Most of the existing collocation extraction systems are based on globally significant statistical behaviors without mechanisms to handle different types of collocations. By taking compositionality, substitutability, modifiability and internal associations into consideration, collocations are categorized into four different types in this work. Based on the analysis for each type of collocation, a multi-stage extraction system is designed using different combinations of discriminative features so as to identify different types of collocations in different stages. Perceptron training is employed to optimize the consolidation of discriminative features from different sources. Experiment results show that the achieved performance is much better than most reported work.
|Name||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|Conference||4th International Conference on Machine Learning and Cybernetics, ICMLC 2005|
|Period||18/08/05 → 21/08/05|
- Theoretical Computer Science
- Computer Science(all)