Alignment verification to improve NMT translation towards highly inflectional languages with limited resources

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 1841–1851 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 1841 Alignment verification to improve NMT translation towards highly inflectional languages with limited resources George Tambouratzis ILSP, Athena R.C. 6 Artemidos Str., Maroussi, 15125, Greece giorg t@athenarc.gr Marina Vassiliou ILSP, Athena R.C. 6 Artemidos Str., Maroussi, 15125, Greece mvas@athenarc.gr Abstract The present article studies translation quality when limited training data is available to trans- late towards morphologically rich languages. The starting point is a neural MT system, used to train translation models with only publicly available parallel data. An initial analysis of the translation output has shown that quality is sub-optimal, mainly due to the insufficient amount of training data. To improve trans- lation, a hybridized solution is proposed, us- ing an ensemble of relatively simple NMT systems trained with different metrics, com- bined with an open source module designed for low-resource MT that measures the align- ment level. A quantitative analysis based on established metrics is complemented by a qual- itative analysis of translation results. These show that over multiple test sets, the proposed hybridized method confers improvements over (i) both the best individual NMT and (ii) the en- semble system provided in the Marian-NMT package. Improvements over Marian-NMT are in many cases statistically significant. 1 Introduction The state of the art in MT involves corpus-based systems developed with machine-learning meth- ods. These methods learn from corpora the models needed for translation. A key strength of this ap- proach is that the system is adapted specifically towards the data it is trained with. For many years, the most successful data-driven approaches were phrase-based and syntax-based Statistical MT (SMT; Koehn, 2009). However, lately Neural MT (NMT) based on the encoder- decoder architecture and the concept of attention (Sutskever et al., 2014; Bahdanau et al., 2016) has become very popular. Indeed, since 2015, in MT shared tasks (Cettolo et al., 2015; Bojar et al., 2015; Bojar et al., 2016) most top-performing systems have been NMT systems. This trend is confirmed in the most recent MT shared task (Barrault et al., 2019), where 80% of participating systems are of NMT type. Though NMT represents the state of the art for MT, specific weaknesses have been re- ported: • NMT performance suffers from the lack of data resources (Koehn and Knowles, 2017), giving lower translation performance, espe- cially when training with out-of-domain rather than in-domain data. • Recent advances in NMT models have been shown (Sennrich and Zhang, 2019) to allow good translations to be achieved with smaller parallel corpora of typically 105 sentences, though substantial improvements are achieved when the corpus size reaches 106 sentences. However, training sets of such sizes are not available for all languages. • Translation performance is affected by non- parallel texts and non-literal translations (Carpuat et al., 2017). • The integration of multiple algorithms into an NMT system does not necessarily improve translation (Denkowski and Neubig, 2017). • The time …

Content truncated to 500 words