Spatial information extraction is to identify spatial entities and their relations in a text. The relations include topology, distance, direction, and motion events relations. The information is useful for question answering, natural language user int...
Spatial information extraction is to identify spatial entities and their relations in a text. The relations include topology, distance, direction, and motion events relations. The information is useful for question answering, natural language user interfaces for robot commanding and navigation system, and so on.
Spatial information is usually extracted by using machine learning methods such as SVM(support vector machine) and CRFs(conditional random fields) which need spatial information annotated corpus for training. As the development of an annotated corpus costs a lot, the corpus is not usually available for the most languages in the early stage of research.
In this , we propose a rule based method to retrieve Korean spatial relations without using annotated corpus and specially focusing on the motion events relation extraction. It uses name entity recognition result, POS tagging result and dependency paring result. The extraction is done in three steps; general triple relations extraction, merging them into relations with multiple arguments, and selecting spatial relations among them with using a spatial word dictionary and semantic information. The extraction rules are derived by analyzing dependency parsing results and considering Korean language characteristics such as Josa (case makers) and long suffix.
Preliminary experiment was done for the 20 spatial information annotated sentences. The result of motion event relation extraction showed 28.00% recall and 26.92% precision, which cannot be compared directly with other research results because they are developed for static relation extractions. We hope that this research may be useful for the bootstrapping of spatial relation annotations and spatial relation machine learning for Korean language.
，韩语论文题目，韩语论文题目 |
