A Computational Mental Lexicon Model based on Principles of Information Processing in Human Brain Kichun Nam, Dept. of Psychology, Korean University Heui-Seok Lim, Dept. of Information & Communication, Cheonan University Abstract: We proposed a computational mental lexicon model which can explain some linguistic effects in human language processing. The proposed model is based on modified trie data structure of which nodes are sorted by phoneme frequency. The model can explain frequency effect, length effect, lexical status effect, and similarity effect with easy because of its inherent characteristics of trie data structure. In addition, when we trained the model with 10 million Eojeol size corpus, the correlation between lexical decision time and frequency of input lexical was very similar that of human language processing. Background: Researches on knowledge representation and its structure of the mental lexicon play very important role in development of computational model in human language processing. There are some findings that are consistently obtained in human language processing such as frequency effect, length effect, lexical status effect, and similarity effect. Frequency effect is that highly frequent word is accessed more rapidly than low frequent word. Length effect is that longer word requires more time to be retrieved than shorter word. Lexical status effect means that non-word takes much time to be retrieved than normal word. Finally similarity effect is that non-word similar to normal word takes much time in lexical decision test. We claim that a computational mental lexicon model should explain the above linguistic characteristics. In this research, we proposed a new computational mental lexicon model and showed the validity of our model through simulation experiment. The proposed model: We use trie data structure as basic structure for our model. The origin of the name, trie is from the middle section of the word "reTRIEeval", and this origin hints on its usage. The trie data structure is based on two principles: a fixed set of indices and hierarchical indexing. The first requirement is usually met when we can index dictionary items by the alphabet. Each of the nodes' elements may point to another 26-element node, and so on. One of the advantages of using the trie as a basic structure for a computational mental lexicon is that length effect, lexical status effect, and similarity effect are explained with easy through the inherent characteristics of its data structure. But it is hard to explain frequency effect with general trie structure. We proposed a modified trie structure called frequency based trie(FB-trie) which enables to explain frequency effect as well as other linguistic effects. The proposed FB-trie is a trie which satisfies the following requirements. 1) The structure of a node is linked list structure. 2) The alphabet in a node is a set of Korean phonemes. 3) The alphabet elements in a node are sorted by descending order of frequencies of the alphabets in a Korean corpus. Nodes in the FB-trie are sorted by descending order of frequencies of the alphabets as described in the above the 3rd requirement. The 3rd requirement is needed to model frequency effect which means more frequent words are accessed more rapidly by visiting the minimal number of nodes. We tried to model language proficiency of Korean by adjusting the size of training corpus in indexing the FB-trie; large corpus for proficiency of an adult or an expert and small for that of an infant or novice. Results of researches: Our experiment was focused on whether the proposed model can explain frequency effect because FB-trie is a kind of trie and it can explain the other linguistic effect naturally. We used three different size of corpora as training data; a million Eojeol size, 5 million Eojeol size, and ten million Eojeol size. We use the same test data in [2] to compare the performance of the proposed model and that of human. The experiment was very promising that correlation between frequency and lexical access time was very similar to that of human language processing. Especially, results with ten million Eojeol corpus, the similarity of the correlation is the most. The reason, we think, is that the training data represents real world text more appropriately as the size of training data is increased [1] Marcus Taft, Reading And The Mental Lexicon, Essays in Cognitive Psychology, Lawrence Erlbaum Associates, Publishers, 1991. [2] 남기춘, 서광준, 최기선, 한글 단어 재인에 서의 단어 길이 효과, 한국인지과학회지:실험 및 인지, Vol. 9, No. 2, 1-18, pp. 1-18, 1997. A Computational Mental Lexicon Model based on Principles of Information Processing in Human Brain 87 Functional Neuroanatomical Study of Inhibitory ResponseControl: Activation and Neurochemical PET Brain Imaging Sang Eun Kim, Department of Nuclear Medicine, Seoul National University Bundang Hospital Abstract: We developed the technologies for constructing neural networks involved in the inhibitory response-control by positron emission tomography (PET) brain mapping and neuropsychological studies, which aimed at the ultimate goal of the project, which is “To construct the neuroanatomical and neurochemical circuits involved in the inhibitory response-control and to develop an integrated neuroanatomicalneurochemical cognitive model for the inhibitory response-control”. Background: Inhibitory response-control is an important cognitive function for the efficient information processing. Thus, the understanding of neural networks involved in the process of inhibitory response-control is essential for the development of human-like artificial intelligence system. Because neurochemical factors are involved in cognitive function, it is necessary to develop a cognitive model