m.a. http://hdl.handle.net/20.500.12537/260). Einmála gögn sem voru bakþýdd og nýtt í þjálfanir eru fengin úr: Risamálheildinni (án íþróttafrétta), Icelandic Common Crawl Corpus (IC3), ritgerðum af skemman.is, fréttum í fréttagrunni Greynis, Wikipedia, Íslendingasögunum, opnum íslenskum rafbókum, Books3, NewsCrawl, Wikipedia, EuroPARL, Reykjavik Grapevine, Iceland Review. Samhliða raungögn eru fengin upp úr European Economic Area (EEA) reglugerðum, samröðuðum útdráttum úr ritgerðum nemenda (IPAC), Stúdentablaðinu, Skýrslu Rannsóknarnefndar Alþingis, Biblíunni og samhliða málheild unna úr Varðturninum (JW300). Útgefin eru líkönin sjálf, orðflísunarlíkan og orðabók fyrir flísunina, ásamt skriptum til að keyra þýðingar frá skipanalínu. Nánari leiðbeiningar eru í README skjalinu.

 Publisher

Miðeind ehf

 Acknowledgement

Ministry of Education, Science and Culture

Project code: Back-translation data selection and filtering (V2b)

Project name: Language Technology for Icelandic 2019-2023

 Subject(s)

nmt machine translation translation

 Collection(s)

Clarin IS

Sýna fulla færslu atriðis