Overview and Results of MixMT Shared-Task at WMT 2022

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 806 - 811 December 7-8, 2022 ©2022 Association for Computational Linguistics Overview and Results of MixMT Shared-Task at WMT 2022 Vivek Srivastava TCS Research Pune, Maharashtra, India srivastava.vivek2@tcs.com Mayank Singh IIT Gandhinagar Gandhinagar, Gujarat, India singh.mayank@iitgn.ac.in Abstract In this paper, we present an overview of the WMT 2022 shared task on code-mixed ma- chine translation (MixMT). In this shared task, we hosted two code-mixed machine translation subtasks in the following settings: (i) monolin- gual to code-mixed translation and (ii) code- mixed to monolingual translation. In both the subtasks, we received registration and partici- pation from teams across the globe showing an interest and need to immediately address the challenges with machine translation involving code-mixed and low-resource languages. 1 Introduction Code-mixing (or code-switching) is an interesting manifestation of multilingualism in communities across the globe. Lately, we observe an uptick in the interest and efforts of the computational linguis- tic community to solve a multitude of challenges with code-mixed languages. Several interesting resources and computational models have been pro- posed for problems such as language identification (Barman et al., 2014; Thara and Poornachandran, 2021), text generation (Gupta et al., 2020; Rizvi et al., 2021), and sentiment analysis (Chakravarthi et al., 2020; Patwa et al., 2020). Machine translation which is an active area of re- search and development for monolingual languages is at the outset for code-mixed languages (Chen et al., 2022). In this shared task, we aim to explore the machine translation task involving a popular code-mixed language i.e., Hinglish (code-mixing of Hindi and English). Through both subtasks, we aim to address the challenges in building a real- world multilingual translation system involving code-mixed language as the source/target. Similar to the recent events on code-mixed lan- guages (Chen et al., 2022; Srivastava and Singh, 2021b; Patwa et al., 2020), the MixMT shared task has received participation and engagement with teams from across the globe. In total, we received registration from 38 teams. Throughout the com- petition, seven teams actively participated and sub- mitted their system for the development phase, test phase, and human evaluation phase. 2 MixMT: Code-mixed Machine Translation 2.1 The two subtasks In the MixMT shared task, we hosted two subtasks involving a code-mixed language i.e. Hinglish. A brief description of both subtasks is given below: 1. Monolingual to code-mixed machine transla- tion (M2CM): In this subtask, Hindi and En- glish are the two source languages and the target language is Hinglish. The source Hindi and En- glish sentences are translations of each other. The Hindi language sentences are written in the Devanagari script whereas the target Hinglish language text is written in the Roman script. 2. Code-mixed to monolingual machine transla- tion (CM2M): In this subtask, Hinglish is the source language and the target language is En- glish. Both the English and Hinglish text are written in Roman script. 2.2 Dataset Training datasets: We provide the following train- ing datasets for both subtasks: 1. M2CM: For this subtask, HinGE …

Content truncated to 500 words