Antonis Koukourikos

Collaborative Filtering Recommendation of Educational Content in Social Environments Utilizing Sentiment Analysis Techniques

Collaborative filtering techniques are commonly used in social networking environments for proposing user connections or interesting shared resources. While metrics based on access patterns and user behaviour produce interesting results, they do not take into account qualitative information, i.e. the actual opinion of a user that used the resource and whether or not he would propose it for use to other users. This is of particular importance on educational repositories, where the users present significant deviations in goals, needs, interests and expertise level. In this paper, we examine the benefits from introducing sentiment analysis techniques on user-generated comments in order to examine the correlation of an explicit rating with the polarity of an associated text, to retrieve additional explicit information from user comments when a standard rating is missing and expand tried recommendation calculation with qualitative information based on the community’s opinion before proposing the resource to another user.

Towards Machines for Measuring Creativity: The Use of Computational Tools in Storytelling Activities

Until now, significant R&D effort in the field of Computational Creativity, has been devoted towards understudying the elements of the creative process from idea conception to production, or towards designing machines which exhibit human level creativity without merely mimicking the human creative process. However, in the effort to determine if an artefact is creative by human standards, it is also important to examine the perception of creativity by humans and to which extend this perception can be formalized and applied on the evaluation of creative works. In this paper, we investigate how the human perception for creativity exhibited in text artefacts can be predicted using appropriate formulations of computational creativity metrics. To this end, we designed and executed a storytelling experiment assisted by the usage of computational tools. We subsequently exploited human-provided rankings of the stories in order to train a model for evaluating creativity as a combination of various characteristics of the produced stories.

From Computational Creativity Metrics to the Principal Components of Human Creativity

Within the field of Computational Creativity, significant effort has been devoted towards identifying variegating aspects of the creative process and constructing appropriate metrics for determining the degree that an artefact exhibits creativity with respect to these aspects. However, in the effort to determine if an artefact is creative by human standards, it is also important to examine the perception of creativity by humans and to which extend this perception can be formalized and applied on the evaluation of creative works. In this paper, we investigate how the human perception for creativity exhibited in text artefacts can be correlated by the usage of appropriate formulations of computational creativity metrics. To this end, we propose a model for transitioning from traditional metrics to a space that adheres to the principal components of human creativity and reflects the way that human approach the assessment of the creative process.

Exploiting unstructured web information for managing linked data spaces

The interlinking, maintenance and updating of different Linked Data repositories is steadily becoming a critical issue as the amount of published data increases. Hence, the automation or the provision of substantial computational support in various phases of the management process is a particularly important topic. The present paper discusses the main considerations of managing and evolving Linked Data repositories and proposes an experimental methodology, based on the usage of unstructured online information, for handling the issues emerging from the Linked Data management process.

Data-Driven Schema Matching in Agricultural Learning Object Repositories

As the wealth of structured repositories of educational content for agricultural object is increasing, the problem of heterogeneity between them on a semantic level is becoming more prominent. Ontology matching is a technique that helps to identify the correspondences on the description schemas of different sources and provide the basis for interesting applications that exploit the information in a linked fashion. The present paper presents a data-driven approach for discovering matches between different classification schemas. The approach is based on content analysis and linguistic processing in order to extract information in the form of relation tuples, use the extracted information to associate the content of different repositories and match their underlying classification schemas based on the degree of content similarity. The preliminary results verified the validity of the approach, as both experiments produced a semantically valid matching in 68% of the examined classes. The results also exposed the need for refinements on the linguistic processing of the available textual information and on the definition of relation similarity, as well as, the need to exploit structural information in order to move from discovering semantically valid matches to effectively handling class specializations and generalizations.

Improving the real-time performance of heterogeneous extremely large datasets

The POWDER protocol is a Semantic Web technology --and W3C Recommendation- that takes advantage of natural groupings of URIs, as identifiers as well as navigational paths, to annotate all the resources in a regular expression-delineated sub-space of the URI space. POWDER was designed as a mechanism for accreditation, trustmarking and resource discovery, emphasizing the publishing of attributed metadata by third parties and trusted authorities. However, its versatility allows the application of POWDER in different use cases such as repository compression. In this paper, we present the POWDER protocol, briefly discuss current implementations and use cases and present how POWDER can be implemented over existing well-tested and robust semantic storage systems. Furthermore, we discuss a novel solution for the scalable storing data summaries in the form of metadata for the purposes of source selection and source schema coordination in large-scale, heterogeneous federations of semantic querying endpoints. Our solution takes advantage of POWDERs ability to exploit naming conventions and other natural groupings of URIs in order to compress instance-level metadata about the nodes of a data service federation, especially in situations where URI hashing cannot be used to efficiently resolve the sources that hold statements regarding a given URI resource.

Cross-Language Ontology Alignment Utilizing Machine Translation Models

In the context of ontology alignment, linguistic analysis is a prominent solution, used by various proposed methodologies. When mapping ontologies that use the same language, the existent approaches have been shown to produce significant results, being able to handle complex descriptions of the enclosed concepts and properties. In order to expand the applied linguistic methods in a cross-language context, i.e. to align ontologies that use different languages, it is essential to automate the process of finding lexical correspondences, beyond simple term translation, between the entity descriptions provided by the involved ontologies. The present paper proposes a machine learning approach to obtain the optimal from a set of translation provided by different automated machine translation services, in order to use it as the basis for aligning ontology pairs that provide complex descriptions expressed in different languages.

Using Open Information Extraction and Linked Open Data towards Ontology Enrichment and Alignment

The interlinking, maintenance and updating of different Linked Data repositories is steadily becoming a critical issue as the amount of published data increases. The wealth of information across the World Wide Web can be exploited in order to provide additional information about the way that an object is described in the real world. This paper proposes a method for discovering new concepts and examining the equivalence of properties in different LOD description schemas by using Open Information Extraction techniques on web resources. The method relies on constructing association graphs from the extracted information, proceeding to a transfer on the conceptual level using information previously known from the LOD repositories and examining the similarities and discrepancies between the produced graphs and the LOD descriptions, as well as between the graphs derived from different repositories.