Unsupervised Induction of Semantic Roles within a Reconstruction-Error Minimization Framework

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1–10, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Unsupervised Induction of Semantic Roles within a Reconstruction-Error Minimization Framework Ivan Titov Ehsan Khoddam University of Amsterdam {titov|e.khoddammohammadi}@uva.nl Abstract We introduce a new approach to unsupervised estimation of feature-rich semantic role la- beling models. Our model consists of two components: (1) an encoding component: a semantic role labeling model which predicts roles given a rich set of syntactic and lexi- cal features; (2) a reconstruction component: a tensor factorization model which relies on roles to predict argument fillers. When the components are estimated jointly to minimize errors in argument reconstruction, the induced roles largely correspond to roles defined in an- notated resources. Our method performs on par with most accurate role induction methods on English and German, even though, unlike these previous approaches, we do not incorpo- rate any prior linguistic knowledge about the languages. 1 Introduction Shallow semantic representations, and semantic role labels in particular, have a long history in linguis- tics (Fillmore, 1968). More recently, with an emer- gence of large annotated resources such as Prop- Bank (Palmer et al., 2005) and FrameNet (Baker et al., 1998), automatic semantic role labeling (SRL) has attracted a lot of attention (Gildea and Jurafsky, 2002; Carreras and M`arquez, 2005; Surdeanu et al., 2008; Hajiˇc et al., 2009; Das et al., 2010). Semantic role representations encode the under- lying predicate-argument structure of sentences, or, more specifically, for every predicate in a sentence they identify a set of arguments and associate each argument with an underlying semantic role, such as an agent (an initiator or doer of the action) or a patient (an affected entity). Semantic roles have many potential applications in NLP and have been shown to benefit question answering (Shen and Lap- ata, 2007; Kaisser and Webber, 2007), textual entail- ment (Sammons et al., 2009), machine translation (Wu and Fung, 2009; Liu and Gildea, 2010; Wu et al., 2011; Gao and Vogel, 2011), and dialogue sys- tems (Basili et al., 2009; van der Plas et al., 2009), among others. Most current statistical approaches to SRL are su- pervised, requiring large quantities of human an- notated data to estimate model parameters. How- ever, such resources are expensive to create and only available for a small number of languages. More- over, when moved to a new domain (e.g., from news corpora to blogs or biomedical texts), the perfor- mance of these models tends to degrade substan- tially (Pradhan et al., 2008). The scarcity of an- notated data has motivated the research into unsu- pervised learning of semantic representations (Swier and Stevenson, 2004; Grenager and Manning, 2006; Lang and Lapata, 2010; Lang and Lapata, 2011a; Lang and Lapata, 2011b; Titov and Klementiev, 2012a; F¨urstenau and Rambow, 2012; Garg and Henderson, 2012). The existing methods have a number of serious shortcomings. First, they make very strong assumptions, for example, assuming that arguments are conditionally independent …

Content truncated to 500 words