Annotating Temporally-Anchored Spatial Knowledge by Leveraging Syntactic Dependencies

Annotating Temporally-Anchored Spatial Knowledge by Leveraging Syntactic Dependencies Alakananda Vempala and Eduardo Blanco Human Intelligence and Language Technologies Lab University of North Texas AlakanandaVempala@my.unt.edu, eduardo.blanco@unt.edu Abstract This paper presents a two-step methodology to annotate temporally-anchored spatial knowledge on top of OntoNotes. We first generate potential knowledge using syntactic dependencies, and then crowdsource annotations to validate the potential knowledge. The resulting annotations indicate how long entities are or are not located somewhere, and temporally anchor this information. Crowdsourcing experiments show that spatial inferences are ubiquitous and intuitive, and experimental results show that they can be done automatically. Keywords: Semantics, Information extraction, Spatial knowledge 1. Introduction Extracting spatial meaning from text is of utmost impor- tance in natural language understanding. Efforts focused on spatial meaning—both corpora development and auto- matic tools—have become popular. Existing approaches to extract spatial knowledge usually focus on extracting loca- tions of events, someone or something. For example, se- mantic role labeling (Palmer et al., 2005) determines who did what to whom, when and where, e.g., Thelma Gutier- rez [went]verb [inside the forensic laboratory where scien- tist are trying to solve this mystery]ARG4, where ARG4 indi- cates the END POINT of event went. Efforts targetting loca- tions of entities include geo-locating Twitter users (Liu and Inkpen, 2015), and pairing companies with the location of their headquarters (Mintz et al., 2009) e.g., [IBM’s]company headquarters in [New York]location. Determining the temporal span where the spatial knowl- edge holds is not extensively researched. From the sen- tence John parked Jamie’s car at the Highland Garage, we can infer that John and the car are certainly located at the Highland Garage minutes before and during parking, and that John will leave shortly after parking whereas the car will be at the garage for a few days but not months. We can also infer that Jamie will probably be at the Highland Garage at some point after parking to pick up his car. This paper presents (1) a two-step methodology to ex- tract temporally-anchored spatial knowledge by manipulat- ing syntactic dependencies, and a (2) crowdsourced corpus annotated with temporally-anchored spatial knowledge. The work presented here extends our previous work (Vem- pala and Blanco, 2016), which only manipulated semantic roles. We show that additional temporally-anchored spatial knowledge can be extracted by leveraging syntactic depen- dencies. We release a new corpus that annotates how long entities are and are not located somewhere, and temporally anchor this spatial information.1 2. Background We work on top of OntoNotes (Hovy et al., 2006) as it is a well known corpus with text from various domains. 1Available at https://alakanandav.bitbucket.io/ Ontonotes contains over 64,000 sentences. It annotates, among other linguistic information, part-of-speech tags, parse trees, named entities and co-reference chains. We use the CoNLL- 2011 Shared Task distribution (Pradhan et al., 2011), and transform the gold parse trees into syn- tactic dependencies using Stanford CoreNLP (Manning et al., 2014). De Marneffe and Manning (2008) present and exemplify the Stanford dependencies, and Weischedel and Brunstein (2005) the named entity types used …

Content truncated to 500 words