BanglaView: A Bangla Image Captioning Dataset

Image description is a more crucial part of the artificial intelligence world because it covers two important sectors, which are computer vision and natural language processing. As the proficiency of computers in interpreting visual data and converting it into textual forms has advanced, there has been a notable surge in research enthusiasm for endeavors such as automated image description in recent years. Although the majority of research efforts are concentrated on the English language within monolingual contexts, less attention has been directed towards resource-constrained languages like Bangla. This oversight primarily stems from the absence of standardized datasets for such languages. To overcome the limited availability of Bangla image captioning data, we propose BanglaView, a novel dataset inspired by Flickr30k. Here, English captions are translated into Bangla using Google Translate and refined by expert annotators. The BanglaView dataset contains a total of 31,783 images with 1,58,915 sentences. The total words, unique words, sentence length mean, and sentence length variance of the BanglaView dataset are 16,68,322, 26,348, 10.49, and 20.82. Each image in these datasets comes with five accompanying descriptions. Native Bangla speakers who are fluent in both Bangla and English painstakingly created the annotations. To maintain high standards, a team of professionals performed review and post-processing to ensure the dataset's quality.

图像描述在人工智能领域占据愈发关键的地位，因其覆盖了计算机视觉（computer vision）与自然语言处理（natural language processing）两大核心研究方向。随着计算机解析视觉数据并将其转换为文本形式的能力不断提升，近年来自动化图像描述等相关研究的热情显著高涨。尽管当前多数研究聚焦于单语种语境下的英语，但包括孟加拉语在内的低资源语言却未得到足够的关注，这一疏漏主要源于此类语言缺乏标准化数据集。为解决孟加拉语图像描述数据稀缺的问题，我们提出了BanglaView数据集——一款以Flickr30k为灵感构建的新型数据集。该数据集的英文标注先通过谷歌翻译（Google Translate）转换为孟加拉语，再由专业标注人员进行润色优化。BanglaView数据集共包含31783张图像，对应158915条语句。该数据集的总词数、独特词数、语句长度均值及语句长度方差分别为1668322、26348、10.49与20.82。每张图像均配有5条配套描述。该数据集的标注工作由兼具孟加拉语与英语流利能力的母语使用者精心完成，为保障数据集质量，专业团队还开展了审核与后期处理环节以确保其高标准品质。