فصلنامه پردازش علائم و داده ها
سال دوازدهم شماره 2 (پیاپی 24، تابستان 1394)

Nowadays summarizers and machine translators have attracted much attention to themselves, and many activities on making such tools have been done around the world. For Farsi like the other languages there have been efforts in this field. So evaluating such tools has a great importance. Human evaluations of machine summarization are extensive but expensive. Human evaluations can take months to finish and involve human labor that cannot be reused.In this paper, we propose a method of automatic machine summarization evaluation that is quick, inexpensive, and language-independent, that correlates highly with human evaluation, and that has little marginal cost per run. This method has the metrics of determining auto summaries’ quality, through comparing them to the summaries produced by Human (ideal summaries). These metrics measures overlapping of system summaries and human ones in number of units like n-tuples, words string and pairs of words. Certainly for semantic comparing of texts in case of review summaries, the appearance of words are not enough and using of WordNet seems to be necessary. In the proposed method words network is used with an appropriate idea and has improved evaluation results significantly. The proposed method is the first method for the Persian language. Performance measurement of the tool was done during a specified and standard procedure and the results indicate acceptable yield of it. We present this method as an automated understudy to skilled human judges which substitutes for them when there is need for quick or frequent evaluations.

Discourse is any section or part of the language used to establish communications. Discourse Relations Recognition System recognizes the relations between textual units of a discourse. Contrast is one of the relations in Persian discourse. Recognition of this relation helps produce and understand the discourse. It can be used in variant systems such as summarization systems, interpretation systems, and so on. The contrast relation can be recognized using certain contrast relation markers such as “اما” and “ولی” Nevertheless, these markers are sometimes omitted, making trouble for relation recognition. In these cases, features such as tense of verbs, word pairs, etc. should be used for recognition. To conduct the experiments, 10000 samples for the contrast recognition and other relations were collected from the Corpus of Research Center of Intelligent Signal Processing as the data set. Then, feature vector was extracted from these samples. Finally, several supervised learning methods such as Supporting Vector Machine, the k-Nearest Neighbors algorithm, Parzen-Window, and the integration of these methods, were used to categorize and recognize contrast relations. The highest accuracy was 87. 13, which belonged to the combination of category-clauses in its best shape.

Discriminative methods are used for increasing pattern recognition and classification accuracy. These methods can be used as discriminant transformations applied to features or they can be used as discriminative learning algorithms for the classifiers. Usually, discriminative transformations criteria are different from the criteria of discriminant classifiers training or their error. In this paper, for relating feature transformation criterion to classification rate, we obtain a feature transformation method using genetic algorithm where we choose fitness function as Support Vectomr Machine(SVM) classification error rate. In addition, we obtain a feature transformation method using multi-objective genetic algorithm in order to consider both between class discrimination (According to feature transformation criterion) and support vector machine classification error rate simultaneously. Experimental results on UCI dataset indicate that using both classification error and between class discrimination in feature transformation improve discriminative feature transformations performance in increasing SVM classification accuracy. Additionally, the use of feature transformation with classification error criterion increases SVM classification more than other conventional feature transformation and proposed two-objective methods.

In this paper we investigate the brain effective networks corresponding to listening to happy and sad music. The connectivity patterns between EEG electrodes were extracted using multivariable autoregressive modeling and partial directed coherence, while participants listened to musical excerpts. Both classical and Iranian music were used as stimuli. The volunteers expressed their self-assessment using software. The results were utilized for classifying the musical excerpts. The corresponding connectivity matrices illustrated variations from happy to sad music. Also the extracted parameters from specific locations of the matrices demonstrated meaningful correlations with the subjective assessments of the emotional content of music pieces. The expressed valence was positively correlated with the input indices to the frontal channels, while this correlation was negative for sad excerpts. The proposed indices can be used for identifying the emotional perception of music pieces.

Today, with the proliferation of media, especially the internet, need for automated translation process has increased noticeably. One of the most famous machine translation methods is statistical method. Parameters of statistical machine translation system are estimated using a large scale training data set (bilingual parallel corpora). but in many languages, basic need for statistical machine translation systems -large scale parallel text - is still unsolved. To resolve this problem, the proposed method is presented for improving the quality of extracted comparable corpora and thus to improve the quality of machine translation systems. We believe that very non-parallel corpora have none or few good sentence pairs; most of their parallel data exists at the sub-sentential level. So proposed system is implemented based on extracting fragment blocks from input related sentences using score calculated from special features such as fragment length, LLR score, relevance path specification in the block and translation coverage percent.According to evaluations, proposed method has very good performance. Not only it outperforms existing methods of fragment extraction in terms of precision and recall, but also adding extracted database to existing data, improved the performance of the basic machine translation system, ranging from 0.33 to 1.4 in Belu metric.

Since it was being predicted that combination of psychophysiological and ERP signals, during the detection of a guilty person's knowledge can increase the performance of integrative lie detection system toward using the separate procedures; Using the knowledge of both aspects, in this study it has been tried to determine the proper Inter-Stimulus Intervals (ISI) together with suitable sequence of stimulations in order to simultaneous recording of P300 component of brain Event Related Potentials and peripheral signals. Also a proper mock crime scenario has been designed; it has the capability of exciting the cognitive aspect of mock crime and also was capable of provoking the subject’s concerns, based on telling lie about the crime.At the next stage, after recording data from 32 participants, features from their ERP and SCR (as one of the most important peripheral signals) signals have been extracted. Then, an LDA classifier was applied on selected features which were selected by Genetic algorithm and these accuracies: 76.67%, 73.33% & 80% have been obtained for EEG, SCR and Combined data respectively.The resulted accuracies at the first show the proper quality of scenario and protocol, in synchronous stimulation and recording of both signal categories, also the improvements which have been resulted by integrated data in compare with separate ones are observable.

This study describes a text summarization system based on cognitive models. The underlying theory is about comprehension and is used to explain comprehending narrative texts. Majority of previous methods have been used statistical approaches for summarization, and this method is different as it tries to build a system based on a cognitive theory and not statistical methods. Main principle of model is that as humans read a text, they will make a mental image based on temporality, causality, intentionality, protagonists, and place. Proposed system extracts five features for each sentence and identifies the most important sentences based on the five features. The results obtained from this method were satisfactory and comparable from modern statistical techniques.

In this paper an statistical speech enhancement method based on Laplacian Mixture Model assumption for speech in order to estimation of clean speech from noisy speech has been proposed. In the proposed method, the short time discrete Fourier transform coefficients (DFT) have been estimated by using the minimum mean square error estimator. We assume the mixture of Laplacians distribution for clean speech DFT coefficients and zero-mean Gaussian distribution for the DFT coefficients of noise. The results of segmental SNR, LLR and PESQ show that the proposed method has better performance than two estimators based on Gaussian and Laplacin model.