• Journal of Intelligence and Information Systems
• /
• v.16 no.3
• /
• pp.77-97
• /
• 2010

Market timing is an investment strategy which is used for obtaining excessive return from financial market. In general, detection of market timing means determining when to buy and sell to get excess return from trading. In many market timing systems, trading rules have been used as an engine to generate signals for trade. On the other hand, some researchers proposed the rough set analysis as a proper tool for market timing because it does not generate a signal for trade when the pattern of the market is uncertain by using the control function. The data for the rough set analysis should be discretized of numeric value because the rough set only accepts categorical data for analysis. Discretization searches for proper "cuts" for numeric data that determine intervals. All values that lie within each interval are transformed into same value. In general, there are four methods for data discretization in rough set analysis including equal frequency scaling, expert's knowledge-based discretization, minimum entropy scaling, and na$\ddot{i}$ve and Boolean reasoning-based discretization. Equal frequency scaling fixes a number of intervals and examines the histogram of each variable, then determines cuts so that approximately the same number of samples fall into each of the intervals. Expert's knowledge-based discretization determines cuts according to knowledge of domain experts through literature review or interview with experts. Minimum entropy scaling implements the algorithm based on recursively partitioning the value set of each variable so that a local measure of entropy is optimized. Na$\ddot{i}$ve and Booleanreasoning-based discretization searches categorical values by using Na$\ddot{i}$ve scaling the data, then finds the optimized dicretization thresholds through Boolean reasoning. Although the rough set analysis is promising for market timing, there is little research on the impact of the various data discretization methods on performance from trading using the rough set analysis. In this study, we compare stock market timing models using rough set analysis with various data discretization methods. The research data used in this study are the KOSPI 200 from May 1996 to October 1998. KOSPI 200 is the underlying index of the KOSPI 200 futures which is the first derivative instrument in the Korean stock market. The KOSPI 200 is a market value weighted index which consists of 200 stocks selected by criteria on liquidity and their status in corresponding industry including manufacturing, construction, communication, electricity and gas, distribution and services, and financing. The total number of samples is 660 trading days. In addition, this study uses popular technical indicators as independent variables. The experimental results show that the most profitable method for the training sample is the na$\ddot{i}$ve and Boolean reasoning but the expert's knowledge-based discretization is the most profitable method for the validation sample. In addition, the expert's knowledge-based discretization produced robust performance for both of training and validation sample. We also compared rough set analysis and decision tree. This study experimented C4.5 for the comparison purpose. The results show that rough set analysis with expert's knowledge-based discretization produced more profitable rules than C4.5.

• Journal of Intelligence and Information Systems
• /
• v.26 no.2
• /
• pp.105-129
• /
• 2020

This study uses corporate data from 2012 to 2018 when K-IFRS was applied in earnest to predict default risks. The data used in the analysis totaled 10,545 rows, consisting of 160 columns including 38 in the statement of financial position, 26 in the statement of comprehensive income, 11 in the statement of cash flows, and 76 in the index of financial ratios. Unlike most previous prior studies used the default event as the basis for learning about default risk, this study calculated default risk using the market capitalization and stock price volatility of each company based on the Merton model. Through this, it was able to solve the problem of data imbalance due to the scarcity of default events, which had been pointed out as the limitation of the existing methodology, and the problem of reflecting the difference in default risk that exists within ordinary companies. Because learning was conducted only by using corporate information available to unlisted companies, default risks of unlisted companies without stock price information can be appropriately derived. Through this, it can provide stable default risk assessment services to unlisted companies that are difficult to determine proper default risk with traditional credit rating models such as small and medium-sized companies and startups. Although there has been an active study of predicting corporate default risks using machine learning recently, model bias issues exist because most studies are making predictions based on a single model. Stable and reliable valuation methodology is required for the calculation of default risk, given that the entity's default risk information is very widely utilized in the market and the sensitivity to the difference in default risk is high. Also, Strict standards are also required for methods of calculation. The credit rating method stipulated by the Financial Services Commission in the Financial Investment Regulations calls for the preparation of evaluation methods, including verification of the adequacy of evaluation methods, in consideration of past statistical data and experiences on credit ratings and changes in future market conditions. This study allowed the reduction of individual models' bias by utilizing stacking ensemble techniques that synthesize various machine learning models. This allows us to capture complex nonlinear relationships between default risk and various corporate information and maximize the advantages of machine learning-based default risk prediction models that take less time to calculate. To calculate forecasts by sub model to be used as input data for the Stacking Ensemble model, training data were divided into seven pieces, and sub-models were trained in a divided set to produce forecasts. To compare the predictive power of the Stacking Ensemble model, Random Forest, MLP, and CNN models were trained with full training data, then the predictive power of each model was verified on the test set. The analysis showed that the Stacking Ensemble model exceeded the predictive power of the Random Forest model, which had the best performance on a single model. Next, to check for statistically significant differences between the Stacking Ensemble model and the forecasts for each individual model, the Pair between the Stacking Ensemble model and each individual model was constructed. Because the results of the Shapiro-wilk normality test also showed that all Pair did not follow normality, Using the nonparametric method wilcoxon rank sum test, we checked whether the two model forecasts that make up the Pair showed statistically significant differences. The analysis showed that the forecasts of the Staging Ensemble model showed statistically significant differences from those of the MLP model and CNN model. In addition, this study can provide a methodology that allows existing credit rating agencies to apply machine learning-based bankruptcy risk prediction methodologies, given that traditional credit rating models can also be reflected as sub-models to calculate the final default probability. Also, the Stacking Ensemble techniques proposed in this study can help design to meet the requirements of the Financial Investment Business Regulations through the combination of various sub-models. We hope that this research will be used as a resource to increase practical use by overcoming and improving the limitations of existing machine learning-based models.

