This work investigates the effective acquisition of lexical knowledge from the Web to perform semantic interpretation. The Web provides an unprecedented amount of natural language from which to gain knowledge useful for semantic interpretation. The knowledge acquired is described as common sense knowledge, information one uses in his or her daily life to understand language and perception. Novel approaches are presented for both the acquisition of this knowledge and use of the knowledge in semantic interpretation algorithms. The goal is to increase accuracy over other automatic semantic interpretation systems, and in turn enable stronger real world applications such as machine translation, advanced Web search, sentiment analysis, and question answering. The major contributions of this dissertation consist of two methods of acquiring lexical knowledge from the Web, namely a database of common sense knowledge and Web selectors. The first method is a framework for acquiring a database of concept relationships. To acquire this knowledge, relationships between nouns are found on the Web and analyzed over WordNet using information-theory, producing information about concepts rather than ambiguous words. For the second contribution, words called Web selectors are retrieved which take the place of an instance of a target word in its local context. The selectors serve for the system to learn the types of concepts that the sense of a target word should be similar. Web selectors are acquired dynamically as part of a semantic interpretation algorithm, while the relationships in the database are useful to iii stand-alone programs. A final contribution of this dissertation concerns a novel semantic similarity measure and an evaluation of similarity and relatedness measures on tasks of concept similarity. Such tasks are useful when applying acquired knowledge to semantic interpretation. Applications to word sense disambiguation, an aspect of semantic interpretation, are used to evaluate the contributions. Disambiguation systems which utilize semantically annotated training data are considered supervised. The algorithms of this dissertation are considered minimallysupervised; they do not require training data created by humans, though they may use humancreated data sources. In the case of evaluating a database of common sense knowledge, integrating the knowledge into an existing minimally-supervised disambiguation system significantly improved results – a 20.5% error reduction. Similarly, the Web selectors disambiguation system, which acquires knowledge directly as part of the algorithm, achieved results comparable with top minimally-supervised systems, an F-score of 80.2% on a standard noun disambiguation task. This work enables the study of many subsequent related tasks for improving semantic interpretation and its application to real-world technologies. Other aspects of semantic interpretation, such as semantic role labeling could utilize the same methods presented here for word sense disambiguation. As the Web continues to grow, the capabilities of the systems in this dissertation are expected to increase. Although the Web selectors system achieves great results, a study in this dissertation shows likely improvements from acquiring more data. Furthermore, the methods for acquiring a database of common sense knowledge could be applied in a more exhaustive fashion for other types of common sense knowledge. Finally, perhaps the greatest benefits from this work will come from the enabling of real world technologies that utilize semantic interpretation.
| Identifer | oai:union.ndltd.org:ucf.edu/oai:stars.library.ucf.edu:etd-2962 |
| Date | 01 January 2011 |
| Creators | Schwartz, Hansen A |
| Publisher | STARS |
| Source Sets | University of Central Florida |
| Language | English |
| Detected Language | English |
| Type | text |
| Format | application/pdf |
| Source | Electronic Theses and Dissertations |
Page generated in 0.0021 seconds