Neural Machine Translation for Low-resource Languages: A Survey

Abstract

Neural Machine Translation (NMT) has seen a tremendous spurt of growth in the last twenty years and has already entered a mature phase. While considered the most widely used solution for Machine Translation, its performance on low-resource language pairs remains sub-optimal compared to the high-resource counterparts due to the unavailability of large parallel corpora. Therefore, the implementation of NMT techniques for low-resource language pairs has been receiving the spotlight recently, thus leading to substantial research on this topic. This paper presents a detailed survey of research advancements in low-resource language NMT (LRL-NMT) and quantitative analysis to identify the most popular techniques. We provide guidelines to select the possible NMT technique for a given LRL data setting based on our findings. We also present a holistic view of the LRL-NMT research landscape and provide recommendations to enhance the research efforts further.

Masakhane Web: A Machine Translation Platform for African Languages

“…Language and communication have been essential for people to get along and work together since the dawn of human civilization [1]. So, translating between different cultures has been very important to the growth of society, both economically and culturally [2,3].…”

“…Consequently, different methods were created to address the MT problem, with statistical machine translation (SMT) being one of the most well-known. However, the performance of the SMT system was largely impacted by the number of parallel sentence pairs available for training, and strong emphasis was placed on creating parallel datasets in addition to research on new MT methods [6]. The advancement of SMT systems was a turning point for many language pairs and made the task of machine translation more achievable [6].…”

“…However, high-resource language pairs (e.g., English, French, and Spanish) do not have dataset size concerns because researchers have created sufficient amounts of parallel databases over the years [10]. However, the need for a lot of parallel data is not realistic for many of the 7,000+ languages that are spoken around the world today, and more than 2,000 of these, or almost 30%, are found in Africa [11,12]. It is a continent reputed for its multiplicity of languages and is most linguistically diverse, which gives reason why many African languages are considered as low-resource languages (LRLs).…”

KLAPrompt: Infusing Semantic Knowledge into Pre-trained Language Models by Long-answer Prompt Learning

“…In recent years, pre-trained language models (PLMs) with external semantic knowledge have shown excellent performance on many natural language processing (NLP) tasks, including named entity recognition [1]- [4], relation extraction [5]- [8], and machine translation [9]- [12]. However, traditional approaches of introducing knowledge are mostly training from scratch, which is time-consuming and computationally expensive, making it infeasible for most users.…”

Leveraging machine translation for cross-lingual fine-grained cyberbullying classification amongst pre-adolescents

“…However, recent advances by neural machine translation (NMT) (Vaswani et al, 2013; Luong, Pham, and Manning, 2015; Lin et al, 2021) have shown impressive results. In addition, NMT survey studies by Wang et al (2021) and Ranathunga et al (2021) suggest that though there is the scope of improvements in NMT systems, development of tools and resources for low-resourced languages (LRLs) (Koehn and Knowles, 2017) have greatly improved. In this study, we leverage advances in NMT to automatically translate labelled user-generated content within the cyberbullying domain from Italian into English language – specifically the dataset by Sprugnoli et al (2018).…”