Multilingual multimodal machine translation for Dravidian languages utilizing phonetic transcription

Multimodal machine translation is the task of translating from a source text into the target language using information from other modalities. Existing multimodal datasets have been restricted to only highly resourced languages. In addition to that, these datasets were collected by manual translation of English descriptions from the Flickr30K dataset. In this work, we introduce MMDravi, a Multilingual Multimodal dataset for under-resourced Dravidian languages. It comprises of 30,000 sentences which were created utilizing several machine translation outputs. Using data from MMDravi and a phonetic transcription of the corpus, we build an Multilingual Multimodal Neural Machine Translation system (MMNMT) for closely related Dravidian languages to take advantage of multilingual corpus and other modalities. We evaluate our translations generated by the proposed approach with human-annotated evaluation dataset in terms of BLEU, METEOR, and TER metrics. Relying on multilingual corpora, phonetic transcription, and image features, our approach improves the translation quality for the underresourced languages.