Automatic Acquisition of the Argument-Predicate Relations from a Frame-Annotated Corpus

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1388–1397, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Automatic Acquisition of the Argument-Predicate Relations from a Frame-Annotated Corpus Ekaterina Ovchinnikova University of Osnabr¨uck eovchinn@uos.de Theodore Alexandrov University of Bremen theodore@ math.uni-bremen.de Tonio Wandmacher University of Osnabr¨uck twandmac@uos.de Abstract This paper presents an approach to au- tomatic acquisition of the argument- predicate relations from a semantically annotated corpus. We use SALSA, a German newspaper corpus manually an- notated with role-semantic information based on frame semantics. Since the rel- atively small size of SALSA does not al- low to estimate the semantic relatedness in the extracted argument-predicate pairs, we use a larger corpus for ranking. Two experiments have been performed in or- der to evaluate the proposed approach. In the first experiment we compare au- tomatically extracted argument-predicate relations with the gold standard formed from associations provided by human sub- jects. In the second experiment we cal- culate correlation between automatic relat- edness measure and human ranking of the extracted relations. 1 Introduction There are many debates in lexical semantics about what kind of world knowledge actually belongs to the meaning of a lexeme. Nowadays, it is widely accepted that predicates impose selectional restrictions on their arguments. For example, since we know that the predicate to be hungry mainly takes expressions describing animate beings as ar- guments, we can correctly resolve the anaphora in the following sentence: We gave the bananas to the monkeys because they were hungry. There exists also multiple linguistic evidence showing that the semantics of arguments can help to pre- dict implicit predicates. For example, the sentence John finished the cigarette usually means John fin- ished smoking the cigarette because the meaning of the noun cigarette is strongly associated with the smoking activity. It has been claimed that information about pred- icates associated with nouns can be helpful for a variety of tasks in natural language processing (NLP), see for example (Pustejovsky et al., 1993; Voorhees, 1994). However, at present there exists no corresponding lexical semantic resource. Sev- eral approaches have been presented that aim at creating a knowledge base containing noun-verb relations. There are two main research paradigms for developing such knowledge bases. The first paradigm assumes manual development of the re- source (Pustejovsky et al., 2006), while the sec- ond one relies on automatic acquisition methods, see for example (Cimiano and Wenderoth, 2007). In this paper we propose a procedure for auto- matic acquisition of argument-predicate relations from a semantically annotated corpus. In line with (Lapata and Lascarides, 2003) our approach is based on the assumption that predicates are omit- ted in a discourse when they are highly predictable from the semantics of their arguments. We exploit SALSA (Burchardt et al., 2006), a German news- paper corpus manually annotated with FrameNet frames based on frame semantics. Using a man- ually annotated corpus for relation extraction has one particular advantage compared to extraction from plain text: the type of an argument-predicate relation is …

Content truncated to 500 words