reflecting both neurochemical and neuroanatomical factors. However, neurochemical approaches to the cognitive mechanisms are very restricted. This project aims to construct neuroanatomical and neurochemical circuits involved in inhibitory response-control and to develop an integrated neuroanatomical-neurochemical cognitive model for inhibitory response-control. ields of Research: This project consists of neuroanatomical mapping studies using O-15 water and fluorodeoxyglucose (FDG) PET, neurochemical mapping studies mainly focusing on the dopamine system using neurotransmitter/receptor PET, neuropsychological evaluation and behavioral studies for the proper interpretation of the mapping studies, and the development of neuropsychological tasks for the assessment of inhibitory response-control function. Based on these studies, an integrated neuroanatomical-neurochemical cognitive model for the inhibitory response-control will be developed. Contents of Research Fields: 1. Mapping of regional cerebral glucose metabolism associated with normal and abnormal processes of inhibitory response-control using FDG PET in healthy controls and patients with attention deficit hyperactivity disorder (ADHD), obsessive compulsive disorder (OCD), and frontotemporal dementia (FTD) 2. Development of neuropsychological tasks and rating scales for the assessment of inhibitory response-control function 3. Study of the central dopaminergic circuitry associated with rewarded motor learning through the mapping of dopamine release with [11C]raclopride PET Results of Researches: 1. Construction of a functional neuroanatomical circuit associated with the inhibitory responsecontrol through the mapping of abnormal metabolism in the frontal lobe, which plays an important role in inhibitory response-control and executive functions, in patients with ADHD, OCD, and FTD (Fig. 1) 2. Development of inhibitory CPT and modified Go/NoGo tasks, which proved useful for the evaluation of the inhibitory response-control function. 3. Standardization of the Korean version of CAARS for the screening of adult ADHD 4. Mapping of dopamine release evoked by rewarded motor learning as a basis for the construction of the central dopaminergic circuitry associated with inhibitory response-control (Fig. 2) Fig. 1. rCMRglu mapping in patients with ADHD and OCD Fig. 2. Mapping of dopamine release evoked by rewarded motor learning Functional Neuroanatomical Study of Inhibitory Response-Control: Activation and Neurochemical PET Brain Imaging 88 Knowledgebase Prototype Construction and its Application for Human Knowledge Processing Modeling Key-Sun Choi, Department of Computer Science & KORTERM, KAIST Abstract The goal of the project is to construct a prototype of a domain-specific knowledgebase for modeling human knowledge processing. We perform research about knowledge extraction, event relation representation and knowledge retrieval and transformation. For its application, we construct question and answering system. In the second phase, we focus on knowledge extraction from documents and modeling knowledge retrieval and transformation. generality and efficiency for knowledge retrieval and construct knowledge. In the representation model, knowledge is represented by concept hierarchy, domain model, and event network Background: Information and knowledge are more and more increasing. Moreover, they are reproduced by human activity. The point of human activity is to analyze information written in texts and to formalize it into knowledge. Human brain is the system that performs language understanding, knowledge acquisition, and knowledge representation effectively. A system for storing and retrieving knowledge is necessary to acquire and represent knowledge. For language understanding, a multi-level language processing system that uses various knowledge resources is needed. This project aims at the system that mirrors human brain processing for language understanding, knowledge acquisition and knowledge representation 3. Modeling for Knowledge retrieval and Knowledge transformation: Design a knowledge retrieval model and a knowledge transformation model. Mirroring Texts Language Understanding Language Resources Knowledge Acquisition Texts reflecting Human knowledge World knowledge Domain-Specific Knowledge Knowledge Retrieval Personal Knowledge Knowledge base Knowledge Transformation Texts containing new knowledge Human Knowledge 2. Tools for Knowledge acquisition: Develop tools for knowledge acquisition and acquire knowledge about 1,000 nodes from the domain specific texts. Knowledge acquisition tool is composed of event extractor, concept mapper, and domain model generator. 4. Modeling for question answering system: Design a question answering (QA) system using knowledge retrieval and transformation model. The QA system generates answer using knowledge acquired from domain specific text. 5. Generalization: Generalize knowledge acquisition/retrieval/transformation model for other domain. Results of Researches: In the second phase, we automatically extract knowledge from medical domain texts - medical encyclopedia (1,000 knowledge nodes are extracted). The acquired knowledge is stored in database with knowledge representation format. Knowledge retrieval and transformation model also designed. A question and answering (QA) system based on the knowledge retrieval and transformation model is developed and is tested using the extracted 1,000 knowledge nodes. Our QA system retrieves knowledge based on similarity between question and extracted knowledge and automatically generates an answer based on scripts.. Alternative ego Fig. 1. System architecture Fields of Research: To perform research, sentence-level semantic categorization, information