Accessing information in a preferred language and interaction format offers significant benefits, particularly for people with accessibility needs and those who do not speak dominant languages. However, underrepresented languages face a critical digital divide, as they lack substantial web presence, leading to poor search engine performance with content in these languages. This project seeks to address these challenges by developing retrieval techniques and datasets tailored to enhance the effectiveness of multimodal search systems for low-resource languages (LRLs).

Training of multimodal retrieval models requires large collections of data in both text and speech modalities and most of these resources are not readily available for LRLs. We are looking for PhD students who will investigate new techniques for creating multimodal datasets consisting of text, speech, audio and video for any group of LRLs including Celtic and African languages. A key aspect of the project will be on developing new techniques for creating resources and training multimodal retrieval models using resources and pre-trained Large Language Models of well-resourced languages.