Graph Alignment for Semi-Supervised Semantic Role Labeling

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 11–20, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Graph Alignment for 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 Unknown lexical items present a major obstacle to the development of broad- coverage semantic role labeling systems. We address this problem with a semi- supervised learning approach which ac- quires training instances for unseen verbs from an unlabeled corpus. Our method re- lies on the hypothesis that unknown lexical items will be structurally and semantically similar to known items for which annota- tions are available. Accordingly, we rep- resent known and unknown sentences as graphs, formalize the search for the most similar verb as a graph alignment prob- lem and solve the optimization using inte- ger linear programming. Experimental re- sults show that role labeling performance for unknown lexical items improves with training data produced automatically by our method. 1 Introduction Semantic role labeling, the task of automatically identifying the semantic roles conveyed by sen- tential constituents, has recently attracted much at- tention in the literature. The ability to express the relations between predicates and their arguments while abstracting over surface syntactic configu- rations holds promise for many applications that require broad coverage semantic processing. Ex- amples include information extraction (Surdeanu et al., 2003), question answering (Narayanan and Harabagiu, 2004), machine translation (Boas, 2005), and summarization (Melli et al., 2005). Much progress in the area of semantic role la- beling is due to the creation of resources like FrameNet (Fillmore et al., 2003), which document the surface realization of semantic roles in real world corpora. Such data is paramount for de- veloping semantic role labelers which are usually based on supervised learning techniques and thus require training on role-annotated data. Examples of the training instances provided in FrameNet are given below: (1) a. If [you]Agent [carelessly]Manner chance going back there, you deserve what you get. b. Only [one winner]Buyer purchased [the paintings]Goods c. [Rachel]Agent injured [her friend]Victim [by closing the car door on his left hand]Means. Each verb in the example sentences evokes a frame which is situation-specific. For instance, chance evokes the Daring frame, purchased the Com- merce buy frame, and injured the Cause harm frame. In addition, frames are associated with semantic roles corresponding to salient entities present in the situation evoked by the predicate. The semantic roles for the frame Daring are Agent and Manner, whereas for Commerce buy these are Buyer and Goods. A system trained on large amounts of such hand-annotated sentences typi- cally learns to identify the boundaries of the argu- ments of the verb predicate (argument identifica- tion) and label them with semantic roles (argument classification). A variety of methods have been developed for semantic role labeling with reasonably good per- formance (F1 measures in the low 80s on standard test collections for English; we refer the interested reader to the …

Content truncated to 500 words