Beyond Plain Spatial Knowledge: Determining Where Entities Are and Are Not Located, and For How Long

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1502–1512, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Beyond Plain Spatial Knowledge: Determining Where Entities Are and Are Not Located, and For How Long Alakananda Vempala and Eduardo Blanco Human Intelligence and Language Technologies Lab University of North Texas Denton, TX, 76203 AlakanandaVempala@my.unt.edu, eduardo.blanco@unt.edu Abstract This paper complements semantic role representations with spatial knowledge be- yond indicating plain locations. Namely, we extract where entities are (and are not) located, and for how long (seconds, hours, days, etc.). Crowdsourced annotations show that this additional knowledge is in- tuitive to humans and can be annotated by non-experts. Experimental results show that the task can be automated. 1 Introduction Extracting meaning from text is crucial for true text understanding and an important component of several natural language processing systems. Among many others, previous efforts have fo- cused on extracting causal relations (Bethard and Martin, 2008), semantic relations between nom- inals (Hendrickx et al., 2010), spatial relations (Kordjamshidi et al., 2011) and temporal relations (Pustejovsky et al., 2003; Chambers et al., 2014). In terms of corpora development and automated approaches, semantic roles are one of the most studied semantic representations (Toutanova et al., 2005; M`arquez et al., 2008). They have been proven useful for, among others, coreference reso- lution (Ponzetto and Strube, 2006) and question answering (Shen and Lapata, 2007). While se- mantic roles provide a useful semantic layer, they capture a portion of the meaning encoded in all but the simplest statements. Consider the sentence in Figure 1 and the semantic roles of drove (solid arrows). In addition to these roles, humans in- tuitively understand that (dashed arrow) (1) John was not located in Berlin before or during drove, (2) he was located in Berlin after drove for a short period of time (presumably, until he was done picking up the package, i.e., for a few minutes to John drove AGENT DESTINATION PURPOSE to Berlin to pick up a package Figure 1: Semantic roles (solid arrows) and addi- tional spatial knowledge (dashed arrow). an hour), and then left Berlin and thus (3) was not located there anymore. Some of this addi- tional spatial knowledge is inherent to the mo- tion verb drive: people cannot drive to the loca- tion where they are currently located, and they will be located at the destination of driving af- ter driving takes place. But determining for how long the agent of drive remains at the destina- tion depends on the arguments of drive: from [John]AGENT [drove]v [home]DESTINATION [after an exhausting work day]TIME, it is reasonable to be- lieve that John will be located at home overnight. This paper manipulates semantic roles in order to extract temporally-anchored spatial knowledge. We extract where entities are and are not located, and temporally anchor this information. Tempo- ral anchors indicate for how long something is (or is not) located somewhere, e.g., for 5 minutes be- fore (or after) an event. We target additional spa- tial …

Content truncated to 500 words