• Journal of Intelligence and Information Systems
• /
• v.25 no.1
• /
• pp.163-177
• /
• 2019

As smartphones are getting widely used, human activity recognition (HAR) tasks for recognizing personal activities of smartphone users with multimodal data have been actively studied recently. The research area is expanding from the recognition of the simple body movement of an individual user to the recognition of low-level behavior and high-level behavior. However, HAR tasks for recognizing interaction behavior with other people, such as whether the user is accompanying or communicating with someone else, have gotten less attention so far. And previous research for recognizing interaction behavior has usually depended on audio, Bluetooth, and Wi-Fi sensors, which are vulnerable to privacy issues and require much time to collect enough data. Whereas physical sensors including accelerometer, magnetic field and gyroscope sensors are less vulnerable to privacy issues and can collect a large amount of data within a short time. In this paper, a method for detecting accompanying status based on deep learning model by only using multimodal physical sensor data, such as an accelerometer, magnetic field and gyroscope, was proposed. The accompanying status was defined as a redefinition of a part of the user interaction behavior, including whether the user is accompanying with an acquaintance at a close distance and the user is actively communicating with the acquaintance. A framework based on convolutional neural networks (CNN) and long short-term memory (LSTM) recurrent networks for classifying accompanying and conversation was proposed. First, a data preprocessing method which consists of time synchronization of multimodal data from different physical sensors, data normalization and sequence data generation was introduced. We applied the nearest interpolation to synchronize the time of collected data from different sensors. Normalization was performed for each x, y, z axis value of the sensor data, and the sequence data was generated according to the sliding window method. Then, the sequence data became the input for CNN, where feature maps representing local dependencies of the original sequence are extracted. The CNN consisted of 3 convolutional layers and did not have a pooling layer to maintain the temporal information of the sequence data. Next, LSTM recurrent networks received the feature maps, learned long-term dependencies from them and extracted features. The LSTM recurrent networks consisted of two layers, each with 128 cells. Finally, the extracted features were used for classification by softmax classifier. The loss function of the model was cross entropy function and the weights of the model were randomly initialized on a normal distribution with an average of 0 and a standard deviation of 0.1. The model was trained using adaptive moment estimation (ADAM) optimization algorithm and the mini batch size was set to 128. We applied dropout to input values of the LSTM recurrent networks to prevent overfitting. The initial learning rate was set to 0.001, and it decreased exponentially by 0.99 at the end of each epoch training. An Android smartphone application was developed and released to collect data. We collected smartphone data for a total of 18 subjects. Using the data, the model classified accompanying and conversation by 98.74% and 98.83% accuracy each. Both the F1 score and accuracy of the model were higher than the F1 score and accuracy of the majority vote classifier, support vector machine, and deep recurrent neural network. In the future research, we will focus on more rigorous multimodal sensor data synchronization methods that minimize the time stamp differences. In addition, we will further study transfer learning method that enables transfer of trained models tailored to the training data to the evaluation data that follows a different distribution. It is expected that a model capable of exhibiting robust recognition performance against changes in data that is not considered in the model learning stage will be obtained.

