Search

This paper addresses the problem of automatic induction of the normalized form (lemma) of regular and mildly irregular words with no direct supervision using language-independent algorithms. More specifically, two string distance metric models (i.e. the Levenshtein Edit Distance algorithm and the Dice Coefficient similarity measure) were employed in order to deal with the automatic word lemmatization task by combining two alignment models based on the string similarity and the most frequent inflectional suffixes. The performance of the proposed model has been evaluated quantitatively and qualitatively. Experiments were performed for the Modern Greek and English languages and the results, which are set within the state-of-the-art, have showed that the proposed model is robust (for a variety of languages) and computationally efficient. The proposed model may be useful as a pre-processing tool to various language engineering and text mining applications such as spell-checkers, electronic dictionaries, morphological analyzers etc.

The development of eye localization and gaze tracking systems without the constraint of dedicated hardware, has lately motivated the scientific community. A lot of "passive" methods have been proposed, however their precision is limited and significantly reduced compared to the commercial, hardware-based, eye tracking devices. In this paper we introduce an automatic, non-intrusive method for precise eye center localization in low resolution images. To this end, the proposed system utilizes radial symmetry transform in both the original eye image and a novel eye map image derived from color information. Experimental results demonstrate great accuracy and robustness even in challenging cases, achieving also significant improvement over existing methods.

The use of visual information derived from accurate lip extraction, can provide features invariant to noise perturbation for speech recognition systems and can be also used in a wide variety of applications. Unlike many current automatic lip reading systems which impose several restrictions on users, our efforts are towards an unconstrained system. In this paper we introduce a method using k-means color clustering with automatically adapted number of clusters, for the extraction of the lip area. The method's performance is improved by previously applying nearest neighbor color segmentation. The extracted lip area is morphologically processed and fitted by a best-fit ellipse. The points of interest (keypoints) of the mouth area are extracted, while a corner detector for fine tuning of mouth corners is applied. Experimental tests have shown that the algorithm works very well under natural conditions and accurate extraction of lip keypoints is feasible.

This paper describes the process of learning taxonomic relations automatically from Modern Greek economic corpora. Supervised learning (Decision trees, Support Vector Machines, Meta-learning) is applied to economic term pairs; each pair is represented through a set of statistical, semantic and syntactic features. The resulting set of feature-value vectors presents a high imbalance in the class distribution, due to the large number of term pairs that do not present a direct semantic relation. This problem is addressed using One-sided Sampling, which reduces the number of the majority class instances by removing examples that are noisy, misleading or redundant. The approach makes use of no external resources (merely an economic corpus that is annotated with elementary morphological and phrase chunking information) and limited language-dependent elements to facilitate its portability to other languages and domains. An overall f-measure of 71 % is achieved.

This article presents a novel technique for suppressing the effect of impulsive noise in the context of automatic speech recognition (ASR). The noise suppression scheme is based on the cooperation of two neural networks. The first one is responsible for the detection of corrupted cepstral vectors due to the presence of impulsive noise. The second one is dedicated to the restoration of the detected problematic vectors solely. The novelty of the method lies in the robust detection of impulsive noise regardless of the noise source and the local restoration of the feature vectors. By avoiding a global act on the original waveform as spectral subtraction and Wiener filters do, we don't inflict any distortions on an already clean part of the original waveform. Extensive experimental valuation on a spoken digit database in the presence of machine gun impulsive noise has proved the robustness of our method.

This paper reports on the evaluation of automatic prosodic phrase break assignment. We utilize two tree-structured predictors, the commonly used CART and a C4.5, to predict break placement from sequences of easy to extract shallow textual features. We are experimenting with two 500-utterance sized prosodic corpora developed by two Greek universities that originate from different domains in order to focus on the differences in prediction between generic and limited domain datasets. The evaluation shows that while the limited dataset achieves better accuracy than the generic one in the CART case, this difference is lowered with the introduction of C4.5. Minor breaks proved to be the most difficult class to predict in CART case, while we achieved a 50% of improvement with C4.5.

Nowadays, tags play an important role in the search and retrieval process in multimedia content sharing social networks. As the amount of multimedia contents explosively increases, it is a challenging problem to find a content that will be appealing to the users. Furthermore, the retrieval of multimedia contents, which can match users' current mood or affective state, can be of great interest. One approach to indexing multimedia contents is to determine the potential affective state, which they can induce in users. In this paper, multimedia content analysis is performed to extract affective audio and visual cues from different music video clips. Furthermore, several fusion techniques are used to combine the information extracted from the audio and video contents of music video clips. We show that using the proposed methodology, a relatively high performance (up to 90%) of affect recognition is obtained.