| Abstract [eng] |
Vision-language models have gained significant popularity in recent years because they can simultaneously address problems related to image and text analysis. They combine computer vision and natural language processing techniques to perform tasks such as image description, picture-based question-answering systems, and multi-modal search. Recently, these models have become increasingly important in developing advanced applications, such as autonomous vehicles, medical diagnostics, and content management. Many Vision-language models are adapted to the most popular languages, such as English, Spanish, and Chinese, but lack integration with less popular languages, like Lithuanian. This study analysed the effectiveness of various Vision-Language models, such as BLIP, Gemma3, Qwen, and others, using pre-prepared data collected from Lithuanian news portals. Thus, to expand the research data, the Flickr8k dataset was selected, and its captions were translated into Lithuanian. The research dataset consists of photos associated with news articles and their corresponding captions below each image. Given that many models cannot generate captions in Lithuanian, a study was conducted to translate captions from Lithuanian to English. Traditional evaluation metrics, such as BLEU, METEOR, ROUGE, BERTScore and Sentence-BERT were used to evaluate the research results. The results of the experimental investigation show that models trained with languages of smaller countries, such as Lithuania, can be sufficiently accurate. |
