Automatically Identifying Implicit Arguments to Improve Argument Linking and Coherence Modeling

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 306–316, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics Automatically Identifying Implicit Arguments to Improve Argument Linking and Coherence Modeling Michael Roth and Anette Frank Department of Computational Linguistics Heidelberg University, Germany {mroth,frank}@cl.uni-heidelberg.de Abstract Implicit arguments are a discourse-level phe- nomenon that has not been extensively stud- ied in semantic processing. One reason for this lies in the scarce amount of annotated data sets available. We argue that more data of this kind would be helpful to improve exist- ing approaches to linking implicit arguments in discourse and to enable more in-depth stud- ies of the phenomenon itself. In this paper, we present a range of studies that empirically val- idate this claim. Our contributions are three- fold: we present a heuristic approach to auto- matically identify implicit arguments and their antecedents by exploiting comparable texts; we show how the induced data can be used as training data for improving existing argument linking models; finally, we present a novel ap- proach to modeling local coherence that ex- tends previous approaches by taking into ac- count non-explicit entity references. 1 Introduction Semantic role labeling systems traditionally process text in a sentence-by-sentence fashion, construct- ing local structures of semantic meaning (Palmer et al., 2010). Information relevant to these structures, however, can be non-local in natural language texts (Palmer et al., 1986; Fillmore, 1986, inter alia). In this paper, we view instances of this phenomenon, also referred to as implicit arguments, as elements of discourse. In a coherent discourse, each utter- ance focuses on a salient set of entities, also called “foci” (Sidner, 1979) or “centers” (Joshi and Kuhn, 1979). According to the theory of Centering (Grosz et al., 1995), the salience of an entity in a discourse is reflected by linguistic factors such as choice of referring expression and syntactic form. Both ex- tremes of salience, i.e., contexts of referential conti- nuity (Brown, 1983) and irrelevance, can also be re- flected by the non-realization of an entity. Altough specific instances of non-realization, so-called zero anaphora, have been well-studied in discourse anal- ysis (Sag and Hankamer, 1984; Tanenhaus and Carl- son, 1990, inter alia), this phenomenon has widely been ignored in computational approaches to entity- based coherence modeling. It could, however, pro- vide an explanation for local coherence in cases that are not covered by current models of Centering (cf. Louis and Nenkova (2010)). In this work, we pro- pose a new model to predict whether realizing an argument contributes to local coherence in a given position in discourse. Example (1) shows a text frag- ment, in which argument realization is necessary in the first sentence but redundant in the second. (1) El Salvador is now the only Latin Ameri- can country which still has troops in [Iraq]. Nicaragua, Honduras and the Dominican Republic have withdrawn their troops [∅]. From a semantic processing perspective, a human reader can easily infer …

Content truncated to 500 words