MUCS@MixMT: IndicTrans-based Machine Translation for Hinglish Text

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1131 - 1135 December 7-8, 2022 ©2022 Association for Computational Linguistics MUCS@MixMT: indicTrans-based Machine Translation for Hinglish Text Asha Hegdea, Hosahalli Lakshmaiah Shashirekhab Department of Computer Science, Mangalore University, Mangalore, India {ahegdekasha,bhlsrekha}@gmail.com Abstract Code-mixing is the phenomena of mixing var- ious linguistic units such as paragraphs, sen- tences, phrases, words, etc., of two or more languages in any text. It is predominantly used to post the comments by social media users who know more than one language. Processing code-mixed text is challenging because of its complex characteristics and lack of tools that support such data. Developing efficient Ma- chine Translation (MT) systems for code-mixed text is challenging due to lack of code-mixed data. Further, existing MT systems developed to translate monolingual data are not portable to translate code-mixed text mainly due to the informal nature of code-mixed data. To address the MT challenges of code-mixed text, this pa- per describes the proposed MT models submit- ted by our team MUCS, to the Code-mixed Machine Translation (MixMT) shared task in the Workshop on Machine Translation (WMT) organized in connection with Empirical mod- els in Natural Language Processing (EMNLP) 2022. This shared task has two subtasks: i) sub- task 1 - to translate English sentences and their corresponding Hindi translations written in De- vanagari script into Hinglish (English-Hindi code-mixed text written in Latin script) text and ii) subtask 2 - to translate Hinglish text into English text. The proposed models that translate English text to Hinglish text and vice versa, comprise of i) transliterating Hinglish text from Latin to Devanagari script and vice versa, ii) pseudo translation generation using existing models, and iii) efficient target gen- eration by combining the pseudo translations along with the training data provided by the shared task organizers. The proposed models obtained 5th and 3rd rank with Recall-Oriented Under-study for Gisting Evaluation (ROUGE) scores of 0.35806 and 0.55453 for subtask 1 and subtask 2 respectively. 1 Introduction In linguistic terms, code-mixing is the practice of switching between two or more languages within or across sentences/words in any text (Joshi, 1982). Due to the widespread use of social media plat- forms like Twitter, Facebook, Reddit, etc., users are generating more and more code-mixed content. In Indian scenario, social media users usually blend English with their mother tongue or local language, for instance, English and Hindi, mainly for the technological limitations of computer keyboard or smartphone keypads to enter text in local languages. Further, as most of the text processing tasks are de- veloped for handling monolingual and formal text, informal and/or code-mixed text such as Hinglish is less explored. As the code-mixed text like Hinglish is increasing day by day, many applications such as MT, sentiment analysis, emotion analysis, etc., are also increasing. This has created a great demand for the tools and resources to process code-mixed data. Sample Hinglish text along with their Hindi and English translations are given in Table 1. In recent years, pre-trained transformer-based …

Content truncated to 500 words