On Target Segmentation for Direct Speech Translation

On Target Segmentation for Direct Speech Translation Mattia A. Di Gangi∗ mdigangi@apptek.com AppTek GmbH, 52062 Aachen, Germany Marco Gaido mgaido@fbk.eu Matteo Negri negri@fbk.eu Marco Turchi turchi@fbk.eu Fondazione Bruno Kessler, Trento, Italy Abstract Recent studies on direct speech translation show continuous improvements by means of data augmentation techniques and bigger deep learning models. While these methods are helping to close the gap between this new approach and the more traditional cascaded one, there are many incongruities among different studies that make it difficult to assess the state of the art. Surpris- ingly, one point of discussion is the segmentation of the target text. Character-level segmenta- tion has been initially proposed to obtain an open vocabulary, but it results on long sequences and long training time. Then, subword-level segmentation became the state of the art in neural machine translation as it produces shorter sequences that reduce the training time, while be- ing superior to word-level models. As such, recent works on speech translation started using target subwords despite the initial use of characters and some recent claims of better results at the character level. In this work, we perform an extensive comparison of the two methods on three benchmarks covering 8 language directions and multilingual training. Subword-level segmentation compares favorably in all settings, outperforming its character-level counterpart in a range of 1 to 3 BLEU points. 1 Introduction The recent surge in direct (or end-to-end) speech translation (ST) research (B´erard et al., 2016; Weiss et al., 2017) has favored a blooming of models and data augmentation techniques that contributed to increase the translation quality scores. However, despite the evidence of a general improving field, the state of the art is still not clearly defined and some contrasting claims are present in literature. One aspect that still needs clarification is the segmentation of the words in the target sentences. Some studies followed the setting of Listen, Attend and Spell (Chan et al., 2016) for automatic speech recognition (ASR) and segmented the target text at character level (Weiss et al., 2017; B´erard et al., 2018; Di Gangi et al., 2019b). Other studies followed Jia et al. (2019) in using subword-level segmentation (Pino et al., 2019; Bahar et al., 2019a; Liu et al., 2019), motivated by the impractical training time when using a character-level segmentation on large datasets due to the increased sequence length. A research strand in neural machine translation (NMT) (Cherry et al., 2018; Kreutzer and Sokolov, 2018; Ataman et al., 2019), based on recurrent models, showed that character-level ∗The first author performed this work while he was a Ph.D. student at FBK. Proceedings of the 14th Conference of the Association for Machine Translation in the Americas October 6 - 9, 2020, Volume 1: MT Research Track Page 137 models can outperform subword-level models, but need architectural changes to make the com- putation more efficient and increase the capacity of the models. Indeed, while the subword-level models can store, to some extent, syntactic and semantic information of words in their embed- ding layers, …

Content truncated to 500 words