• Journal of Intelligence and Information Systems
• /
• v.24 no.1
• /
• pp.205-225
• /
• 2018

Convolutional Neural Network (ConvNet) is one class of the powerful Deep Neural Network that can analyze and learn hierarchies of visual features. Originally, first neural network (Neocognitron) was introduced in the 80s. At that time, the neural network was not broadly used in both industry and academic field by cause of large-scale dataset shortage and low computational power. However, after a few decades later in 2012, Krizhevsky made a breakthrough on ILSVRC-12 visual recognition competition using Convolutional Neural Network. That breakthrough revived people interest in the neural network. The success of Convolutional Neural Network is achieved with two main factors. First of them is the emergence of advanced hardware (GPUs) for sufficient parallel computation. Second is the availability of large-scale datasets such as ImageNet (ILSVRC) dataset for training. Unfortunately, many new domains are bottlenecked by these factors. For most domains, it is difficult and requires lots of effort to gather large-scale dataset to train a ConvNet. Moreover, even if we have a large-scale dataset, training ConvNet from scratch is required expensive resource and time-consuming. These two obstacles can be solved by using transfer learning. Transfer learning is a method for transferring the knowledge from a source domain to new domain. There are two major Transfer learning cases. First one is ConvNet as fixed feature extractor, and the second one is Fine-tune the ConvNet on a new dataset. In the first case, using pre-trained ConvNet (such as on ImageNet) to compute feed-forward activations of the image into the ConvNet and extract activation features from specific layers. In the second case, replacing and retraining the ConvNet classifier on the new dataset, then fine-tune the weights of the pre-trained network with the backpropagation. In this paper, we focus on using multiple ConvNet layers as a fixed feature extractor only. However, applying features with high dimensional complexity that is directly extracted from multiple ConvNet layers is still a challenging problem. We observe that features extracted from multiple ConvNet layers address the different characteristics of the image which means better representation could be obtained by finding the optimal combination of multiple ConvNet layers. Based on that observation, we propose to employ multiple ConvNet layer representations for transfer learning instead of a single ConvNet layer representation. Overall, our primary pipeline has three steps. Firstly, images from target task are given as input to ConvNet, then that image will be feed-forwarded into pre-trained AlexNet, and the activation features from three fully connected convolutional layers are extracted. Secondly, activation features of three ConvNet layers are concatenated to obtain multiple ConvNet layers representation because it will gain more information about an image. When three fully connected layer features concatenated, the occurring image representation would have 9192 (4096+4096+1000) dimension features. However, features extracted from multiple ConvNet layers are redundant and noisy since they are extracted from the same ConvNet. Thus, a third step, we will use Principal Component Analysis (PCA) to select salient features before the training phase. When salient features are obtained, the classifier can classify image more accurately, and the performance of transfer learning can be improved. To evaluate proposed method, experiments are conducted in three standard datasets (Caltech-256, VOC07, and SUN397) to compare multiple ConvNet layer representations against single ConvNet layer representation by using PCA for feature selection and dimension reduction. Our experiments demonstrated the importance of feature selection for multiple ConvNet layer representation. Moreover, our proposed approach achieved 75.6% accuracy compared to 73.9% accuracy achieved by FC7 layer on the Caltech-256 dataset, 73.1% accuracy compared to 69.2% accuracy achieved by FC8 layer on the VOC07 dataset, 52.2% accuracy compared to 48.7% accuracy achieved by FC7 layer on the SUN397 dataset. We also showed that our proposed approach achieved superior performance, 2.8%, 2.1% and 3.1% accuracy improvement on Caltech-256, VOC07, and SUN397 dataset respectively compare to existing work.

• Journal of Internet Computing and Services
• /
• v.21 no.4
• /
• pp.17-23
• /
• 2020

Biometric information indicating measurement items related to human characteristics has attracted great attention as security technology with high reliability since there is no fear of theft or loss. Among these biometric information, fingerprints are mainly used in fields such as identity verification and identification. If there is a problem such as a wound, wrinkle, or moisture that is difficult to authenticate to the fingerprint image when identifying the identity, the fingerprint expert can identify the problem with the fingerprint directly through the preprocessing step, and apply the image processing algorithm appropriate to the problem. Solve the problem. In this case, by implementing artificial intelligence software that distinguishes fingerprint images with cuts and wrinkles on the fingerprint, it is easy to check whether there are cuts or wrinkles, and by selecting an appropriate algorithm, the fingerprint image can be easily improved. In this study, we developed a total of 17,080 fingerprint databases by acquiring all finger prints of 1,010 students from the Royal University of Cambodia, 600 Sokoto open data sets, and 98 Korean students. In order to determine if there are any injuries or wrinkles in the built database, criteria were established, and the data were validated by experts. The training and test datasets consisted of Cambodian data and Sokoto data, and the ratio was set to 8: 2. The data of 98 Korean students were set up as a validation data set. Using the constructed data set, five CNN-based architectures such as Classic CNN, AlexNet, VGG-16, Resnet50, and Yolo v3 were implemented. A study was conducted to find the model that performed best on the readings. Among the five architectures, ResNet50 showed the best performance with 81.51%.