extraction, representing event relation, generation, human brain processing based question answering system will be necessary. Contents of Research Fields: 1. Research on a knowledge representation format: Design a knowledge representation model that has Knowledgebase prototype construction and its application for human knowledge processing modeling 89 Intelligent Visual System Design Based on Text Extraction in Natural Scene Images Yeong-Woo Choi, Department of Computer Science, Sookmyung Women’s University Abstract: In this research we develop text extraction methods on natural scene images relatively independent to text font style, location, orientation, size, color and background complexities. And we will apply the developed methods to design intelligent visual systems with speed and accuracy of minimizing user’s inconveniences. In the second year of this research, we develop a hierarchical feature combining and verification method to improve the text extraction accuracy and speed in natural scene images. Background: Texts in the natural scene images often contain important summarized information about the scene. If we can find these items accurately in real time, we can design the vision systems for assisting navigations of the moving robots or the blinds. Many of the previous researches have focused on extracting texts that are artificially inserted into the image like captions, thus the developed methods have many difficulties to be used in the real applications. In this research, we are developing the text extraction methods on the natural scene images that can be used in the real world application systems. Proposed Method: <Fig. 1> shows an overview of the research. In the first year of this research, text extraction methods based on the low level image processing are developed. In the second year, we developed the text extraction method both using the low-level image features of color continuity, gray-level variation and color variance and by verifying the extracted regions with the high-level text feature such as stroke. And the two level features are combined hierarchically. The color variance feature is added since the text strokes are distinctive in their color values to the background, and this value is more sensitive than the gray-level variations. The text level stroke features are extracted using a multi-resolution wavelet transforms on the local image areas and the feature vectors are input to a SVM(Support Vector Machine) classifier for the verification. The proposed method can even extract the text regions in the images of uneven illumination, different text orientations, and skewed texts. Contents and Results: Developing methods for information selection: 1) In the color reduced images by clustering the colors of the small number of histograms to the nearest colors and by improving the heuristic rules of the clustering method, we obtained the clustering results that are not sensitive to the initial selection of the colors. Thus, the clustering results are stabilized and the extraction accuracies are improved. 2) In the edge images the long line removal method was improved by considering the 4-directional histogram and the X/Y projections of the lines. Also, the emphasizing and combining the candidate text regions are improved. These improvements make possible the extraction of the vertical text regions, and also improve the text extraction accuracy. 3) To use the color variance of the text regions in the image, both the horizontal and vertical color variances are measured and their results are logically ANDed to remove noisy objects. With this feature the text extraction accuracy was improved a lot. 4) The combining method for the three candidate text regions is developed to improve the accuracy of the extraction. 5) Using multi-resolution wavelets transformations 36 features are extracted from the 16x16 image blocks that represent the directions of the character strokes. Only 12 distinctive features were selected for the verification purposes with Baysian method. Also, the SVM classifier was used for verifying the text candidate regions. 1. Improving the extraction efficiency: 1) Developed the skew correction method using the shearing transformation by clustering the similar skew lines according to the line locations. Also, developed the method for correcting the perspective of the text regions by finding the one or two vanishing points. 2) Improved the text quality up to the level of possible recognition by enlarging the low-resolution images, by developing the filtering method to remove the noises, and by developing the proper binarization method. 3) Improved the processing time of the Canny edge detection up to 1/5, 1/10 by improving the non-maximum suppression step. The quality of the edges in the image is maintained by controlling the variances of the Gaussian filter. We have tested the developed method using various kinds of the natural images and have confirmed that the extraction rates are very high even in complex images. Fig. 1. The proposed method Intelligent Visual System Design Based on Text Extraction in Natural Scene Images 90 A Study on Classifiers for Face Recognition and Face Detection using Color Model Hyeran Byun, Dept. of Computer Science, Yonsei University Abstraction: Face recognition at a distance requires both efficient face detection technique and classifier with good generalization performance. For real time facd detecting using skin-color information, we proposed multi-channel skin-color model and applied automatic whitening technique for the proposed method to be robust to color temperature. Then, we applied gaussian mixture model for real time detection. Face images have a large variations in appearances, so we should adopt a