Improving Machine Translation in the E-commerce Luxury Space. A case study

Improving Machine Translation in the E-commerce Luxury Space. A case study Jos´e-Manuel de-la-Torre-Vilari˜no Acclaro jvilarino@acclaro.com Juan-Luis Garc´ıa-Mendoza Universit´e Sorbonne Paris Nord garciamendoza@lipn.univ-paris13.fr Alessia Petrucci Acclaro apetrucci@acclaro.com Abstract This case study presents a Multilingual e- commerce Project, which principal aim is to create an improved system that trans- lates product titles and descriptions, plus other content in multiple languages. The project consisted of two main phases; a research-intensive solution using state-of- the-art Machine Translation systems and baseline language models for two lan- guage pairs, and the development of a Ma- chine Translation system. The features implemented included Quality Estimation, model benchmarking, entity recognizers, and automatic domain detection. mBART model was used to create the system for the specific domain of e-commerce, for luxury items. 1 Introduction Machine Translation (MT) is the automatic trans- lation of text from one language to another with- out human intervention (Stahlberg, 2020). When this translation is performed using Deep Neural Networks (DNN), it is referred to as Neural Ma- chine Translation (NMT) (Stahlberg, 2020). NMT technology has made significant progress in recent years, however, they have generally been trained on domain-general data, directly affecting domain- specific translations (Martins et al., 2022; Saun- ders, 2022). According to Martins et al. (2022), most methods for adapting MT systems to a spe- cific domain focus on fine-tuning. One of those specific domains is e-commerce. In today’s globalized world, e-commerce has be- © 2023 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. come a crucial part of the economy. With the rise of online shopping, businesses must be able to communicate with customers in their native lan- guages to provide a seamless shopping experience. However, translating product titles and descrip- tions, reviews, and other content while maintain- ing formal and informal styles, and dealing with lengthy and very short sentences, can be a daunt- ing task. In this work, an e-commerce Multilingual Project aimed to improve machine translation quality is introduced. This improvement was car- ried out with mBART model (Liu et al., 2020), which is leading the way in Multilingual Transla- tion. This model is designed to handle multiple languages simultaneously, making it ideal for e- commerce applications where content needs to be translated quickly and accurately. The project has been led by Acclaro1, a leading company with ex- tensive experience in professional translation ser- vices. 2 Arquitectures In the state-of-the-art (SoTA) of NMT, the archi- tectures proposed by Radford et al. (2019), Liu et al. (2020) and Tang et al. (2020) stand out. On the one hand, GPT-2 model proposed by Radford et al. (2019) is based on the architecture of large transformers (Vaswani et al., 2017). Besides, GPT- 2 follows the details of the OpenAI GPT model proposed by Radford et al. (2018). On the other hand, according to Liu et al. (2020), mBART is “a multilingual sequence-to-sequence denoising auto-encoder” that uses BART (Lewis et al., 2020) large-scale monolingual corpora across many lan- guages. This model was …

Content truncated to 500 words