Spatial information extraction, a field of information extraction, is a study that searches for toponyms expressed in a sentence, spatial nominals, relation information among spatial words, and spatially-related dynamic event information. While this study has a considerable potential for being used in a variety of application systems such as internet of things, geographic information system, and question answering system, its research on Korean language is still fairly new.
Spatial information extraction has been conducted with the development of various spatial information annotation schemes such as SpatialML (Spatial Markup Language), SpRL (Spatial Role Labeling), and ISOspace (ISO Space). The recent study is based on ISOspace, which is the ISO (International Standard Organization) standard for spatial information annotation. In addition, various spatial information extraction studies have been conducted under the name ‘SpaceEval’ in the shared task of CoNLL (Conference on Natural Language Learning). SpaceEval, which mainly targets English texts, extracts the information using various machine learning methods and rules.
Korean is a rich morphological language, and it has both features of agglutinative and inflected languages, which is significantly different from English. Moreover, because Korean has a relatively free word order, neighboring context words do not always have significant meanings like English where they play important roles in spatial word classification. Furthermore, as words in Korean are frequently omitted, and not much semantic resource exists, such as a semantic parser that provides word sense category information, Korean cannot use the same feature values as those in previous studies based on English. Therefore, this thesis proposes automatic extraction methods for spatial information in Korean texts defined in ISOspace considering Korean linguistic characteristics.
The spatial information extraction task can be divided into spatial entity and spatial relation extraction tasks. In Korean spatial entity extraction, we use CRFs (Conditional Random Fields), a machine learning model, which performed the best in SpaceEval. We have newly defined the feature values useful to spatial entity extraction by reflecting the characteristics of Korean such as word spacing, word phrase, and dependency relationships in a sentence. Further, we suggest methods that use multiple sub-models by grouping similar spatial entity tags, applying a down-sampling method to mitigate data skewness, and using word vector similarities to reduce over-generation. The suggested methods showed approximately 86% of the performance in F1-score.
Spatial relation extraction is a study that extracts spatial relationships in the triple entity form as <trajector, trigger, landmark> from the spatial entities given in the previous step. In the case of English texts, the rule-based study showed the best performance, which means the relation information is relatively easily acquired because of the rigid word order in English sentences. Since Korean has a relatively free word order and frequently omits the basic components that constitute the sentence, extracting relation information is more difficult compared to English. Further, applying machine learning models is not easy owing to its lack of feature values necessary to distinguish the correct relation information.
In this thesis, we have developed new rules to extract relation information and suggested a new probabilistic model that uses the result of dependency parsing. The first is a manually-written relation generating rules that reflect grammatical structure and lexical characteristics of Korean. In addition, after analyzing the spatial relation information appearing in the corpus, we have developed the rules by using pattern information. The second method is extracting information by using the Bayesian probability model, which uses the structure of dependency relation information, which solves the problems of free word order and a long distance between words involved in spatial relationship. The relation probability, as well as the role probability of each argument, is calculated by the transition probabilities of dependency parse tree labels between arguments involved in the spatial relation. In terms of the relation information extraction, the performance was recorded to be 46% in the F1-score. In addition, the Bayesian probability model that used dependency parse tree rather than the rule-based extraction method was shown to have a higher performance.
This thesis is the first study that extracts the ISOspace-based spatial information in Korean texts. By analyzing the characteristics of Korean, we suggested the effective methods to extract the spatial information in Korean texts and verified them experimentally.