Gengfeng Wu

Sentiment classification using phrase patterns

This paper presents a phrase pattern-based method in classifying sentiment orientation of text. That is to analyze whether the text expresses a favorable or unfavorable sentiment for a specific subject. In our method, we construct some phrase patterns and calculate their sentiment orientation by unsupervised learning algorithm. When we classify a document, we first add special tags to some words in the text, then match the tags within a sentence with some phrase patterns to get the sentiment orientation of the sentence. At last, we add up the sentiment orientation of each sentence. We classify the text according to this summation. The method achieves an accuracy rate of 86% when used to evaluate sports reviews from some Websites.

Using Bilingual Lexicon to Judge Sentiment Orientation of Chinese Words

It is a challenging task to identify sentiments (the affective parts of options) of reviews. One of the most important problems is to predict the sentiment orientation of the words. This paper proposes a new method for determining the sentiment orientation of the Chinese words by using bilingual lexicons. Given a Chinese word, we observe the occurrences of English sentiment words in its interpretations, to predict the sentiment orientation of the Chinese word. The whole process can be illustrated logically as follows: (1) translate a Chinese word into Chinese-English interpretation; (2) generate an English word sequence by parsing the interpretation; (3) calculate the sentiment vector from the English word sequence; (4) use a classifier to predict the sentiment orientation for the Chinese word. The performance of two kinds of classifiers (SVM and C4.5) is studied. The experiments show that the proposed method performed well and achieved high accuracy.

Ontology Based User Profiling in Personalized Information Service Agent

Personalized information service agents have emerged in the recent years to help users to cope with the increasing amount of information available on the Internet. The effectiveness of agents depends mainly on profile completeness and accuracy. In the existing agents, although the performance of these systems improves after learning a user profile, it is difficult to share user profile and adapt user profile to user interests. In order to solve these problems in agents, we present ontology based user profiling methods. The approaches in user profiling, such as representation, acquisition, learning, adaptation and re-ranking, are discussed. Moreover, a personalized information service agent is designed based on these approaches, and the performance of agent is evaluated.

3D triangle mesh smoothing via adaptive MMSE filtering

This paper introduces an effective mesh smoothing method for 3D noisy shapes via the adaptive MMSE (minimum mean squared error) filter. The adaptive MMSE filter is applied to modify the face normals of triangle meshes and then mesh vertex positions are reconstructed in order to satisfy the modified normals. We also compare quantitatively and visually the adaptive MMSE filter to 3D triangle meshes with the conventional and simple Laplacian smoothing and mean and median filtering schemes. The experiments demonstrate that the adaptive MMSE filter applied to triangle meshes with round shapes and low frequency noise outperforms the existing smoothing schemes mentioned above.

Correlation Between Forehead EEG and Sensorimotor Area EEG in Motor Imagery Task

Electrodes connection on the scalp needs to apply gel or paste on the scalp and fit EEG-cap on the head and this procedure also needs to deal with the hair on it. By comparison, to fit EEG electrodes on the forehead area is much easier because there is no hair on it. If correlations of the EEGs generated from the forehead area are high with respect to EEGs from the sensorimotor area, it is then possible to achieve relatively high classification accuracy in motor imagery tasks just using EEG from forehead channels. In this way, it will help to make the procedure of motor imagery tasks much easier and convenient. Because correlation coefficients is often used to measure the similarity of two signals, it is necessary to study whether there is a high correlation between forehead channels’ EEG and EEG from sensorimotor area during MI(Motor Imagery) tasks. In this paper, EEG data from three subjects were used in the tests. Firstly, a test was conducted on the correlation between EEGs from forehead 8(Fp1-Af8) channels and EEGs from the sensorimotor area channel C3, C4 during the MI tasks. The correlation is calculated with respect to ERP (Event-Related-Potential), Spectral Power between 6-25Hz, and ERSP (Event-Related-Spectral-Perturbation) at Alpha rhythms 8-12Hz. The results of the correlation tests are mostly above 70%. In particular, for subject Sl1, the correlation coefficient of ERSP between forehead channels and C3, C4 are as high as 0.9 during the left hand movement imagery trials. Secondly, we did a test on the classification of imagined left/right hand movement tasks using EEGs from 8 electrodes in the forehead. Classification results show that the accuracy of the forehead 8 channels’ EEGs are as high as 81% for subject Sk6 comparing to 90% using 29 channels’ EEG signal neighboring to C3, C4. For subject Sk3 and subject Sl1, the accuracies are 65% and 79% comparing to 80% and 83% using EEG signals from 29 channels neighboring to C3, C4. So there are high correlation between EEGs from the forehead area and EEGs from the sensorimotor area. That is to say, we can use EEGs from forehead in some situations, such as classification of left/right imagery, because they are much easier to measure than EEG form the sensorimotor area. This will make BCI system more portable and more convenient to use.

