A Study on Action Adverb Recognition on Korean Adverb Dataset

Abstract

This research introduces a specialized dataset for Korean action adverb recognition, particularly tailored for the cooking domain. Current studies in this field typically strive to either enhance the performance of recognition models or to compile novel datasets for action adverb recognition. Our dataset addresses the shortcomings of existing collections by focusing on Korean-labeled adverbs from culinary videos, thereby filling a void left by datasets that are largely derived from English content. These datasets not only lack linguistic diversity but also amalgamate disparate domains, leading to inconsistencies in how adverbs impact the interpretation of the same actions across different contexts. To bridge this gap, we meticulously analyzed contemporary models for action adverb recognition, applying them to our Korean dataset and juxtaposing the findings with those obtained from English datasets. The results were promising, indicating that our dataset not only holds its own in terms of performance when compared to its English counterparts but also underscores the importance of contextual and linguistic specificity in action adverb recognition. The study confirms the potential of our dataset in supporting the refinement and development of more nuanced recognition models, and in contributing to a more comprehensive understanding of natural language processing in Korean.