Extracting Time Information from Korean Documents

Most of the documents or writings we see in our daily lives contain information about time, and it has been important to develop a model for extracting the time information from unstructued texts. As the time information appears differently in different languages, there have been studies that proposed language-specific models. This study aims to extract timex3 tags of time expresisons (e.g., May 30) and event tags of eventual expressions (e.g., go, expect) from Korean documents. The Korean dataset used in this study is manually annotated, and used to finetune a pre-trained language model for the extraction task of timex3 and event tags. By experiments, we found that the model gave better performance when it is finetued for extracting the two tags simultaneously. The model achieved 0.5563 of F1 score.