Can Selectional Preferences Help Automatic Semantic Role Labeling?

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 222–227, Denver, Colorado, June 4–5, 2015. Can Selectional Preferences Help Automatic Semantic Role Labeling? Shumin Wu Department of Computer Science University of Colorado Boulder shumin@colorado.edu Martha Palmer Department of Linguistics University of Colorado Boulder mpalmer@colorado.edu Abstract We describe a topic model based approach for selectional preference. Using the topic fea- tures generated by an LDA model on the ex- tracted predicate-arguments over the Chinese Gigaword corpus, we show improvement to our state-of-the-art Chinese SRL system by 2.34 F1 points on arguments of nominal pred- icates, 0.40 F1 point on arguments of verb predicates, and 0.66 F1 point overall. More over, similar gains were achieved on out-of- genre test data, as well as on English SRL us- ing the same technique. 1 Introduction It’s long been theorized that selectional preferences (SP)/semantic constraints can improve automatic se- mantic role labeling (SRL). And while there have been several publications showing positive effects of SP, the evaluations have been dominated by pseudo- disambiguation. Zapirain et al. (2013) demonstrated end-to-end SRL improvement on arguments of En- glish verb predicates by using a combination of lex- ical resources and distributional similarity based SP. However, the margin of improvement is a modest 0.4 F1 point (on WSJ) over a baseline system with performance over 4 F1 points lower than the top sys- tem in CoNLL-2005 (Carreras and M`arquez, 2005). These results may not be convincing enough to mo- tivate the incorporation of SP when building an SRL system. One reason for the small improvement may be that arguments of a verb predicate are highly con- strained by the underlying syntactic parse, and SP features that could disambiguate between role types are often negated by parse errors. With the recent extension of PropBank SRL to nominal and adjec- tive predicates, preposition relationships, light-verb constructions, and abstract meaning representation (Bonial et al., 2014; Banarescu et al., 2013), it may be time to revisit SP for SRL. We hypothesize that SP will provide a greater benefit to nominal SRL, es- pecially on a language with lower parsing accuracy. In this paper, we apply SP to Chinese SRL (which has few morphological clues that impacts parsing accuracy) for arguments of both verb and nominal predicates using Chinese Gigaword. Our hypothe- sis, that SP will provide a greater benefit for nomi- nal predicates than for verbal predicates, is verified by our results. We achieve a 2.34 F1 point improve- ment to our Chinese SRL system on arguments of nominal predicates, 0.40 F1 point on arguments of verb predicates, and 0.66 F1 point overall. 2 Previous Work on Selectional Preference Inducing selectional preferences from corpus data was first proposed by Resnik (1997) for sense dis- ambiguation. He generalized seen words using the WordNet (Fellbaum, 1998) hierarchy. Gildea and Jurafsky (2002) applied SP to automatic SRL by clustering extracted verb-direct object pairs, result- ing in modest improvements. This syntactic signa- ture based selectional preference technique has also been successfully extended …

Content truncated to 500 words