complex classifier such as SVM. We studied fundamentals of SVM and proposed a new method for multiclass problem. Motivations: For real-time face detection, it is essential to use color information. Since skin-color is classified with the color of artifact, we can perform face detection efficiently though color information. Unfortunately, because object recognition based on color information is very sensitive not only illumination but also color temperature. It causes the problem of Color Constant. Th conventional face recognition uses the simplest classifier such as Nearest Neighbor. It cannot apply efficiently to the variation of facial images. Therefore, We need to adopt complex classifier to have robust performance against such variations. SVM is one of solutions. However, we cannot apply it to a muti-class problem like face recognition. Real-time Skein-color Detection using Gaussian Mixture Model: We made a skin-color model containing only chrometic components to achieve robust detection against illumination changes. In this process, thorough analyzing the feature of skin-color, we can use combination of each color of red, blue and green channel. Th red component is distributed widly among the face skin-colors while blue and green components are concentrated on narrow regions, so we made the skin-color model using gaussian mixture model based on blue and green components after extracting candidate region by red component. The skin-color is very sensitive not only illumination but also color temperature, so we adopt automatic whitening technique to handle the variations due to the color temperature. We can separate skin-color region efficiently using multichannel color model with whitened images. The Methods adopting SVM as Multiclass classifiers: We adopt output coding method for applying SVM to multiclass problems. The output coding method can be described as follows: a complex multiclass problem is decomposed into a set of binary problems and then the outputs of binary classifiers for each binary problem are reconstructed. The conventional output coding methods are OPC and All-Pairs. The OPC method separates one class from all other classes and the All-Pairs method separates only two classes for each possible pair of classes. We propose a new method, the N-Division output coding method, which is a generalized form of OPC and All-Pairs. It divides problems produced by OPC into parts of N. In this way, we can make the decomposition which varies from OPC to All-Pairs by controlling the value of N. As N decreases, the decomposition becomes similar to OPC. If N is increasing, it becomes similar to All-Pairs. Fig. 2 shows the framework of our N-Division method. Results: The Fig. 1 showed skin-color detection results by skin-color model which is based on proposed color. In order to show the performance of our proposed output coding method, we conducted experiment on the ORL face dataset. The proposed method have good properties such as problem complexity although the difference in classification performance is not significant. And we presented that a value of 2 or 3 for N is desirable. Fig. 1 Face skin-color detection by using skin moded and gaussian mixture model. proposed Fig 2. N-Division Output Coding Method. A Study on Classifiers for Face Recognition and Face Detection Using Color Model 91 Noise-induced Multimode Behavior in Excitable System D. E. Postnov1 , O. V. Sosnovtseva1, S.K. Han2, and W.S. Kim2, 1Dept. of Physics, Saratov State University, Russia, 2Dept. of Physics, Chungbuk Nat'l. University, Korea Abstract: Based on experiments with electronic circuits, we show how a system of coupled excitable units can posses several noise-induced oscillatory mode. We characterize the multimode organization in term of the coherence resonance effect. Background: While generation and entrainment of single-mode deterministic or stochastic oscillations are well understood, the dynamics of systems with many oscillations with different modes is less studied Many living systems perform oscillations with different modes [1,2,3]. We focus on time scales that are delivered and controlled by noise and that did not exist in the deterministic case. [1] X.J. Wang, Neuroscience 59, 21, 1994 [2] A. Neiman and D.F. Russel, Phys. Rev. Lett. 86, 3443, 2001 [3] S.K.Han, T.G.Yim, D.Postnov, and O.Sosnovtseva, Phys. Rev. Lett. 83, 1771, 1999 [4] D.Postnov, O Sosnovtseva, S.K.Han, and W.S.Kim, Phys. Rev. E 66, 016203, 2002 Results: We examined different implementations of coupled identical excitable units with different types of coupling as shown in Figure 1 [4]. Fig. 1. Different implementation of coupled exci-table units. (a) Monovibrator;(b)mutually coupled units;(c) circle configuration Figure 2. demonstrates different collective behaviors when the coupling strength and noise intensity of two symmetrically coupled excitable units is varied. It is clearly seen that two-mode behavior is observed through the resonant and nonresonant ration between the noise-induced frequencies. In Figure3, three different frequencies is observed when three functional excitable units are coupled as circle shape (Figure1(c)). From this results, we can state that the three-unit system is able to generate three-mode stochastic behavior. For more realistic neuronal excitable system, we consider the Figure4(a) which emulate the simplest circuit model of snail’s breathing rhythm generator. As shown in Figure 4(c) , it can generate two-mode stochastic oscillation behavior. Discussions: We have shown that a simple system of coupled excitable functional units can generate a few oscillatory modes that are induced and controlled by noise and coupling. Noise-induced multimode behavior in excitable system Fig. 2. Two-mode collective response in the system of two mutually coupled monovibrators[Figure1(b)]. (a) D=0.475V2. (b)D=0.77V2. Fig. 3. Power spectrum illustration three-mode collective behavior in a system of three interacting excitable unit[figure1(c)] at D=0.35V2 and g=0.03. Fig. 4. (a) Two monovibrators with delayed inhibitory couplings imitate the simple neural circuit.