International Journal of Computer Processing of Languages

This paper proposes a new, promising character recognition system with a category-dependent shape normalization technique that normalizes an input pattern against each reference pattern adaptively using global affine transformation (GAT) as follows. (1) An input character pattern is fed to "the basic OCR", the most powerful of the conventional OCRs. (2) The basic OCR plays the role of rough classification and outputs a small set of candidate categories for the input pattern. (3) GAT normalizes the input pattern against each candidate's reference pattern adaptively. (4) Each adaptively normalized input pattern is fed again to the basic OCR. (5) The final recognition result is obtained using the updated "distance" values within candidate categories. In experiments, our basic OCR linked to GAT adaptive normalization is successfully applied to 28,694 patterns of totally unconstrained handwritten characters, including Kanji, Kana, and alphanumerics, written by 300 people, with substantial improvements in recognition accuracy.

In this paper, we propose a grapheme verifier based on a neural network (NN) for on-line handwritten Hangul (Korean script) recognition. The verifier complements the global information and discrimination power to hidden Markov models (HMMs) by penalizing incorrect inputs and passing correct ones intact. It takes structural features from partial inputs aligned to HMM states as global information and gives the probability of their validity; one if they belong to the grapheme and zero if they do not. Hangul syllable characters are recognized with the network of grapheme models which are composed of HMM recognizers and NN verifiers. Probabilities of verifiers are integrated with those of HMMs during the search for the most probable path in the network. Experimental result showed that the proposed verifiers effectively reduced recognition errors of HMM recognizers for Hangul syllable characters by 39.2% on average.

A context-free grammar-based language model for string recognition has been developed. The developed language model is implemented as a set of graphs which are equivalent to a recursive transition networks. This graph-structured model is not only more compact in memory size than the trie-structured model but also more powerful in describing variations of recognition-target phrases, such as use of characters, selection of words, and structures of phrases. This can improve recognition accuracy. We also developed a method to semi-automatically generate the language model from a collection of sample phrase texts and predefined variation rules. An algorithm to match this language model to the character-recognition results was devised. The model and the generation method have been successfully applied to practical applications such as address recognition. And experimental tests with freely-handwritten addresses show the effectiveness for implementation of the language model and the improvement of the recognition accuracy.

A new method for information extraction from document images is proposed in this paper as the basis for a document reader which can extract the required keywords and their logical relationship from various printed documents. The proposed method consists of robust keyword matching, global document matching, and postprocessing for matching errors. First, robust keyword matching between two-dimensional OCR results consisting of a set of possible character candidate lists and a set of keywords defined in the keyword dictionary is carried out. This keyword dictionary includes incorrect words with typical OCR errors and segments of words in order to deal with OCR errors. Next, document matching is invoked between keyword matching results in an input document and document models. Each document model consists of a set of word models with their logical relationship described in terms of a tree structure. This model matching extracts the required keywords and their logical relationship from the input document and determines the most suitable model for the input document. Finally, postprocessing using heuristic rules defined in the model is applied to document matching results to recover keyword matching errors and to modify keyword matching results. This comprehensive approach solves word segmentation problems accurately even if a document has unknown words, compound words, or incorrect words due to OCR errors. Experimental results obtained for 100 documents show that the method is robust and effective for various document structures.

This paper proposes a method to generate realistic character images from on-line patterns. From the pen trajectory of an on-line pattern, the proposed method generates images of various stroke shapes using four painting modes: constant line mode, proportional mode and two calligraphic modes. Particularly, the calligraphic modes combine the pen trajectory with real stroke-shape images so that the generated images resemble the characters produced with brush pen. In the calligraphic painting mode based on primitive stroke identification (PSI), the strokes of on-line patterns are classified into different classes and each class of strokes is painted with the corresponding stroke shape template; while in the calligraphic painting mode based on stroke component classification (SCC), each stroke is decomposed into ending, bending and connecting parts, and each part is painted with a stroke shape template. Decomposing strokes into parts is helpful to deal with connected strokes in cursive writing. Our method of image transformation serves two purposes: supplying image samples for off-line recognition, and application of off-line recognition methods to on-line recognition. The experimental results show that our methods are well suited for both purposes. We show that calligraphic painting mode is appropriate for off-line recognition while constant line mode and proportional mode are appropriate for on-line recognition.

Temporal information is a significant semantic carrier in natural languages. In a broader sense, when we speak of the temporal information contained in sentences, we do not simply mean those explicit specifications, such as time clauses or phrases associated with the actions addressed. It also includes the information which is implicitly embedded in verbs. Thus, modelling of the temporal aspects of a language is more complicated than a physical time-dependent system. Over the past years, temporal information processing and extraction have received increasing attentions. Nevertheless, only a few researchers have investigated these areas in Chinese. The aim of this paper is to investigate both the explicit expressions and the inherent temporal properties of Chinese sentences. Furthermore, an algorithm to automatically classify a situation described by a sentence into three temporal categories, namely event, process and state, is proposed. This algorithm is the fundamental building block for effective Chinese temporal information processing.