An Unsupervised Model for Instance Level Subcategorization Acquisition

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 278–289, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics An Unsupervised Model for Instance Level Subcategorization Acquisition Simon Baker Computer Laboratory University of Cambridge sb895@cam.ac.uk Roi Reichart Technion, IIT Haifa, Israel roiri@ie.technion.ac.il Anna Korhonen Computer Laboratory University of Cambridge alk23@cam.ac.uk Abstract Most existing systems for subcategoriza- tion frame (SCF) acquisition rely on su- pervised parsing and infer SCF distribu- tions at type, rather than instance level. These systems suffer from poor portability across domains and their benefit for NLP tasks that involve sentence-level process- ing is limited. We propose a new unsuper- vised, Markov Random Field-based model for SCF acquisition which is designed to address these problems. The system relies on supervised POS tagging rather than parsing, and is capable of learning SCFs at instance level. We perform eval- uation against gold standard data which shows that our system outperforms several supervised and type-level SCF baselines. We also conduct task-based evaluation in the context of verb similarity prediction, demonstrating that a vector space model based on our SCFs substantially outper- forms a lexical model and a model based on a supervised parser 1. 1 Introduction Subcategorization frame (SCF) acquisition in- volves identifying the arguments of a predicate and generalizing about its syntactic frames, where each frame specifies the syntactic type and number of arguments permitted by the predicate. For example, in sentences (1)-(3) the verb distin- guish takes three different frames, the difference between which is not evident when considering the phrase structure categorization: (1) Direct Transitive: [They]NP [distin- guished]VP [the mast]NP [of [ships on the horizon ]NP ]PP . 1The verb similarity dataset used for the evaluation of our model is publicly available at ie.technion.ac.il/∼roiri/. (2) Indirect Transitive: [They]NP [distin- guished]VP [between [me and you]ADVP ]PP . (3) Ditransitive: [They]NP [distinguished]VP [him]NP [from [the other boys]NP ]PP. As SCFs describe the syntactic realization of the verbal predicate-argument structure, they are highly valuable for a variety of NLP tasks. For example, verb subcategorization information has proven useful for tasks such as parsing (Carroll and Fang, 2004; Arun and Keller, 2005; Cholakov and van Noord, 2010), semantic role labeling (Bharati et al., 2005; Moschitti and Basili, 2005), verb clustering, (Schulte im Walde, 2006; Sun and Korhonen, 2011) and machine translation (hye Han et al., 2000; Hajiˇc et al., 2002; Weller et al., 2013). SCF induction is challenging. The argument- adjunct distinction is difficult even for humans, and is further complicated by the fact that both ar- guments and adjuncts can appear frequently in po- tential argument head positions (Korhonen et al., 2000). SCFs are also highly sensitive to domain variation so that both the frames themselves and their probabilities vary depending on the meaning and behavior of predicates in the domain in ques- tion (e.g. (Roland and Jurafsky, 1998; Lippincott et al., 2010; Rimell et al., 2013), Section 4). Because of the strong impact of domain vari- ation, SCF information is best acquired automat- ically. …

Content truncated to 500 words