Inducing Example-based Semantic Frames from a Massive Amount of Verb Uses

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 58–67, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Inducing Example-based Semantic Frames from a Massive Amount of Verb Uses Daisuke Kawahara† Daniel W. Peterson‡ Octavian Popescu§ Martha Palmer‡ †Kyoto University, Kyoto, Japan ‡University of Colorado at Boulder, Boulder, CO, USA §Fondazione Bruno Kessler, Trento, Italy dk@i.kyoto-u.ac.jp, {Daniel.W.Peterson, Martha.Palmer}@colorado.edu, popescu@fbk.eu Abstract We present an unsupervised method for in- ducing semantic frames from verb uses in giga-word corpora. Our semantic frames are verb-specific example-based frames that are distinguished according to their senses. We use the Chinese Restau- rant Process to automatically induce these frames from a massive amount of verb in- stances. In our experiments, we acquire broad-coverage semantic frames from two giga-word corpora, the larger comprising 20 billion words. Our experimental results indicate the effectiveness of our approach. 1 Introduction Semantic frames are indispensable knowledge for semantic analysis or text understanding. In the last decade, semantic frames, such as FrameNet (Baker et al., 1998) and PropBank (Palmer et al., 2005), have been manually elaborated. These resources are effectively exploited in many nat- ural language processing (NLP) tasks, includ- ing not only semantic parsing but also ma- chine translation (Boas, 2002), information ex- traction (Surdeanu et al., 2003), question answer- ing (Narayanan and Harabagiu, 2004), paraphrase acquisition (Ellsworth and Janin, 2007) and recog- nition of textual entailment (Burchardt and Frank, 2006). There have been many attempts to automati- cally acquire frame knowledge from raw corpora with the goal of either adding frequency informa- tion to an existing resource or of inducing simi- lar frames for other languages. Most of these ap- proaches, however, focus on syntactic frames, i.e., subcategorization frames (e.g., (Manning, 1993; Briscoe and Carroll, 1997; Korhonen et al., 2006; Lippincott et al., 2012; Reichart and Korhonen, 2013)). Since subcategorization frames represent argument patterns of verbs and are purely syn- tactic, expressions that have the same subcatego- rization frame can have different meanings (e.g., metaphors). Semantics-oriented NLP applications based on frames, such as paraphrase acquisition and machine translation, require consistency in the meaning of each frame, and thus these subcatego- rization frames are not suitable for these semantic tasks. Recently, there have been a few studies on au- tomatically acquiring semantic frames (Materna, 2012; Materna, 2013). Materna induced seman- tic frames (called LDA-Frames) from triples of (subject, verb, object) in the British National Corpus (BNC) based on Latent Dirichlet Allo- cation (LDA) and the Dirichlet Process. LDA- Frames capture limited linguistic phenomena of these triples, and are defined across verbs based on probabilistic topic distributions. This paper presents a method for automati- cally building verb-specific semantic frames from a large raw corpus. Our semantic frames are verb- specific like PropBank and semantically distin- guished. A frame has several syntactic case slots, each of which consists of words that are eligible to fill the slot. For example, let us show three seman- tic frames of the verb “observe”:1 observe:1 nsubj:{we, author, ...} dobj:{effect, …

Content truncated to 500 words