Cross-Domain Named Entity Recognition with Image-aware Contexts: Leveraging Image Captions and Chain-of-Thought

ACL ARR 2025 February Submission3989 Authors

Abstract: Cross-Domain Named entity recognition is a crucial task in natural language processing that helps extract meaningful entities from text when transferring across different domains. However current cross-domain NER methods are often limited in leverage heterogeneous information from other modalities, which limits the ability of cross-domain knowledge discovery and data mining, thereby constraining the application potential of large-scale information systems. To address these challenges, we propose a cross-domain NER method that utilizes image-aware contexts, consisting of Domain-specific Dynamic Image Captioning(DDC) and Cross-domain Reasoning Chain(CRC). DDC generates contextualized image captions by aligning the semantics of text and captions conditioned on textual domain cues. Then CRC identifies potential entities and classifies them using captions generated by DDC and chain-of-thought. Experimental results demonstrate that our method achieves a remarkable 6.23\% average F1 improvement across all tested domains. Particularly notable are the performance gains in the political and scientific domains, where our approach surpasses the best baseline model with F1-score increases of 8.22\% and 9.58\%.

Paper Type: Long

Research Area: Information Extraction

Research Area Keywords: named entity recognition and relation extraction, zero/few-shot extraction,

Contribution Types: Approaches to low-resource settings

Languages Studied: English

Submission Number: 3989