Automatically Tagging Constructions of Causation and Their Slot-Fillers

Automatically Tagging Constructions of Causation and Their Slot-Fillers Jesse Dunietz Computer Science Department Carnegie Mellon University Pittsburgh, PA 15213, USA jdunietz@cs.cmu.edu Lori Levin and Jaime Carbonell Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {lsl,jgc}@cs.cmu.edu Abstract This paper explores extending shallow seman- tic parsing beyond lexical-unit triggers, using causal relations as a test case. Semantic pars- ing becomes difficult in the face of the wide variety of linguistic realizations that causation can take on. We therefore base our approach on the concept of CONSTRUCTIONS from the linguistic paradigm known as CONSTRUCTION GRAMMAR (CxG). In CxG, a construction is a form/function pairing that can rely on arbi- trary linguistic and semantic features. Rather than codifying all aspects of each construc- tion’s form, as some attempts to employ CxG in NLP have done, we propose methods that offload that problem to machine learning. We describe two supervised approaches for tag- ging causal constructions and their arguments. Both approaches combine automatically in- duced pattern-matching rules with statistical classifiers that learn the subtler parameters of the constructions. Our results show that these approaches are promising: they significantly outperform na¨ıve baselines for both construc- tion recognition and cause and effect head matches. 1 Introduction Historically, shallow semantic parsing has focused on tagging predicates expressed by individual lexical units. While this paradigm has been fruitful, tying meaning to lexical units excludes some essential se- mantic relationships that cannot be captured in such a representation. One domain that highlights the problem is causal relations. Causation can be expressed in a tremen- 1. THIS BILL promotes consolidation and coopera- tion among regulatory agencies. 2. SUCH SWELLING can impede breathing. 3. WE DON’T HAVE MUCH TIME, so let’s move quickly. 4. She’s mad because I HID THE CAR KEYS. 5. He died from A BLOCKED ARTERY. 6. Making money is contingent on FINDING A GOOD-PAYING JOB. 7. THIS DECISION opens the way for much broader application of the law. 8. For market discipline to work, BANKS CANNOT EXPECT TO BE BAILED OUT. 9. Judy’s comments were SO OFFENSIVE that I left. Table 1: Examples of causal language, reflecting the anno- tation scheme described in §2.1 (with connectives in bold, CAUSES in small caps, and effects in italics). dous variety of linguistic forms (Wolff et al., 2005). As exemplified in Table 1, possibilities include verbs (1, 2), prepositions/conjunctions (3, 4, 5), adjectives (6), and much more complex expressions. Some of these trickier cases can be handled as idiomatic multi- word expressions (MWEs; 7). Others (8, 9), however, are more structured than typical MWEs: they depend on particular configurations of syntactic relations and slot-fillers, placing them closer to the grammatical end of the continuum of lexicon and grammar. This diversity presents a problem for most se- mantic parsers, which inherit the restrictions of the representational schemes they are based on. Many semantic annotation schemes limit themselves to the argument structures of particular word classes. For example, the Penn Discourse Treebank (PDTB; 117 Transactions of the Association for Computational Linguistics, …

Content truncated to 500 words