Improving Machine Translation Formality Control with Weakly-Labelled Data Augmentation and Post Editing Strategies

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 351 - 360 May 26-27, 2022 c⃝2022 Association for Computational Linguistics Improving Machine Translation Formality Control with Weakly-Labelled Data Augmentation and Post Editing Strategies Danial Zhang*, Jiang Yu*, Pragati Verma*, Ashwinkumar Ganesan* & Sarah Campbell Alexa AI, Amazon {dyz, janyu, vpragati, gashwink, srh}@amazon.com Abstract This paper describes Amazon Alexa AI’s im- plementation for the IWSLT 2022 shared task on formality control. We focus on the uncon- strained and supervised task for en→hi (Hindi) and en→ja (Japanese) pairs where very lim- ited formality annotated data is available. We propose three simple yet effective post editing strategies namely, T-V conversion, utilizing a verb conjugator and seq2seq models in order to rewrite the translated phrases into formal or in- formal language. Considering nuances for for- mality and informality in different languages, our analysis shows that a language-specific post editing strategy achieves the best perfor- mance. To address the unique challenge of limited formality annotations, we further de- velop a formality classifier to perform weakly- labelled data augmentation which automati- cally generates synthetic formality labels from large parallel corpus. Empirical results on the IWSLT formality testset have shown that pro- posed system achieved significant improve- ments in terms of formality accuracy while re- taining BLEU score on-par with baseline. 1 Introduction Although neural machine translation (NMT) mod- els have achieved state-of-the-art results with high BLEU scores1, given a language pair, they are trained on generic parallel corpora that are ex- tracted from various open source datasets such as the Europarl corpus (Koehn; Iranzo-Sánchez et al., 2019). These datasets make an implicit assumption that there is a single translation in the target lan- guage to a sentence from the source language. But the style of the language generated, through which meaning is conveyed, is also important (Heylighen et al., 1999). Thus, there is a need to control certain attributes of the text generated in a target language such as politeness or formality. *Equal contribution. 1http://nlpprogress.com/english/machine_translation. html In this paper, we present our system for the IWSLT 2022 formality control task for machine translation.2 We focus on the unconstrained and su- pervised scenario for en→hi and en→ja language pairs. In the proposed system, we explore post edit- ing strategies that correct or alter textual formality once the translation has been completed. Post edit- ing strategies can be language specific or language agnostic. We propose three strategies, T-V conver- sion (deterministically converting the informal or T-form of a pronoun to its corresponding formal or V-form) , verb conjugation, and a seq2seq model that learns to transform input text to be of a for- mal or informal nature. The T-V conversion and verb conjugation are language-specific strategies that are applied to en→hi, and en→ja pairs respec- tively. These two methods are compared against an alternative seq2seq model (Enarvi et al., 2020) that is language agnostic. We show that compared to a baseline translation model provided in task, a finetuned mBART model (Liu et al., 2020) …

Content truncated to 500 words