Zaninello and Birch (2020): augments the training data with MWE translations from dictionaries, back translating (Sennrich et al., 2016), target-side sentences with MWEs, or wrapping source MWEs with special tokens

Fadaee et al. (2018): prepends a special token in idiom source sentences
 
Gamallo and Garcia (2019): unsupervised translation of MWEs by composing cross-lingual word embeddings

Idioms in NMT 


Stanovsky et al., (2019), Fade et al., (2018): use word alignments to compare reference and hypothesis matches that translate source idioms

Zaninello and Birch (2020): use edit-distance to align words in hypothesis and reference matches and score aligned words with character-level matching

Shao et al., (2018)’s method estimates the frequency of such measures but requires creating language-specific hand-crafted lists


Targeted Evaluation 


Related work in Idioms in NMT and Targeted Evaluation. 


University of California, Berkeley

The translation of idiomatic expressions such as “under the weather” is a current problem in Neural Machine Translation. NMT models incorrectly translate idiomatic expressions’ meanings literally or word-by-word. 

Translation of Idiomatic Expressions 


The lack of targeted and automatic methods for quantifying these errors limits current research on translating idiomatic expressions. 


Obstacles for idiomatic expressions in NMT 


Global Metrics (e.g. BLEU) consider the full translation and obscure the effects of idiom translation


Global metrics for NMT of idiomatic expression 


Targeted evaluation isolates the idiom translation with word alignments or word edit distance, which measures idiom translation accuracy but doesn’t account for literal translation errors. Shao et al. (2018) proposed a method for estimating these errors’ frequency but it requires hand-crafted blocklists with words corresponding to literal translation errors. 


Learn Before

Related

Learn After