Unsupervised Induction of Semantic Roles

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 939–947, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Unsupervised Induction of Semantic Roles Joel Lang and Mirella Lapata School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB, UK J.Lang-3@sms.ed.ac.uk, mlap@inf.ed.ac.uk Abstract Datasets annotated with semantic roles are an important prerequisite to developing high- performance role labeling systems. Unfortu- nately, the reliance on manual annotations, which are both difficult and highly expen- sive to produce, presents a major obstacle to the widespread application of these systems across different languages and text genres. In this paper we describe a method for induc- ing the semantic roles of verbal arguments di- rectly from unannotated text. We formulate the role induction problem as one of detecting alternations and finding a canonical syntactic form for them. Both steps are implemented in a novel probabilistic model, a latent-variable variant of the logistic classifier. Our method increases the purity of the induced role clus- ters by a wide margin over a strong baseline. 1 Introduction Semantic role labeling (SRL, Gildea and Jurafsky 2002) is the task of automatically classifying the ar- guments of a predicate with roles such as Agent, Pa- tient or Location. These labels capture aspects of the semantics of the relationship between the predicate and the argument while abstracting over surface syn- tactic configurations. SRL has received much atten- tion in recent years (Surdeanu et al., 2008; M`arquez et al., 2008), partly because of its potential to im- prove applications that require broad coverage se- mantic processing. Examples include information extraction (Surdeanu et al., 2003), question answer- ing (Shen and Lapata, 2007), summarization (Melli et al., 2005), and machine translation (Wu and Fung, 2009). Given sentences (1-a) and (1-b) as input, an SRL system would have to identify the verb predicate (shown in boldface), its arguments (Michael and sandwich) and label them with semantic roles (Agent and Patient). (1) a. [Michael]Agent eats [a sandwich]Patient. b. [A sandwich]Patient is eaten [by Michael]Agent. Here, sentence (1-b) is an alternation of (1-a). The verbal arguments bear the same semantic role, even though they appear in different syntactic posi- tions: sandwich is the object of eat in sentence (1-a) and its subject in (1-b) but it is in both instances as- signed the role Patient. The example illustrates the passive alternation. The latter is merely one type of alternation, many others exist (Levin, 1993), and their computational treatment is one of the main challenges faced by semantic role labelers. Most SRL systems to date conceptualize semantic role labeling as a supervised learning problem and rely on role-annotated data for model training. Prop- Bank (Palmer et al., 2005) has been widely used for the development of semantic role labelers as well as FrameNet (Fillmore et al., 2003). Under the Prop- Bank annotation framework (which we will assume throughout this paper) each predicate is associated with a set of core roles (named A0, A1, A2, and so on) …

Content truncated to 500 words