Leveraging Verb-Argument Structures to Infer Semantic Relations

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 145–154, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Leveraging Verb-Argument Structures to Infer Semantic Relations Eduardo Blanco and Dan Moldovan Lymba Corporation Richardson, TX 75080 USA {eduardo,moldovan}@lymba.com Abstract This paper presents a methodology to in- fer implicit semantic relations from verb- argument structures. An annotation effort shows implicit relations boost the amount of meaning explicitly encoded for verbs. Experimental results with automatically obtained parse trees and verb-argument structures demonstrate that inferring im- plicit relations is a doable task. 1 Introduction Automatic extraction of semantic relations is an important step towards capturing the meaning of text. Semantic relations explicitly encode links be- tween concepts. For example, in The accident left him a changed man, the ‘accident’ is the CAUSE of the man undergoing some ‘change’. A question answering system would benefit from detecting this relation when answering Why did he change? Extracting all semantic relations from text is a monumental task and is at the core of language understanding. In recent years, approaches that aim at extracting a subset of all relations have achieved great success. In particular, previous re- search (Carreras and M`arquez, 2005; Punyakanok et al., 2008; Che et al., 2010; Zapirain et al., 2010) focused on verb-argument structures, i.e., relations between a verb and its syntactic arguments. Prop- Bank (Palmer et al., 2005) is the corpus of refer- ence for verb-argument relations. However, rela- tions between a verb and its syntactic arguments are only a fraction of the relations present in texts. Consider the statement [Mr. Brown]NP1 suc- ceeds [Joseph W. Hibben, who retired last August]NP2 and its parse tree (Figure 1). Verb- argument relations encode that NP1 is the AGENT and NP2 is the THEME of verb ‘succeeds’ (Prop- Bank uses labels ARG0 and ARG1). Any se- mantic relation between ‘succeeds’ and concepts dominated in the parse tree by one of its syntac- tic arguments NP1 or NP2, e.g., ‘succeeds’ oc- S NP1 VP Mr. Brown VBZ NP2 succeeds AGENT THEME TIME-AFTER [Joseph W. Hibben, who]AGENT [retired]v [last August]TIME Figure 1: Example of parse tree and verb- argument structures (solid arrows). The relation between ‘succeeds’ and ‘last August’ is missing, but a TIME-AFTER holds (dashed arrow). curred after ‘last August’, are missing. Note that in this example, verb-argument structures encode that ‘retired’ has TIME ‘last August’, and this knowledge could be exploited to infer the miss- ing relation. The work presented here stems from two observations: (1) verbs are semantically con- nected with concepts that are not direct syntac- tic arguments (henceforth, implicit relations); and (2) verb-argument structures can be leveraged to infer implicit relations. This paper goes beyond verb-argument struc- tures and targets implicit relations like the one depicted above. TIME, LOCATION, MANNER, PURPOSE and CAUSE are inferred without im- posing syntactic restrictions between their argu- ments: systems trained over PropBank do not at- tempt to extract these relations. An annotation ef- fort demonstrates implicit relations reveal …

Content truncated to 500 words