Seong Joon Yoo

Senti-lexicon and improved Naïve Bayes algorithms for sentiment analysis of restaurant reviews

The existing senti-lexicon does not sufficiently accommodate the sentiment word that is used in the restaurant review. Therefore, this thesis proposes a new senti-lexicon for the sentiment analysis of restaurant reviews. When classifying a review document as a positive sentiment and as a negative sentiment using the supervised learning algorithm, there is a tendency for the positive classification accuracy to appear up to approximately 10% higher than the negative classification accuracy. This creates a problem of decreasing the average accuracy when the accuracies of the two classes are expressed as an average value. In order to mitigate such problem, an improved Naive Bayes algorithm is proposed. The result of the experiment showed that when this algorithm was used and a unigrams+bigrams was used as the feature, the gap between the positive accuracy and the negative accuracy was narrowed to 3.6% compared to when the original Naive Bayes was used, and that the 28.5% gap was able to be narrowed compared to when SVM was used. Additionally, the use of this algorithm based on the senti-lexicon showed an accuracy that improved by a maximum of 10.2% in recall and a maximum of 26.2% in precision compared to when SVM was used, and by a maximum of 5.6% in recall and a maximum of 1.9% in precision compared to when Naive Bayes was used.

Intelligent Multimedia Information Retrieval for Identifying and Rating Adult Images

We applied an intelligent multimedia information retrieval technique to devise an algorithm identifying and rating adult images. Given a query, ten most similar images are retrieved from an adult image database and a non-adult image database in which we store existing images of each class. If majority of the retrieved are adult images, then the query is determined to be an adult image. Otherwise, it is determined to be a non-adult class. Our experiment shows 99% true positives with 23% false positives with a database containing 1,300 non-adult images, and 93.5% correct detections with 8.4% false positives when experimented with a database containing 12,000 non-adult images. 9,900 adult images are used for both experiments. We also present an adult image rating algorithm which produces results that can be used as a reference for rating images.

Neural network based adult image classification

Digital multimedia data is dramatically being increased everyday since the Internet became popular. This increment in multimedia data increases adult image contents to the Internet as well. Consequently, a large number of children are exposed to these X-rated contents. In this paper, we propose an efficient classification system that can categorize input images into adult or non-adult images. The simulation shows that this system achieved 95% of the true rate whereas it reduces the false positive rate below 3%.

Composition of MPEG-7 visual descriptors for detecting adult images on the internet

We describe a similarity based adult image detection technique (SID) exploiting composed MPEG-7 visual descriptors. The technique with a large set of training adult image database and a smaller set of training non-adult image database is practically useful in detecting adult images with little false negatives. SID achieved 99% correct detections with 23% false positives when experimented with a database containing 1,300 training non-adult images, and 93.5% correct detections with 8.4% false positives when experimented with a database containing 12,000 training non-adult images. 9,900 training adult images are used for both experiments. Given a query, ten most similar images are retrieved. If majority of the retrieved are adult images, then the query is determined to be an adult image. Otherwise, it is determined to be a non-adult image. SID can detect adult Internet content with the aid of text filtering system as described in the later section.

Modeling Web Crawler Wrappers to Collect User Reviews on Shopping Mall with Various Hierarchical Tree Structure

Along with the increased number of internet shopping mall users, high quantities of reviews on products are often found in many shopping malls. In order to extract effective information from those reviews, many studies on opinion mining have been actively performed. Due to the various type of structure of shopping malls, it is difficult to apply a single web crawler on all the shopping malls, so proper web crawler models need to be implemented for each shopping mall. The core technique of constructing the appropriate web crawler is the Wrapper, and in this study, wrapper models for product reviewing web crawlers are suggested, designed, and implemented for four large shopping malls.

Accessing Positive and Negative Online Opinions

Nowadays, an increasing number of people review the comments on each item before they will purchase the commodities and services offered by online shopping malls, Internet blogs, or cafes. However, it is somewhat challenging to routinely read trough all of the comments. The purpose of this study is to introduce some methods to classify the positive or negative review pertaining to the blog comments on a movie written in Korean. For this purpose, a variety of algorithms was used to classify the reviews and allow feature-selection by applying the traditional machine learning method for classifying literature.

Detecting adult images using seven MPEG-7 visual descriptors

In this paper we introduce an effective method of the adult image classification via MPEG-7 descriptors. The proposed system uses MPEG-7 descriptors as the main feature of the adult image classification systems. The simulation shows that the proposed image classification system performs the 5 class classification task with success rate of above 70%.

BLITE-SVR: New forecasting model for late blight on potato using support-vector regression

Various simple statistical methods have been used for the prediction of plant-disease epidemics. However, the need to develop a new model, reflecting many changed environmental factors and applicable to the Korean domestic farmhouse, has been raised. Given this point, we developed the potato late blight prediction model called BLITE-SVR, after which we predicted and verified the first date of occurrence with the data from 1976 to 1985 and from 2009 to 2012 through support-vector regression (SVR), a statistical method offering good performance. For the prediction model, we collected 13 kinds of weather data, including temperature, humidity, evaporation, and so on, which displayed very high correlation to the first date of the occurrence of late blight. The performance of BLITE-SVR has been evaluated through comparison with the conventional moving-average method that was previously used, as well as through pace regression and linear regression. The accuracy of prediction for the first date of occurrence was 64.3% by BLITE-SVR, thus showing a higher degree of accuracy compared with 42.9% by the conventional moving-average method, 42.9% by pace regression and 35.7% by linear regression. This study will enable farmers to match the targeted fungicide application to the time of greatest need and thereby achieve a reduction in chemical use.

Real-Time Face-Detection Engine for Robustness to Variable Illumination and Rotated Faces

In this paper, we proposes a novel hardware architecture and FPGA implementation method of high performance real-time face-detection engine for robustness to variable illumination and rotation. The proposed face detection algorithm improved its performance by using MCT (Modified Census Transform), rotation transformation and AdaBoost learning algorithm. For implementation, we used a QVGA class camera, LCD display, and Virtex5 LX330 FPGA made by Xilinx Corporation. The verification results showed that it is possible to detect at least 32 faces in a wide variety of sizes at a maximum speed of 43 frames per second in real time. This finding can be applied to artificial intelligence robots for human recognition, conventional security systems for identity certification, and cutting-edge digital cameras using image processing techniques.

Rules for Mining Comparative Online Opinions

The study of comparative online opinions is about sorting comparative sentences out of given sentences. This study, which is focused on the documents in Korean, may be the first of its kind in Korea although there have been a few of such studies in English spoken countries. In this study, 39 words –the most frequently used in the comparative sentences– were identified for the extraction of comparative sentences; especially, of the 39 words, this study is focused on the word ‘boda’, which is the most frequently occurring in Korean comparative sentences in identifying the rules for distinguishing the comparative sentences. The Korean word ‘boda’ is a proposition that has the same role as the English word, ‘than’; and if used as an adverb, ‘more’. In total, 11 rules were found in the observation of commodity review documents in blogs using the word ‘boda’. The study might be applied to a comparative search, on the internet, of a commodity or other object as well as an elementary technology of Opinion Mining.