Machine translation

organization that conducts research and provides expertise

Translation Memory, database of previous translations

Factory Acceptance User Testing, product is tested by end-users at manufacturer's facility

CAT tool, provides translation memory, terminology and machine translation, finds out how much of the content is repetitive

simultaneous shipping, launch in all countries at the same time so pirating is not possible

content is: flawed (slang, source uttered by non-natives, typos), quirky (humor, references), massive , always different , pieced-together

to rebrand them, from translators to Intercultural & Linguistic Mediators/Professionals

syntactic, semantic, pragmatic strategies

rendering of source language content into target language in written form

editing and correcting machine translation output

examination of target language content carried out by translator

bilingual examination of target language content against source language content for its suitability for the agreed purpose

monolingual examination of target language content for its sustainability for the agreed purpose

examination of the revised target language content and applying corrections before printing

Globalization (e.g. market research), Internationalization (generalizing product so it can handle languages and cultural conventions), Localization (linguistical and cultural appropriateness) , Translation (source content to target content)

CAT Tool, puts text in segments on the left, right side is for translation, reminds you of how you already translated before

allow to work with different files
extract translatable content thanks to file filters
break up content into segments
update project translation memories when you confirm
allow to save terminological databases
allow quality assurance checks

translation unit identified by CAT tool according to language-specific segmentation rules

open source online CAT tool

xml format

translation of segment is found in bilingual file (xliff: XML Localization Interchange File Format)

not only exact match but also segment before and/or after matches

both content and formatting of new source segment are identical to segment that’s already stored in the TM

part of the new segment matches segment stored in TM

(usually) under 49%, similarity is not useful enough

source and target segments, metadata, layout info

during project: translating in CAT tool and confirming translation, before project: importing external TM, aligning source and target files, crawling web and aligning segments

iate, wipo (world intellectual property organization), RWS Trados Multiterm

automatic translation of text from one natural language to another using a computer application

Language Service Company

grammatical, interlingual transformational rules and bilingual dictionaries, created manually

Apertium Simpleton

source sentence parsed into constituents: part-of-speech-tagging, phrase identification

statistical MT, neural MT

table of probable correspondences (e.g. ich-I= 0.3. Katze-cat=0.7)
the more unlikely the correspondence the lower the score

Translation Model + Language Model + optional model like glossary

indicates how many words are in one segment

sentence structure still problematic, no highlighting of non-translated words

recurrent and transformer architecture (input ->encoders->decoders->output)

existing translations in form of parallel, sentence-aligned corpora

array of numbers

one hot vectors: one dimension (=number in vector) per word in the vocabulary, vector dimension 1 or 0, or vector representations with fixed dimensionality, vector size does not correspond to vocabulary size

vector representations of words/sentences, based on belief that similar words appear in similar contexts

statistical: phrase table (how probable is it that one word equals the other in the target language: Katze=cat=0.7), neural: representation learning, no probabilities

Moses

words are separated into letters and recombines them so most frequently occur together

no translation if similarity is too low

omission: not translating parts of sentence, addition/hallucination: : generating unrelated output words, inaccuracies more difficult to spot than in earlier approaches, less robust to errors in training, difficulties in dealing with names, elimination of hedging devices (=more affirmative or stronger negation)

NMT: new model to each language direction needed, rule-based: works in both directions

connections between neurons

determines which part of sentence is important for translating

difference between actual output and reference output

training method, in beginning system output is random because it has not learned yet, training algorithm updates network parameters to minimize difference

aligned at segment level (usually not longer than sentence) in UTF-8 TMT or tmx format

emergency approach, monolingual data is being translated by machine, flipped and added to training data

using multilingual data to train multilingual NMT systems

training data: 80%, validation set: selection of model that performs best, 10% test set: in the end unbiased estimation of performance 10%

one whole pass through the training data

when loss doesn't decrease anymore or starts increasing, metric (BLEU), amount of epochs, time

best loss, BLEU

Bilingual Evaluation Understudy, statistic for measuring the accuracy of machine translations compared to human translators, high score is good

trains NM engines

SMT

existing aligned corpora, manually aligned TMs, synthetic data, glossaries, automatic crawling

accuracy, validity (expected type?), consistency (contradicting?) , uniformity (same format?), completeness (enough data?)

correct encoding, correct file format, removing/ relabling segments, removing tags, fixing misalignments, removing long segment-pairs, anonymizing data

offers MT engine or creation of own MT

data quality, domain specificity

tool for linguistic resource quality assurance (QA)

GIGO , NMT quickly degraded by noise in training data, training and retraining NMT has environmental impact

open parallel corpora

MT researcher (produced good system?), language professionals (choosing system), potential MT users (which costs can I save)

slow, expensive, dependent on guidelines, more meaningful

compares MT to one or several reference translations, produces score at corpus level, accurate and correct translation can score low because they use different words than reference translation

gives score of how many edits are needed to get output segment to a reference translation segment

quick translation of a text to help users understand the general meaning without focusing on full accuracy or linguistic

when text is fluent we accept it and overlook errors

linguistic device to express uncertainty or politeness

little changes make big difference in output, syntactic vs. pragmatic translation, propagation of societal biases present in training data, offensive content, loss of coherence beyond segment level (Du and Sie to refere to same person), MT influencing general language, priming effect (seeing MT influences translation process, you can't unsee)

fairslator

new in-domain data, duplicate existing data to give it prominence in training set

comprehensibility and accuracy

comprehensibility, accuracy, style, syntax grammar and punctuation

identifying errors, reasonable speed, follow guidelines

controlled language, edit textspeak and pdfs

post-editing the input: look at output first, correct errors in source and see if they improved, not correcting everything in general

fix Optical Character Recognition errors, improve quality of original writing, standardize non-standard writing

over-editing, under-editing

increased productivity, ergonomics, accessibility and output naturalness

news aggregator for translation industry

risk of introducing errors e.g. homophones, accessibility issues e.g. not all languages supported, potential increase in cognitive load, more colloquial and informal translation choices

an automatic subtitling tool