Corpus annotation with a linguistic analysis of the associations between event mentions and spatial expressions

Corpus annotation with a linguistic analysis of the associations between event mentions and spatial expressions Jin-Woo Chung Jinseon You Jong C. Park* School of Computing Korea Advanced Institute of Science and Technology 291 Daehak-ro, Daejeon, Republic of Korea {jwchung, jsyou, park}@nlp.kaist.ac.kr Abstract Recognizing spatial information associated with events expressed in natural language text is essential for the proper interpretation of such events. However, the associations between events and spatial information found throughout the text have been much less studied than other types of spatial association as looked into in SpatialML and ISO-Space. In this paper, we present an annotation framework for the linguistic analysis of the associations between event mentions and spatial expressions in broadcast news articles. Based on the corpus annotation and analysis, we discuss which information should be included in the guidelines and what makes it difficult to achieve a high inter-annotator agreement. We also discuss possible improvements on the current corpus and annotation framework for insights into developing an automated system. 1 Introduction Every event is situated within some real-world space and textual descriptions that refer to events in documents also convey such spatial information. Such information is important not only for the interpretation of single events but also for the understanding of the relations among them. Spatial information can be used for various applications such as information extraction, textual entailment, and question answering. For instance, if we want to answer the question “Where did traffic accidents happen most frequently in 2014?,” we would need a * Corresponding author method to access and collect spatial information associated with all traffic accidents from the relevant textual descriptions. However, such information is usually not provided explicitly in text since humans can intuitively understand from the context where each event occurs. In general, two factors make it difficult to automatically recognize the location of events in text. First, there are usually more event mentions in text than expressions containing information about the location of events. A system must thus choose the most appropriate spatial expression for a given event mention. Second, such expressions are not always syntactically close to event mentions, which may make it less obvious to recognize their semantic association. The following three sentences exemplify different levels of difficulties in determining whether particular event mentions and spatial expressions are associated, i.e., whether a spatial expression refers to the space where an event occurred. (1) A fire [broke out]EVENT at [a refrigerated warehouse]SPACE yesterday. (2) A North Korean fishing vessel intruded 10 miles across [the Northern Limit Line]SPACE and South Korea’s Navy [fired]EVENT 6 warning shots. (3) He searched all over [the room]SPACE for his [missing]EVENT ring. Sentence (1) shows that the event mention and the spatial expression are syntactically connected in a single clause, which can probably be identified in a straightforward manner with a conventional semantic role labeler. Sentence (2) shows that they exist in the same sentence but not in the same clause, and that there must be some inference in order to PACLIC 29 535 …

Content truncated to 500 words