WebNLG Challenge 2023: Domain Adaptive Machine Translation for Low-Resource Multilingual RDF-to-Text Generation (WebNLG 2023)

Proceedings of the Workshop on Multimodal, Multilingual Natural Language Generation and Multilingual WebNLG Challenge (MM-NLG 2023), pages 93–94 September 12, 2023 ©2023 Association for Computational Linguistics WebNLG Challenge 2023: Domain Adaptive Machine Translation for Low-Resource Multilingual RDF-to-Text Generation Kancharla Aditya Hari IIIT Hyderabad Bhavyajeet Singh IIIT Hyderabad Anubhav Sharma IIIT Hyderabad Vasudeva Varma IIIT Hyderabad Abstract This paper presents our submission to the WebNLG Challenge 2023 for generating text in several low-resource languages from RDF- triples. Our submission focuses on using ma- chine translation for generating texts in Irish, Maltese, Welsh and Russian. While a sim- ple and straightfoward approach, recent works have shown that using monolingual models for inference for multilingual tasks with the help of machine translation (translate-test) can out- perform multilingual models and training mul- tilingual models on machine-translated data (translate-train) through careful tuning of the MT component. Our results show that this ap- proach demonstrates competitive performance for this task even with limited data. 1 Introduction The multilingual RDF-to-Text generation task in- volves generating verbalizations of RDF triples in different languages. It represents an important problem, as it can allow for the rapid generation and enrichment of informative texts in different lan- guages by leveraging existing knowledge graphs and databases. This is of particular relevance for low-resource languages which are typically less represented. Recent methods for RDF-to-Text generation rely on neural methods, which require availability of high quality data for training. However, this is challenging for low-resource languages due to the lack of such data. Manual curation of such datasets requires expert human supervision, and is expen- sive and time-consuming. Thus, utilizing machine translated data available in different languages is naturally motivated. The translate-train method relies of translating the training data to the target language, while the translate-test method translate the test data to English and using a monolingual model for inference. Recent works have shown that through careful adaption of the MT component, the translate-test method can perform significantly better than previ- ously shown results, even outperforming translate- train and multilingual models across different clas- sification tasks (Artetxe et al., 2023). Our submis- sion investigates the performance of translate-test method for generation. In particular, we use do- main adaption to improve the MT component and generate high quality texts in the target language with very little training data. We first generate text in English using a monolingual, English generation model, and then translate the generated text to the target text using the MT component. 2 Experimental Setup 2.1 Pre-trained model We use the T5 model (Raffel et al., 2020) for RDF-to-Text generation in English. The task is thus modeled as a text-to-text generation task, with the input RDF triples serialized to strings. The small, 60M parameter variant of T5 is used. The model is finetuned on the training split of WebNLG 2020 challenge’s English dataset. 2.2 MT component For the MT component, we use the distilled variant of NLLB (Team et al., 2022) with 600M parame- ters. Beam search with a beam size of 5 …

Content truncated to 500 words