Extraction of if-then rules from trained neural network and its application to earthquake prediction

This paper presents a supervised ART neural network named impulse force based ART (IFART) neural network. It enhances the prediction accuracy of the supervised ART neural network using impulse forces on attributes optimized by genetic algorithm, which identify the different effect of input attributes on category results. However, the IFART neural network is still a black box and difficult to understand, which is the disadvantage of artificial neural network. In this paper, a method to extract if-then rules from the trained IFART neural network according to its architecture is proposed to interpret the neural network. Furthermore, the rules are refined in terms of their used frequency. Finally, IFART neural network is applied to predict the magnitude of earthquake.

A Personalized Semantic Search Method for Intelligent e-Learning

Leukocyte migration is an important phenomenon in the inflammatory tissue. The migration process includes the rolling velocity decreasing and the leukocytes adhesion. However, the analysis of in vivo microscopy video is a labor-intensive and time consuming task. Several approaches have been proposed for tracking leukocyte movements. However, these approaches can either only track leukocytes that roll along the centerline of the blood vessel, or can only handle leukocytes with fixed morphologies. In addition, the camera/subject movement is a severe problem which occurs frequently while analyzing in vivo microscopy videos. In this paper, we proposed a new method for automatic recognition of non-adherent and adherent leukocytes. The experimental results demonstrate the effectiveness of the proposed method.In intelligent e-Learning, personalized search support becomes even more and more important now. This paper shows how to achieve the personalized search in e-learning environments. We present a semantic-based search method for personalized e- Learning. In the method, learning resources ontology and learner ontology are designed for semantic analysis and algorithm for ontology learning. The method is applied in an e-learning system and evaluated by some users. The result shows that the personalized semantic search method can improve the precision and recall of learning resources retrieval effectively in e-learning system, and can provide more intelligent e-learning services for the learners.

Walking Stability by Age A Feature Analysis Based on a Fourteen-Linkage Model

More and more researchers have been interested in walking analysis for health management of elderly people. Before understanding pathological gait, it is necessary to understand normal gait, since this provides a quantitative standard to judge or evaluate the gait of a patient. A starting point to do this is to study the relationship between walking features and ages. For this purpose, this paper proposes a walking model called Fourteen-Linkage Walking (FLW) model to detect some useful features relative with age from 21 to 65 year old. The features of walking stability are then defined and extracted based on the FLW model. Two kinds of stability features, footprint features and walking cycle features, are researched in this paper. To evaluate the walking pattern, a new concept, stability coefficient, is suggested. Our results show that the footprint and walking cycle features have close correlation with ages.

Earthquake Prediction by RBF Neural Network Ensemble

Earthquake Prediction is one of the most difficult subjects in the world. It is difficult to simulate the non-linear relationship between the magnitude of earthquake and many complicated attributes arising the earthquake. In this paper, RBF neural network ensemble was employed to predict the magnitude of earthquake. Firstly, the earthquake examples were divided to several training sets based on Bagging algorithm. Then a component RBF neural network, which was optimized by Adaptive Genetic Algorithm, was trained from each of those training sets. The result was obtained by majority voting method, which combined the predictions of component neural networks. Experiments demonstrated that the prediction accuracy was increased through using RBF neural network ensemble.

Super Parsing:Sentiment Classification with Review Extraction

This paper describes the sentiment classification with review extraction. Whole process can be illustrated logically as: (1) extract the review expressions on specific subjects and attach sentiment tag and weight to each expression; (2) calculate the sentiment indicator of each tag by accumulating the weights of all the expression with the corresponding tag; (3) given the indicators on different tags, use a classifier to predict the sentiment label of the text. A system approximate text analysis (ATA) is used for review extraction in stage 1. It follows the idea of super parsing, which enables non-adjacent constituents to be merged to deduce a new one. To traverse the valid constituent combinations in super parsing, an algorithm named candidate list algorithm (CLA) is proposed. Then the performance of three kinds of classifiers (a simple linear classifier, SVM and decision tree) in stage 3 is studied. The experiments on on-line documents show that the SVM algorithm achieves the best performance