Footprints

Vagueness is a common human knowledge and linguistic phenomenon, typically manifested by predicates that lack clear applicability conditions and boundaries such as High, Expert or Bad. In the context of ontologies and semantic data, the usage of such predicates within ontology element definitions (classes, relations etc.) can hamper the latter’s quality, primarily in terms of shareability and meaning explicitness. With that in mind, we present in this paper a vague word sense classifier that may help both ontology creators and consumers to automatically detect vague ontology definitions and, thus, assess their quality better.

Given a set of texts discussing a particular entity (e.g., customer reviews of a smartphone), aspect based sentiment analysis (ABSA) identifies prominent aspects of the entity (e.g., battery, screen) and an average sentiment score per aspect. We focus on aspect term extraction (ATE), one of the core processing stages of ABSA that extracts terms naming aspects. We make publicly available three new ATE datasets, arguing that they are better than previously available ones. We also introduce new evaluation measures for ATE, again arguing that they are better than previously used ones. Finally, we show how a popular unsupervised ATE method can be improved by using continuous space vector representations of words and phrases.

Valencia, Spain

Atlanta, Georgia, U.S.A.

"This paper describes the systems with which we participated in the task Sentiment Analysis in Twitter of SEMEVAL 2013 and specifically the Message Polarity Classification. We used a 2-stage pipeline approach employing a linear SVM classifier at each stage and several features including BOW features, POS based features and lexicon based features. We have also experimented with Naive Bayes classifiers trained with BOW features."

We propose a framework for predicting the ranking position of a Web page based on previous rankings. Assuming a set of successive top-k rankings, we learn predictors based on different methodologies. The prediction quality is quantified as the similarity between the predicted and the actual rankings. Extensive experiments were performed on real world large scale datasets for global and query-based top-k rankings, using a variety of existing similarity measures for comparing top-k ranked lists, including a novel and more strict measure introduced in this paper. The predictions are highly accurate and robust for all experimental setups and similarity measures.

The goal of this study is the development of an assessment system with the support of a Neural Network approach optimized with the use of Genetic Programming. The data used as training data are real data derived from an educational project. The developed system is proved capable of assessing data from both single select and multiple choice questions in an e-learning environment. The final result is the assessment of the learners’ answers through various criteria.

As resulted from an ongoing collaboration with Panos Alexopoulos

I designed a statistical model to measure recursively the strength of nodes in an argumentation graph. This was integrated to an ontology built by George Anadiotis & Adam Westerski.