Explorations in disambiguation using XML text representation

Explorations in Disambiguation Using XML Text Representation Kenneth C. Litkowski CL Research 9208 Gue Road Damascus, MD 20872 ken@clres.com Abstract In SENSEVAL-3, CL Research participated in four tasks: English all-words, English lexical sample, disambiguation of WordNet glosses, and automatic labeling of semantic roles. This participation was performed within the development of CL Research’s Knowledge Management System, which massively tags texts with syntactic, semantic, and discourse characterizations and attributes. This System is fully integrated with CL Research’s DIMAP dictionary maintenance software, which provides access to one or more dictionaries for disambiguation and representation. Our core disambiguation functionality, unchanged since SENSEVAL-2, performed at a level comparable to our previous performance. Our participation in the SENSEVAL-3 tasks was concerned primarily with text processing and representation issues and did not advance our disambiguation capabilities. Introduction CL Research participated in four SENSEVAL-3 tasks: English all-words, English lexical sample, disambiguation of WordNet glosses, and automatic labeling of semantic roles. We also ran the latter two tasks, but since their test sets were generated blindly, our results did not involve use of any prior information. Our participation in these tasks is a continuation and extension of our efforts to perform NLP tasks within an integrated text processing system known as the Knowledge Management System (KMS). KMS parses and processes text into an XML representation tagged with syntactic, semantic, and discourse properties. This representation is then used for such tasks as question answering and text summarization (Litkowski, 2004a; Litkowski, 2004b). The SENSEVAL-3 tasks were performed as part of CL Research’s efforts to extend and improve the semantic characterizations in the KMS XML representations. For each SENSEVAL-3 task, the corresponding texts in the test sets were processed using the general KMS functionality. However, since the texts involved in the SENSEVAL tasks were quite small, the amount of processing was quite minimal. The descriptions below focus on the integration of disambiguation technology in a larger system and do not present any advancements in this technology. 1 The SENSEVAL-3 All-Words Task Our procedures for performing this task and our results were largely unchanged from SENSEVAL-2 (Litkowski, 2001; Litkowski, 2002). Our system is unsupervised, instead relying on information in whatever dictionary is being used to disambiguate the words. In this case, as in SENSEVAL-2, WordNet 1.7.1 was used. The main types of information used are default sense selection, idiomatic usage, syntactic and semantic clues, subcategorization patterns, word forms, syntactic usage, context, and topics or subject fields. As pointed out in Litkowski (2002), the amount of information available in WordNet is problematic. Additional information suitable for disambiguation is available in WordNet 2.0, but we were unable to test the effect of the changes, even though we could have easily switched our system to use this later version. In performing this task, we spent some time cleaning the text files, removing extraneous material and creating a more natural text file (e.g., joining contractions). Use of a preprocessed file is somewhat difficult. Since some tokens to be disambiguated were unnatural (e.g., “that’s” broken into two …

Content truncated to 500 words