Entity Linking Korean Text: An Unsupervised Learning Approach using Semantic Relations

Proceedings of the 19th Conference on Computational Language Learning, pages 132–141, Beijing, China, July 30-31, 2015. c⃝2015 Association for Computational Linguistics Entity Linking Korean Text: An Unsupervised Learning Approach using Semantic Relations Youngsik Kim KAIST / Korea, The Republic of twilight@kaist.ac.kr Key-Sun Choi KAIST / Korea, The Republic of kschoi@kaist.edu Abstract Although entity linking is a widely re- searched topic, the same cannot be said for entity linking geared for languages other than English. Several limitations includ- ing syntactic features and the relative lack of resources prevent typical approaches to entity linking to be used as effectively for other languages in general. We describe an entity linking system that leverage se- mantic relations between entities within an existing knowledge base to learn and per- form entity linking using a minimal en- vironment consisting of a part-of-speech tagger. We measure the performance of our system against Korean Wikipedia ab- stract snippets, using the Korean DBpe- dia knowledge base for training. Based on these results, we argue both the feasibil- ity of our system and the possibility of ex- tending to other domains and languages in general. 1 Introduction A crucial step in creating the Web of Data is the process of extracting structured data, or RDF (Adida et al., 2012) from unstructured text. This step enables machines to read and understand unstructured Web pages that consist the major- ity of the Web. Three tasks play a part in ex- tracting RDF from unstructured text: Named en- tity recognition(NER), where strings representing named entities are extracted from the given text; entity linking(EL), where each named entity rec- ognized from NER is mapped to a appropriate re- source from a knowledge base; and relation ex- traction(Usbeck et al., 2014). Although entity linking is an extensively researched field, most research done is aimed primarily for the English language. Research about entity linking for lan- guages other than English have also been per- formed (Jakob et al., 2013), but most state-of-art entity linking systems are not fully language inde- pendent. The reason for this is two-fold. Firstly, most entity linking systems depend on an existing sys- tem to perform named entity recognition before- hand. For instance, the system proposed by Us- beck (2014) uses FOX (Speck et al., 2014) to perform named entity recognition as the starting point. The problem with this approach is that an existing named entity recognition system is re- quired, and thus the performance of entity linking is bound by the performance of the named entity recognition system. Named entity recognition sys- tems for English achieve high performance even by utilizing a simple dictionary-based approach augmented by part-of-speech annotations, but this approach does not work in all languages in gen- eral. CJK1 languages in particular are difficult to perform named entity recognition on because lan- guage traits such as capitalization and strict token separation by white-space do not exist. Other ap- proaches to named entity recognition such as sta- tistical models or projection models respectively require a large amount of annotated …

Content truncated to 500 words