Studies of Radical Model for Retrieval of Cursive Chinese Handwritten Annotations Matthew Ma 1 and Chi Zhang 2 and Patrick Wang, IAPR Fellow 3 1 Panasonic Information and Networking Technologies Laboratory Panasonic Technologies, Inc. Two Research Way, Princeton, NJ 08540 USA 2,3 College of Computer Science, Northeastern University, Boston, MA 02115 Abstract. Our research focuses on Chinese online ink matching that tries to match handwritten annotations with handwritten queries without attempting to recognize them. Previously, we proposed a semantic matching scheme that uses elastic matching with a dynamic programming approach based on the radical model of Chinese characters. By means of semantic matching, a handwritten annotation may also be retrieved independently of writers via typed text query, or stored texts can be retrieved by handwritten queries. This work concerns with the behavior of the previously proposed radical model in several aspects including character normalization, stroke segmentation, structural information, dynamic programming costs and schemes. Based on our study, a new radical model is proposed. As a result, the recall of retrieval by handwritten query reaches 90% for the first hit (an improvement of 20% over previous results) and the recall by text query reaches 80% when top 20 matches are returned. 1 Introduction and Motivation In language computing, both on-line and off-line handwritten Chinese character recognition (HCCR) have been existing for several decades. Although online recognition has the advantage over offline because the temporal order of the input points and strokes is provided, it still has proved to be a more difficult problem than most people anticipated because of the variations of the way people write and a complex training process involved [1]. In addition, a large lexicon is to be incorporated due to the large number of characters (3,000 – 5,000) that are daily used. Instead of handwriting recognition, some research work has been conducted on online ink matching that tries to match a handwritten query against raw ink data without attempting to recognize them [4]. This technique can be used in a document annotating and browsing system, which enables users to search their personal notes by a handwritten query. Similar work and various applications also appear elsewhere [6,7]. Recently, a semantic matching method was proposed by Ma et al. [5]. By extending Wang’s Learning by Knowledge paradigm [8], this method focuses on the 1 Corresponding author. E-mail: mma@research.panasonic.com