Multimodal Word Sense Disambiguation in Creative Practice

摘要: Language is ambiguous; many terms and expressions can convey the same idea. This especially true in creative practice, where ideas design intents are highly subjective. We present a dataset, Ambiguous Descriptions of Art Images (ADARI), contemporary workpieces, which aims to provide foundational resource for subjective image description multimodal word disambiguation context practice. The dataset contains total 240k images labeled with 260k descriptive sentences. It additionally organized into sub-domains architecture, art, design, fashion, furniture, product technology. In description, labels not deterministic: example, ambiguous label dynamic might correspond hundreds different images. To understand this complexity, we analyze ambiguity relevance text respect using state-of-the-art pre-trained BERT model sentence classification. baseline multi-label classification tasks demonstrate potential approaches understanding intentions. hope that ADARI baselines constitute first step towards