• Journal of Intelligence and Information Systems
• /
• v.21 no.4
• /
• pp.1-16
• /
• 2015

Traffic accident is one of the major cause of death worldwide for the last several decades. According to the statistics of world health organization, approximately 1.24 million deaths occurred on the world's roads in 2010. In order to reduce future traffic accident, multipronged approaches have been adopted including traffic regulations, injury-reducing technologies, driving training program and so on. Records on traffic accidents are generated and maintained for this purpose. To make these records meaningful and effective, it is necessary to analyze relationship between traffic accident and related factors including vehicle design, road design, weather, driver behavior etc. Insight derived from these analysis can be used for accident prevention approaches. Traffic accident data mining is an activity to find useful knowledges about such relationship that is not well-known and user may interested in it. Many studies about mining accident data have been reported over the past two decades. Most of studies mainly focused on predict risk of accident using accident related factors. Supervised learning methods like decision tree, logistic regression, k-nearest neighbor, neural network are used for these prediction. However, derived prediction model from these algorithms are too complex to understand for human itself because the main purpose of these algorithms are prediction, not explanation of the data. Some of studies use unsupervised clustering algorithm to dividing the data into several groups, but derived group itself is still not easy to understand for human, so it is necessary to do some additional analytic works. Rule based learning methods are adequate when we want to derive comprehensive form of knowledge about the target domain. It derives a set of if-then rules that represent relationship between the target feature with other features. Rules are fairly easy for human to understand its meaning therefore it can help provide insight and comprehensible results for human. Association rule learning methods and subgroup discovery methods are representing rule based learning methods for descriptive task. These two algorithms have been used in a wide range of area from transaction analysis, accident data analysis, detection of statistically significant patient risk groups, discovering key person in social communities and so on. We use both the association rule learning method and the subgroup discovery method to discover useful patterns from a traffic accident dataset consisting of many features including profile of driver, location of accident, types of accident, information of vehicle, violation of regulation and so on. The association rule learning method, which is one of the unsupervised learning methods, searches for frequent item sets from the data and translates them into rules. In contrast, the subgroup discovery method is a kind of supervised learning method that discovers rules of user specified concepts satisfying certain degree of generality and unusualness. Depending on what aspect of the data we are focusing our attention to, we may combine different multiple relevant features of interest to make a synthetic target feature, and give it to the rule learning algorithms. After a set of rules is derived, some postprocessing steps are taken to make the ruleset more compact and easier to understand by removing some uninteresting or redundant rules. We conducted a set of experiments of mining our traffic accident data in both unsupervised mode and supervised mode for comparison of these rule based learning algorithms. Experiments with the traffic accident data reveals that the association rule learning, in its pure unsupervised mode, can discover some hidden relationship among the features. Under supervised learning setting with combinatorial target feature, however, the subgroup discovery method finds good rules much more easily than the association rule learning method that requires a lot of efforts to tune the parameters.

• Journal of Intelligence and Information Systems
• /
• v.21 no.4
• /
• pp.93-110
• /
• 2015

