فصلنامه پردازش علائم و داده ها
سال نهم شماره 2 (پیاپی 18، پاییز و زمستان 1391)

Valency is the key concept in dependency grammar. Among different lexical categories, verbs play the most important role in syntax and semantics of the sentence in dependency grammar. Verbs are known to be the central point of sentences and reveal the main concept of the sentence. In this article, by investigating different methods of unsupervised Persian verb valence induction, we discuss specific issues in recognizing verbs in linguistics terms and ambiguities in Persian verb valence induction. We propose possible solutions for the mentioned challenges. According to our experiments, Expectation Maximization (EM) algorithm has the highest performance among all induction methods in Persian. We show that the performance of EM algorithm is more than twice as of the binomial hypothesis testing.

In this paper, a new approach is presented for improving image quality which it provides a new outlook on how to apply the enhancement methods on images. Image enhancement techniques may be performed based on the illumination, resolution, or distribution. In image capturing, issues such as the illumination of the scene and reflectance of objects affect on image captures. The effect of each of these issues on the resulting image can be considered as a separate component or image and they can be reviewed using homomorphic filter where the original image is decomposed into illumination and reflectance components. Generally, the pixels value of an image is proportional to the illumination of point in the scene and the reflectance of the object. Based on this, impairment in the image may be due to each of the illumination or reflectance images. In this paper it is shown that various types of impairments have a different effect on the illumination and reflectance of image components. Studies show that impairment effect on an image depending on the type of the impairment on one component is different to another component. Unlike conventional methods which dose enhancement process on the original image for any type of impairment, in this paper it is suggested that based on type of impairments, enhancement of image is done on corresponding components after image decomposition into illumination and reflectance components. Hence final improved image can be achieved using image components that they include improved component. Results of this research show that image components enhancement has better results comparing to directly enhancement of original image to eliminate the image impairment. In this paper, it is used the Yale Face B database for facial images and different gray level images for non facial images.

In recent years, many efforts have been done to design ontology learning methods and automate ontology construction process. The ontology construction process is a time-consuming and costly procedure for almost all domains/applications, so automating this process is a solution to overcome the knowledge acquisition bottleneck in information systems and reduce the construction cost. In this article a novel intelligent ontology learning method is proposed which can be used in many domains and applications. The proposed learning system has no need for initial common or specialized input ontologies or predefined semantic terms; indeed, the initial database anonly consists of input texts sets. The proposed learning system could extract associated ontologies of various domains using combined methods. To do this, a combination of linguistic, statisticaland machine learning methods based on the C-value method, the TF-IDF one, the neural network, and co occurance analysis are applied. So, first domain-related documents were collected. Then natural language processing methods such as C-value method were implemented for extracting meaningful terms from documents. Next, ART (Adaptive Resonance Theory) neural network was used to cluster documents and associated weight of terms was calculated by TF–IDF method in order to find candidate keyword for each cluster. Finally, co-occurrence analysis was used to construct concept hierarchy and complete the ontology. Results show that the proposed ontology learning method has a high precision comparing to similar studies.

In this study, Recurrence Plots (RPs) has been applied as a nonlinear approach in order to detection of guilty subjects’ knowledge based on their single-trial ERPs. The dataset were recorded from 49 human subjects who were participated in a Concealed Information Test (CIT). According to the test protocol, guilty subjects denied their information about familiar faces, so the aim was to detect the concealed faces in these subjects. Recurrence Quantification Analysis (RQAs) were employed in feature extraction stage. Brain's dynamic and it's trajectories in phase space are two important issues that can be indicated in these measures. Results demonstrate, that the appearance of P300 signals in guilty subjects (because of denying a familiar face), can increase the determinism and predictability of their brain’s signals. Also using Genetic Algorithm (GA) in feature selection level together with Linear Discriminant Analysis (LDA) classifier and a new method named: “variable threshold”, we achieved an accuracy about 89.7% on combining information of Fz, Cz and Pz channels.

The aim of this article is design a lie detector system using GSR and PPG. The data set was including of photoplethysmograph signals and galvanic skin response record through an inductive test and using classic polygraph device. Thenceforth, features of time and frequency were extracted. Consequently data were classified and accuracy sensitivity coefficients were calculated by applying these features to linear Discriminant analysis LDA, MLP and Elman neural networks. 20 people participated in the study. The mean age of participants was 36 years. Finally determination of lie was done with accuracy coefficient of 88% by applying Elman neural network. According to the findings in this study, the new method has introduced while offering a more comfortable recording and less diagnostic costs. This new method can be suggested for use as a lie screening system

In this paper, we fit a function on probability density curve representing an information stream using artificial neural network. This methodology results in a specific function which represents a memorize able probability density curve. We then use the obtained function for information compression by Huffman algorithm. The difference between the proposed method with the general methods is, to use the Huffman algorithm in several times. In every time, the probability density function is fitted, estimated and then the information representing the function is added to end of the information stream. We next propose two different algorithms for information encoding and decoding using time variable estimation of probability density function. In order to evaluate the proposed algorithm, the percentage of the compression resulting our method has been compared with two popular methods named FDR code and Golomb at the end.