(b) Stochastic spike trains generated by the first and second exitable units. 92 Neural Network Modeling for Intelligent Novelty Detection Sungzoon Cho, Department of Industrial Engineering, Seoul National University Abstract: In this research, we developed three novelty detection models based on auto associative neural networks (AANNs) and support vector machines (SVMs). The first was applied to typing pattern based password identity verification problem, the second to trading of index future and the third to default company detection problem. All three problems share the characteristic fit for novelty detection in that training data set is heavily or completely imbalanced. The first reduced the error rate 5 times while the third reduced the error by half. The second resulted in an excessive profit of 50 points in a simulated trading for a year. Background: Novelty detection is to identify novel patterns out of normal ones. There are many situations where normal patterns can be easily obtained yet novel ones can not. For instance, the way one types his password can be captured and stored as timing vectors. However, it is neither easy nor desirable to obtain such patterns from the rest of the population including potential imposters. Although the problem can be cast as a binary classification problem, scarcity or lack of data from one class prevents one from using typical two class classifiers. A traditional neural network algorithm handling this novelty detection problem is auto associative neural network. The AANN, which was trained only with the normal patterns, produces a small error on a normal pattern while a large error on a novel pattern. Employing a proper threshold, one can distinguish a normal from a novel pattern based on the error value. Lately, SVM has gained popularities because of its clear cut theory and good performance. Another algorithm which can be applied to novelty detection problem is one-class SVM. In this research, we developed novelty detection models based on both AANNs and SVMs. The models were applied to typing pattern based password identity verification, index future trading and default company detection. The password identity verification system detects imposters’ typing patterns from user’s. There were no training patterns available of non-user. We additionally developed a feature selection algorithm and integrated it into the system. For financial applications, a sudden change in future price trends or a company having abnormal financial status was defined as novelty. In both cases, the number of novel patterns was relatively quite small, thus novelty detection approach was employed. The research details were published in the papers [1] through [4]. Approaches: (A) Typing pattern based password identity verification: We collected normal typing patterns from 21 users and novel ones from 15 imposters. Novelty detection algorithms based on AANN and SVM were implemented independently, and the results were compared [1]. In addition, we studied feature selection algorithms based on wrapper approach [2]. In the proposed wrapper feature selection algorithm, SVM was used as the induction algorithm and GA was adopted as the search mechanism. Use of SVM as the induction algorithm allowed us to adopt wrapper approach since its training time was 3 orders of magnitude smaller than that of AANN. (B) Future trading and default company detection: First, KOSPI future prices from 1999.01 to 2001.12 were used for training and simulation [3]. We extracted four technical indicators (Volume Ratio, Relative Strength Index, Rate Of Change, Stochastic Slow %D), which were used as input variables. Up-trend and down-trend were separately trained and later combined to produce trading signals. Second, the balance sheets of 2000 solvent companies were used for training. [4]. A neural network based bagging model was built as a novelty detector. Results: (A) Typing pattern based password identity verification: Fast learning SVM allowed a wrapper based feature selection, which resulted in error reduction to FRR=3.54% from FRR=15.78% (when FAR=0). (B) Future trading and default company detection: The simulated trading made an annual profit of more than 50 points without considering trading cost [3]. For default company detection, novelty detection AANN reduced the error by half compared with the conventional approach where imbalanced data set was used [4]. [1] E. Yu and S. Cho, "Novelty Detection Approach for Keystroke Dynamics Identity Verification", Fourth International Conference on Intelligent Data Engineering and Automated Learning, IDEAL 2003, HongKong. [2] E. Yu, and S. Cho, "GA-SVM Wrapper Approach for Feature Subset Selection in Keystroke Dynamics Identity Verification", 2003 INNS-IEEE International Joint Conference on Neural Networks, July 2003, Portland, USA. [3] J. Lee, S. Cho, and J. Baek, “Trend Detection Using Auto-Associative Neural Networks: Intraday KOSPI 200 Futures,” IEEE 2003 International Conference on Computational Intelligence for Financial Engineering (CIFEr2003), Mar 21-23, 2003, Hong Kong. [4] J. Baek and S. Cho, “`Bankruptcy Prediction for Credit Risk Using an Auto-Associative Neural Network in Korean firms,” IEEE 2003 International Conference on Computational Intelligence for Financial Engineering (CIFEr2003), Mar 21-23, 2003, Hong Kong. Neural Network Modeling for Intelligent Novelty Detection 93