Semi-Supervised Semantic Role Labeling

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 220–228, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Semi-Supervised Semantic Role Labeling Hagen F¨urstenau Dept. of Computational Linguistics Saarland University Saarbr¨ucken, Germany hagenf@coli.uni-saarland.de Mirella Lapata School of Informatics University of Edinburgh Edinburgh, UK mlap@inf.ed.ac.uk Abstract Large scale annotated corpora are pre- requisite to developing high-performance semantic role labeling systems. Unfor- tunately, such corpora are expensive to produce, limited in size, and may not be representative. Our work aims to reduce the annotation effort involved in creat- ing resources for semantic role labeling via semi-supervised learning. Our algo- rithm augments a small number of man- ually labeled instances with unlabeled ex- amples whose roles are inferred automat- ically via annotation projection. We for- mulate the projection task as a generaliza- tion of the linear assignment problem. We seek to find a role assignment in the un- labeled data such that the argument sim- ilarity between the labeled and unlabeled instances is maximized. Experimental re- sults on semantic role labeling show that the automatic annotations produced by our method improve performance over using hand-labeled instances alone. 1 Introduction Recent years have seen a growing interest in the task of automatically identifying and labeling the semantic roles conveyed by sentential constituents (Gildea and Jurafsky, 2002). This is partly due to its relevance for applications ranging from infor- mation extraction (Surdeanu et al., 2003; Mos- chitti et al., 2003) to question answering (Shen and Lapata, 2007), paraphrase identification (Pad´o and Erk, 2005), and the modeling of textual entailment relations (Tatu and Moldovan, 2005). Resources like FrameNet (Fillmore et al., 2003) and Prop- Bank (Palmer et al., 2005) have also facilitated the development of semantic role labeling methods by providing high-quality annotations for use in train- ing. Semantic role labelers are commonly devel- oped using a supervised learning paradigm1 where a classifier learns to predict role labels based on features extracted from annotated training data. Examples of the annotations provided FrameNet are given in (1). Here, the meaning ofin predicates (usually verbs, nouns, or adjectives) is conveyed by frames, schematic representations of situations. Semantic roles (or frame elements) are defined for each frame and correspond to salient entities present in the situation evoked by the pred- icate (or frame evoking element). Predicates with similar semantics instantiate the same frame and are attested with the same roles. In our exam- ple, the frame Cause harm has three core semantic roles, Agent, Victim, and Body part and can be in- stantiated with verbs such as punch, crush, slap, and injure. The frame may also be attested with non-core (peripheral) roles that are more generic and often shared across frames (see the roles De- gree, Reason, and Means, in (1c) and (1d)). (1) a. [Lee]Agent punched [John]Victim [in the eye]Body part. b. [A falling rock]Cause crushed [my ankle]Body part. c. [She]Agent slapped [him]Victim [hard]Degree [for his change of mood]Reason. d. [Rachel]Agent injured [her friend]Victim [by closing the car door …

Content truncated to 500 words