BPE vs. Morphological Segmentation: A Case Study on Machine Translation of Four Polysynthetic Languages

Findings of the Association for Computational Linguistics: ACL 2022, pages 961 - 971 May 22-27, 2022 c⃝2022 Association for Computational Linguistics BPE vs. Morphological Segmentation: A Case Study on Machine Translation of Four Polysynthetic Languages Manuel Mager♢ Arturo Oncevay♡ Elisabeth Mager♯ Katharina Kann♠ Ngoc Thang Vu♢ ♢University of Stuttgart ♯Universidad Nacional Autónoma de México ♠University of Colorado Boulder ♡University of Edinburgh Abstract Morphologically-rich polysynthetic languages present a challenge for NLP systems due to data sparsity, and a common strategy to handle this issue is to apply subword segmentation. We investigate a wide variety of supervised and un- supervised morphological segmentation meth- ods for four polysynthetic languages: Nahu- atl, Raramuri, Shipibo-Konibo, and Wixarika. Then, we compare the morphologically in- spired segmentation methods against Byte-Pair Encodings (BPEs) as inputs for machine trans- lation (MT) when translating to and from Span- ish. We show that for all language pairs except for Nahuatl, an unsupervised morphological segmentation algorithm outperforms BPEs con- sistently and that, although supervised meth- ods achieve better segmentation scores, they under-perform in MT challenges. Finally, we contribute two new morphological segmenta- tion datasets for Raramuri and Shipibo-Konibo, and a parallel corpus for Raramuri–Spanish. 1 Introduction Polysynthetic languages are known because of their rich morphology, that encodes most parts of the semantics into verbs, leading to a high morpheme- per-word rate. The resulting combinations of mor- phemes and roots result in extreme type sparsity. Thus, polysynthetic languages represent a challeng- ing environment for NLP methods (Klavans, 2018). Subword segmentation has been a common method to reduce sparsity (Vania and Lopez, 2017). More- over, as these languages are mostly extremely low- resource (ELR), the challenge is even harder. Some of the reasons behind this is that most of them are endangered and spoken by minority groups (Mager et al., 2018; Littell et al., 2018). But what impact does morphological segmen- tation have on downstream tasks like machine translation (MT), when translating from or into fusional languages? Linguistically inspired seg- mentation was considered to be the best option to handle rich morphology (Koehn et al., 2005; Virpi- oja et al., 2007) until the appearance of Byte-Pair Encodings (BPEs; Sennrich et al., 2016) and has been adopted as the default segmentation technique. BPEs earned this status for its good results, unsu- pervised training and language independence. Sal- eva and Lignos (2021) show that there is no signifi- cant gain when using an unsupervised morphologi- cal segmentation for the input over BPEs when eval- uating those methods in moderate LR scenarios for Nepali–English and Kazakh–English, contradict- ing initial findings of Ataman and Federico (2018). However, how would BPEs perform for polysyn- thetic languages in ELR scenarios? Schwartz et al. (2020) compare BPE, with Morfessor (Smit et al., 2014) and Rule-Based morphological analyzers for medium resourced Inuktitut–English, and for the ELR Yupik–English and Guarani–Spanish. Their results show that BPEs outperform Morfessor and the morphological analyzer in all MT cases (but with better Language Modeling capabilities of mor- phological models over BPEs). However, most of these studies only …

Content truncated to 500 words