CLR: Integration of FrameNet in a Text Representation System

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 113–116, Prague, June 2007. c⃝2007 Association for Computational Linguistics CLR: Integration of FrameNet in a Text Representation System Ken. Litkowski CL Research 9208 Gue Road Damascus, MD 20872 ken@clres.com Abstract In SemEval-2007, CL Research participated in the task for Frame Semantic Structure Extraction. Participation in this task was used as the vehicle for efforts to integrate and exploit FrameNet in a comprehensive text processing system. In particular, this involved steps to build a FrameNet dictionary with CL Research’s DIMAP dictionary software and to use this dictionary (along with its semantic network processing capabilities) in processing text into XML representations. Implementation of the entire integrated package is only in its initial stages and was used to make only a bare submission of frame identification. On this task, over all texts, a recall of 0.372, a precision of 0.553, and an F-score of 0.445 were achieved. Considering only targets included in the DIMAP FrameNet dictionary, the overall F-score is 0.605. These results, competitive with the top scoring system, support continued attempts at a dictionary-based approach to frame structure extraction. 1 Introduction CL Research participated in the SemEval-2007 task for Frame Semantic Structure Extraction. In participating in this task, we integrated the use of FrameNet in the Text Parser component of the CL Research Knowledge Management System (KMS). In particular, we created a FrameNet dictionary from the FrameNet databases with the CL Research DIMAP dictionary software and used this dictionary as a lexical resource. This new lexical resource was integrated in the same manner as other lexical resources (including WordNet and the Oxford Dictionary of English (ODE, 2004)). As such, the FrameNet dictionary was available as the basis for sense disambiguation. In the CL Research Text Parser, this integration was seamless, in which disambiguation can be performed against several lexical resources. This work attempts to expand on semantic role labeling experiments in Senseval-3 (Litkowski, 2004a, and Litkowski, 2004b). In the following sections, we first describe the overall structure of the CL Research Knowledge Management System and Text Parser, describing their general parsing and text analysis routines. Next, we describe the creation of the FrameNet dictionary, particularly identifying design considerations to exploit the richness of the FrameNet data. In section 4, we describe our submission for the SemEval task. In section 5, we describe our results. Finally, we identify next steps that can be taken within the CL Research KMS and DIMAP environments to extend the FrameNet data. 2 CL Research Text Processing The CL Research Knowledge Management System (KMS) is an integrated environment for performing several higher level applications, particularly question answering and summarization. The underlying architecture of KMS relies on an XML representation of texts that captures discourse structure and discourse elements, particularly noun phrases, verbs, and semantic roles (predominantly as reified in prepositions). The texts that are represented include primarily full texts as they may appear in several forms, but also include questions, topic specifications for which summaries are desired, …

Content truncated to 500 words