Papers by Josef Jon
End-to-End Lexically Constrained Machine Translation for Morphologically Rich Languages (2021.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to enforce word forms in translations struggle to make them agree with the rest of the output. |
| Approach: | They propose to train neural machine translation models with lemmatized constraints to infer correct word inflection. |
| Outcome: | The proposed model reduces errors in translation of constrained terms in automatic and manual evaluations on English-Czech language pairs. |
GAATME: A Genetic Algorithm for Adversarial Translation Metrics Evaluation (2024.lrec-main)
Copied to clipboard
| Challenge: | a genetic algorithm generates adversarial translations for arbitrary metrics . produced translations score well in an arbitrary MT evaluation metric, despite serious, deliberately introduced errors. |
| Approach: | They propose a method for decoding translation candidates from a machine translation model via a genetic algorithm to generate adversarial translations to test and challenge MT evaluation metrics. |
| Outcome: | The proposed method scores very well in an arbitrary MT evaluation metric despite serious, deliberately introduced errors. |
Breeding Machine Translations: Evolutionary approach to survive and thrive in the world of automated evaluation (2023.acl-long)
Copied to clipboard
| Challenge: | a genetic algorithm (GA) based method improves MT quality and identifies weaknesses in evaluation metrics. |
| Approach: | They propose a genetic algorithm-based method for modifying n-best lists produced by a machine translation system using a fitness function. |
| Outcome: | The proposed method improves translation quality and identifies weaknesses in evaluation metrics. |
Thesis proposal: Are We Losing Textual Diversity to Natural Language Processing? (2026.eacl-srw)
Copied to clipboard
| Challenge: | Using Neural Machine Translation, we examine whether the algorithms used in NMT have inherent inductive biases that are beneficial for most types of inputs but might harm the processing of untypical texts. |
| Approach: | They propose to use a set of measures to quantify text diversity based on its statistical properties to determine whether NMT systems struggle with maintaining the diversity of such texts. |
| Outcome: | The proposed approaches maintain the diversity and complexity of language and allow for better global planning of the output generation. |