Extracting Spatial Entities and Relations in Korean Text

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2389–2396, Osaka, Japan, December 11-17 2016. Extracting Spatial Entities and Relations in Korean Text Bogyum Kim and Jae Sung Lee Dept. of Computer Science Chungbuk National University Chungdae-ro 1, Seowon-gu, Cheongju, 28644, Korea bogyum@cbnu.ac.kr, jasonlee@cbnu.ac.kr Abstract A spatial information extraction system retrieves spatial entities and their relationships for ge- ological searches and reasoning. Spatial information systems have been developed mainly for English text, e.g., through the SpaceEval competition. Some of the techniques are useful but not directly applicable to Korean text, because of linguistic differences and the lack of language resources. In this paper, we propose a Korean spatial entity extraction model and a spatial rela- tion extraction model; the spatial entity extraction model uses word vectors to alleviate the over generation and the spatial relation extraction model uses dependency parse labels to find the proper arguments in relations. Experiments with Korean text show that the two models are ef- fective for spatial information extraction. 1 Introduction A spatial information extraction system retrieves spatially related lexical items and their relationships and then provides the information in a normalized form. This information is used for geological searches and reasoning, and ultimately, for understanding natural language text. For example, from the spatial relations that A is on B and B is on C, a human can simply infer the fact that A is on C. A spatial information extraction system retrieves the relations ‘on (A, B)’ and ‘on (B, C)’ from the text; then, a reasoning program can infer the relation ‘on (A, C)’ from the relations. This enables the system to build a knowledge base with a compact size from text for many intelligent systems such as robot navigation and question-answering systems. A spatial information extraction task is usually carried out by two sub tasks: spatial entity extraction and spatial relation extraction. Because spatial entity extraction retrieves the entities to be used for spa- tial relationships, it is different from a place extraction task in named entity recognition systems (Lee et al. 2011), which only retrieves place-related entities. This implies that spatial entity extraction deals with all of the entities involved in spatial relations, such as trajectors, landmarks, and spatial signals. Moreover, spatial signals are usually articles or particles that do not have explicit arguments for spatial relations. Whether some entities or relations are extracted or not depends on the semantic roles in a sentence, which makes the task more complicated and challenging. Many spatial information extraction systems have been developed for English text, especially those developed through the competition at the SpaceEval conference (Pustejovsky et al. 2015). The applica- tion of the techniques directly to Korean text is not simple because of its different linguistic features. The Korean language is a morphologically rich and agglutinative language, which is very different from English (Kim et al. 2016). Moreover, because Korean has a relatively free word order and words are frequently omitted, the order of neighboring words does …

Content truncated to 500 words