Bingquan Liu

Subtopic segmentation of Chinese document: an adapted dotplot approach

An adapted dotplot model based on Chinese word sense quantization is presented to find the boundaries of subtopics in a document. The data reduction techniques of rough sets are introduced for the purpose of selecting axis words for word space. For discrete and filter data in the information table, the mutual information between axis words and feature words is calculated. Then the adapted model is constructed by replacing the counting identical words with the calculation of similarity between feature words. As a submodule of our InsunAbs Chinese auto-summarization system, its performance is indirectly evaluated through a quantitative evaluation. By comparison this adapted model outperforms the baseline and original dotplot model in the test experiments.

An approach to machine learning of Chinese Pinyin-to-character conversion for small-memory application

Chinese Pinyin-to-character conversion is used in Chinese character input through keyboard and Chinese speech recognition. The key of this kind of system is machine learning that fits system for specific user. In this paper, an effective approach of machine learning of Chinese Pinyin-to-character conversion for small-memory application is presented. The approach is based on iterative new word identification and word frequency increasing that results in more accurate segmentation of Chinese character gradually and satisfy the need of user finally. Applying proposed machine learning to Chinese character input system through keyboard improves accuracy of Pinyin-to-character conversion from 90% up to 98%. Such a system can run in very small memory (limited in 120 K) and satisfy the need of small-memory platform. With rapid development of digital appliances such as PDA, mobile telephone, intelligent refrigerator and etc., and with development of embedded operating system, Pinyin-to-character conversion presented in this paper has found its new application.

Combining multiple classifiers based on statistical method for handwritten Chinese character recognition

In various application areas of pattern recognition, combining multiple classifiers is regarded as a method for achieving a substantial gain in performance of systems. The paper presents a method for handwritten Chinese character recognition to combine multiple classifiers based on statistics. Fusion strategies are discussed for providing a basis for combining classifiers. These combination strategies are experimentally tested on an online handwritten Chinese character recognition system. In our experiments, other combination approaches are also involved for comparison.

A gradual combining method for multi-SVM classifiers based on distance estimation

A fusion algorithm based on multi SVM classifiers is presented in order to improve the performance of SVMs (support vector machines). Different SVM classifiers are trained with special instances. A gradual method based on distance estimation is utilized to combine different SVM classifiers into a sole learner. Instances that are easy to be categorized mistakenly by present classifier will be handed to the next classifier. These instances are chosen according to their distance to the optimal discrimination hyperplane. Evaluation on efficacy of the proposed multi-SVM classifier is carried on Chinese personal name recognition. Experiments show this multi SVM classifiers achieve better performance than that of single SVM learner and SVM ensemble using weighted voting scheme.