Abstract
We present an automated method for generating concreteness ratings that achieves beyond human-level reliability across multiple languages and expression types. Our approach combines multimodal transformers with emotion-finetuned language models and achieves correlations of 0.93 for single British words and 0.85 for multiword expressions with existing corpora of human raters. We demonstrate general applicability through successful cross-lingual generalization to an entirely unseen corpus of Estonian single- and multi-word expressions (N = 35,979), achieved via automated language detection and translation. By leveraging both visual and emotional information in context-aware language embeddings, our method effectively captures the full spectrum from concrete to abstract concepts. Our automated system offers a context sensitive, reliable alternative to traditional human ratings, eliminating the need for time-consuming and costly human rating collection. We provide an easy to access web-based interface for research to use our tool under concreteness.eu .