The recommender system is a system which recommends products to the customers who are likely to be interested in. Based on automated information filtering technology, various recommender systems have been developed. Collaborative filtering (CF), one of the most successful recommendation algorithms, has been applied in a number of different domains such as recommending Web pages, books, movies, music and products. But, it has been known that CF has a critical shortcoming. CF finds neighbors whose preferences are like those of the target customer and recommends products those customers have most liked. Thus, CF works properly only when there's a sufficient number of ratings on common product from customers. When there's a shortage of customer ratings, CF makes the formation of a neighborhood inaccurate, thereby resulting in poor recommendations. To improve the performance of CF based recommender systems, most of the related studies have been focused on the development of novel algorithms under the assumption of using a single profile, which is created from user's rating information for items, purchase transactions, or Web access logs. With the advent of big data, companies got to collect more data and to use a variety of information with big size. So, many companies recognize it very importantly to utilize big data because it makes companies to improve their competitiveness and to create new value. In particular, on the rise is the issue of utilizing personal big data in the recommender system. It is why personal big data facilitate more accurate identification of the preferences or behaviors of users. The proposed recommendation methodology is as follows: First, multimodal user profiles are created from personal big data in order to grasp the preferences and behavior of users from various viewpoints. We derive five user profiles based on the personal information such as rating, site preference, demographic, Internet usage, and topic in text. Next, the similarity between users is calculated based on the profiles and then neighbors of users are found from the results. One of three ensemble approaches is applied to calculate the similarity. Each ensemble approach uses the similarity of combined profile, the average similarity of each profile, and the weighted average similarity of each profile, respectively. Finally, the products that people among the neighborhood prefer most to are recommended to the target users. For the experiments, we used the demographic data and a very large volume of Web log transaction for 5,000 panel users of a company that is specialized to analyzing ranks of Web sites. R and SAS E-miner was used to implement the proposed recommender system and to conduct the topic analysis using the keyword search, respectively. To evaluate the recommendation performance, we used 60% of data for training and 40% of data for test. The 5-fold cross validation was also conducted to enhance the reliability of our experiments. A widely used combination metric called F1 metric that gives equal weight to both recall and precision was employed for our evaluation. As the results of evaluation, the proposed methodology achieved the significant improvement over the single profile based CF algorithm. In particular, the ensemble approach using weighted average similarity shows the highest performance. That is, the rate of improvement in F1 is 16.9 percent for the ensemble approach using weighted average similarity and 8.1 percent for the ensemble approach using average similarity of each profile. From these results, we conclude that the multimodal profile ensemble approach is a viable solution to the problems encountered when there's a shortage of customer ratings. This study has significance in suggesting what kind of information could we use to create profile in the environment of big data and how could we combine and utilize them effectively. However, our methodology should be further studied to consider for its real-world application. We need to compare the differences in recommendation accuracy by applying the proposed method to different recommendation algorithms and then to identify which combination of them would show the best performance.

• The Korean Journal of Elementary Counseling
• /
• v.11 no.1
• /
• pp.51-69
• /
• 2012

This research is a fundamental research to provide the foundation that psychological test can be utilized efficiently. It investigated and studied the reality of the psychological test used in elementary school, teacher's degrees of understanding and utilization of psychological test, requirements for improving the psychological test that is performed currently, and the direction of the testing tools that are required by elementary school. First, the main purpose of psychological test was different. The tests that were usually utilized in elementary schools were career inventory test and aptitude test, followed by personality test, learning test, creativity test and intelligence test in order. Second, it showed that teachers read the testing manual before carrying out the psychological test and understood it individually, the degree of understanding of T score showed to be the lowest. As a factor that hampered elementary school teacher's understanding of the psychological test, there was insufficiency of the teacher's professional knowledge, and it showed that the preparatory training for teachers about the tools of psychological test was necessary in order to improve this. Teacher's degree of utilization of psychological test showed to be lower than that of understanding. As a factor that influenced this kind of utilization by teachers, the lack of methodical understanding in afterwards guidance after the testing was the biggest factor. To increase the teacher's degree of utilization of psychological test, it was investigated that advice for more specific methods of guidance after the testing was necessary. Third, cross analysis was executed to find out if there was difference in the degree of understanding and utilization of psychological test according to whether a teacher completed the education course regarding counseling or not, and as a result, there was difference in the degree of understanding regarding some tests, but there was no difference found in the degree of utilization. Besides these, the point to be complemented in psychological test that was currently executed elementary schools and the direction of the testing tools that elementary schools require were investigated.

