Towards the Unsupervised Acquisition of Implicit Semantic Roles

Proceedings of Recent Advances in Natural Language Processing, pages 570–578, Hissar, Bulgaria, Sep 7–9 2015. Towards the Unsupervised Acquisition of Implicit Semantic Roles Niko Schenk, Christian Chiarcos, Maria Sukhareva Applied Computational Linguistics Lab Goethe University Frankfurt am Main, Germany {n.schenk,chiarcos,sukhareva}@em.uni-frankfurt.de Abstract This paper describes a novel approach to find evidence for implicit semantic roles. Our data-driven models generalize over large amounts of explicit annotations only, in order to acquire information about im- plicit roles. We establish a generic back- ground knowledge base of probablistic predicate-role co-occurrences in an unsu- pervised manner, and estimate thresholds which trigger the prediction of a missing role. Our approach outperforms the state- of-the-art in terms of recognition rate and offers a more flexible alternative to rule- based solutions which rely on costly, lan- guage and domain-specific lexica. 1 Introduction In its classical form, an automated semantic role labeling (SRL) system (Gildea and Jurafsky, 2002) detects events (verbal or nominal predicates), to- gether with their associated participants within the local context. Semantic roles are assigned to syn- tactic elements, such as A0 for the agent of an event, A1 for the patient (i.e. the entity which undergoes the action), etc.1 The output of SRL systems have proven to offer a good approxima- tion to a deeper semantic modeling of natural lan- guage. However, given its inherent complexity, re- cent efforts for improvement have tried to extend traditional SRL from the sentence-internal context to the surrounding discourse. As an illustration, consider the following biomedical example from Ruppenhofer et al. (2010). [A0Twenty-two month old] with history of re- current right middle lobe infiltrate. Increased [A0∅] cough, [A0∅] tachypnea, and [A0∅] work of breathing. 1For details on the PropBank labels used in our study, see Palmer et al. (2005). In the second sentence, a standard SRL system would ideally identify cough, tachypnea and work of breathing as nominal predicates. However, the A0 (experiencer/agent) role of these predicates is unfilled in the current sentence, and only explic- itly realized in the preceding one (cf. twenty-two month old). Its identification is thus beyond the scope of the traditional parser, which is restricted to an isolated per-sentence analysis. More precisely, the agent (argument) role of cough, for example, is non-overt or implicit, i.e. locally unexpressed in the second sentence and can only be resolved from the wider context. In general, many role realizations of this sort are sup- pressed on the surface level. These implicit roles are also called null instantiations (NIs) (Fillmore, 1986; Ruppenhofer, 2005) and have been exten- sively studied in the literature, cf. zero anaphora (Levinson, 1987; Hangyo et al., 2013). The automated detection of such implicit roles (iSRL) and their fillers is a challenging task. Yet, if uncovered, NIs provide highly beneficial ‘sup- plementary’ information: These can in turn be in- corporated into practical, downstream applications in the context of Natural Language Understand- ing, like text summarization, recognizing textual entailment or question answering. Current issues in iSRL Corpus data with man- ually annotated implicit roles is extremely sparse …

Content truncated to 500 words