• Korean Security Journal
• /
• no.24
• /
• pp.91-123
• /
• 2010

The threat factors available for occurrence given G20 Summit Meeting are expected leader terrorism, hostage terrorism, bomb terrorism, public facilities terrorism, and aircraft terrorism. As for the threat groups, which are expected in Korea, the North Korea, Islam extremist group, and the group such as NGO organization of being opposed to international meeting are regarded as having possibility of causing hazard. Thus, the purpose of this study is to suggest VIP Security-measure plans in the main site in preparation for G20 Summit Meeting. Accordingly, each country in the world is adopting 'the principle of Triple Ring' in common. Thus, it elicited a coping plan by 1st line(inner ring) 2nd line(middle ring) 3rd line(outer ring) based on this principle, and proposed even an opinion together that will need to be reflected in light of policy for the VIP security measures. In conclusion, as for the VIP Security-measure plans in the main site in preparation for G20 Summit Meeting, In the inner ring(safety sector), first, an intercepting measure needs to be devised for a spot of getting into and out of vehicles given the Straight Street. Second, the Walking Formation needs to be reinforced boldly in the exposed area. In the middle ring(security sector), first, the control plan needs to be devised by considering particularity of the main site. Second, there is necessity for adopting the efficient security badge operation plan that is included RFID function within security badge. In the outer ring(aid protective sector), first, there is necessity of preparing for several VIP terrorisms, of collecting information and intelligence, and of reinforcing the information collection system against terrorism under the cooperation with the overseas information agency. Second, the urgent measure training in time of emergency needs to be carried out toward security agent event manpower. Third, to maintain the certain pace in VIP motorcade, the efficient traffic control system needs to be operated. Finally, as for what will need to be reflected in light of policy for VIP security measures, first, there is necessity for allowing VIP residence to be efficiently dispersed to be distributed and controlled. Second, there is necessity for allowing impure element to misjudge or attack to be failed by utilizing diverse deception operations. Third, according to the reorganization in North Korea's Organization of the South Directed Operations, the powerful 'military-support measure' needs to be driven from this G20 Summit Meeting. For this, the necessity was proposed for further reinforcing the front back defense posture under the supervision of the Ministry of National Defense and for positively coping even with detecting and removing poison in preparation for CBR (chemical, biological, and radio-logical) terrorism.

• Journal of Intelligence and Information Systems
• /
• v.21 no.2
• /
• pp.131-150
• /
• 2015

Omission of noun phrases for obligatory cases is a common phenomenon in sentences of Korean and Japanese, which is not observed in English. When an argument of a predicate can be filled with a noun phrase co-referential with the title, the argument is more easily omitted in Encyclopedia texts. The omitted noun phrase is called a zero anaphor or zero pronoun. Encyclopedias like Wikipedia are major source for information extraction by intelligent application systems such as information retrieval and question answering systems. However, omission of noun phrases makes the quality of information extraction poor. This paper deals with the problem of developing a system that can restore omitted noun phrases in encyclopedia documents. The problem that our system deals with is almost similar to zero anaphora resolution which is one of the important problems in natural language processing. A noun phrase existing in the text that can be used for restoration is called an antecedent. An antecedent must be co-referential with the zero anaphor. While the candidates for the antecedent are only noun phrases in the same text in case of zero anaphora resolution, the title is also a candidate in our problem. In our system, the first stage is in charge of detecting the zero anaphor. In the second stage, antecedent search is carried out by considering the candidates. If antecedent search fails, an attempt made, in the third stage, to use the title as the antecedent. The main characteristic of our system is to make use of a structural SVM for finding the antecedent. The noun phrases in the text that appear before the position of zero anaphor comprise the search space. The main technique used in the methods proposed in previous research works is to perform binary classification for all the noun phrases in the search space. The noun phrase classified to be an antecedent with highest confidence is selected as the antecedent. However, we propose in this paper that antecedent search is viewed as the problem of assigning the antecedent indicator labels to a sequence of noun phrases. In other words, sequence labeling is employed in antecedent search in the text. We are the first to suggest this idea. To perform sequence labeling, we suggest to use a structural SVM which receives a sequence of noun phrases as input and returns the sequence of labels as output. An output label takes one of two values: one indicating that the corresponding noun phrase is the antecedent and the other indicating that it is not. The structural SVM we used is based on the modified Pegasos algorithm which exploits a subgradient descent methodology used for optimization problems. To train and test our system we selected a set of Wikipedia texts and constructed the annotated corpus in which gold-standard answers are provided such as zero anaphors and their possible antecedents. Training examples are prepared using the annotated corpus and used to train the SVMs and test the system. For zero anaphor detection, sentences are parsed by a syntactic analyzer and subject or object cases omitted are identified. Thus performance of our system is dependent on that of the syntactic analyzer, which is a limitation of our system. When an antecedent is not found in the text, our system tries to use the title to restore the zero anaphor. This is based on binary classification using the regular SVM. The experiment showed that our system's performance is F1 = 68.58%. This means that state-of-the-art system can be developed with our technique. It is expected that future work that enables the system to utilize semantic information can lead to a significant performance improvement.