Machine Learning (ML), among other things, facilitates Text Classification, the task of assigning classes to textual items. Classification performance in ML has been significantly improved due to recent developments, including the rise of Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRUs), and Transformer Models. Internal memory states with dynamic temporal behavior can be found in these kinds of cells. This temporal behavior in the LSTM cell is stored in two different states: “Current” and “Hidden”. In this work, we define a modification layer within the LSTM cell which allows us to perform additional state adjustments for either state, or even simultaneously alter both. We perform 17 state alterations. Out of these 17 single-state alteration experiments, 12 involve the Current state whereas five involve the Hidden one. These alterations are evaluated using seven datasets related to sentiment analysis, document classification, hate speech detection, and human-to-robot interaction. Our results showed that the highest performing alteration for Current and Hidden state can achieve an average F1 improvement of 0.5% and 0.3%, respectively. We also compare our modified cell performance to two Transformer models, where our modified LSTM cell is outperformed in classification metrics in 4/6 datasets, but improves upon the simple Transformer model and clearly has a better cost efficiency than both Transformer models.

This article presents LinkZoo, a web-based, linked data enabled tool that supports collaborative management of information resources. LinkZoo addresses the modern needs of information-intensive collaboration environments to publish, manage, and share heterogeneous resources within user-driven contexts. Users create and manage diverse types of resources into common spaces such as files, web documents, people, datasets, and calendar events. They can interlink them, annotate them, and share them with other users, thus enabling collaborative editing, as well as enrich them with links to externally linked data resources. Resources are inherently modeled and published as resource description framework (RDF) and can be explicitly interlinked and dereferenced by external applications. LinkZoo supports creation of dynamic communities that enable web-based collaboration through resource sharing and annotating, exposing objects on the linked data Cloud under controlled vocabularies and permissions. The authors demonstrate the applicability of the tool on a popular collaboration use case scenario for sharing and organizing research resources.

The Web has been identified to consist of a large portion of content that cannot be crawled by general-purpose search engines because it is only generated after a valid submission to a search interface. Accessing such content, however, requires the location and identification of search interfaces. Towards the automation of this task, many approaches have been proposed that involve the manual definition of rules for the identification of query interfaces. In this paper, we propose a rule induction approach to automatically construct a set of rules by searching the most promising subspace of all possible rules with a brute-force method and information theoretic criteria. To specify the features for the rules, we initially make a descriptive analysis of Yahoo L11, a specialized dataset containing complex interfaces, which to the best of our knowledge has not been used in previous works. We perform a series of evaluations and present the rules constructed by running the algorithm on a random sample of the Yahoo L11 dataset and another dataset used in similar works. The resulting rules yield high classification accuracy in predicting the functionality of new, previously unseen forms and since humans can easily interpret them, they can be easily ported to any application as-is.

SPARQL query execution in state of the art RDF engines depends on, and is often limited by the underlying storage and indexing schemes. Typically, these systems exhaustively store permutations of the standard three-column triples table. However, even though RDF can give birth to datasets with loosely defined schemas, it is common for an emerging structure to appear in the data. In this paper, we introduce a novel indexing scheme for RDF data, that takes advantage of the inherent structure of triples. To this end, we define the Extended Characteristic Set (ECS), a schema abstraction that classifies triples based on the properties of their subjects and objects, and we discuss methods and algorithms for the identification and extraction of ECSs. We show how these can be used to assist query processing, and we implement axonDB, an RDF storage and querying engine based on ECS indexing. We perform an experimental evaluation on real world and synthetic datasets and observe that axonDB outperforms the competition by a few orders of magnitude.

Multidimensional data are published in the web of data under common directives, such as the Resource Description Framework (RDF). The increasing volume and diversity of these data pose the challenge of finding relations between them in a most efficient and accurate way, by taking into advantage their overlapping schemes. In this paper we define two types of relationships between multidimensional RDF data, and we propose algorithms for efficient and scalable computation of these relationships. Specifically, we define the notions of containment and complementarity between points in multidimensional dataspaces, as different aspects of relatedness, and we propose a baseline method for computing them, as well as two alternative methods that target speed and scalability. We provide an experimental evaluation over real-world and synthetic datasets and we compare our approach to a SPARQL-based and a rule-based alternative, which prove to be inefficient for increasing input sizes. © 2016, Copyright is with the authors.

Abstract The Data Web refers to the vast and rapidly increasing quantity of scientific, corporate, government and crowd‐sourced data published in the form of Linked Open Data, which encourages the uniform representation of heterogeneous data items on the web and the creation of links between them. The growing availability of open linked datasets has brought forth significant new challenges regarding their proper preservation and the management of evolving information within them. In this paper, we focus on the evolution and preservation challenges related to publishing and preserving evolving linked data across time. We discuss the main problems regarding their proper modelling and querying and provide a conceptual model and a query language for modelling and retrieving evolving data along with changes affecting them. We present in details the syntax of the query language and demonstrate its functionality over a real‐world use case of evolving linked dataset from the biological domain.

Abstract Virtual Worlds evolution is breaking the barriers of virtual isolation, thus allowing users to participate in geographically dispersed and culturally diverse places. At the same time, mobile agents have been established as a solid programming method for heterogeneous networking and computing environments. Our work focuses on the definition of a distributed Virtual World reference platform for enhanced users' experience. Towards interconnecting virtual worlds with mobile agents, we have further enriched the concept of a human‐like appearance avatar. We propose two distributed virtual world architectures, namely loose and tight. In parallel, we present a relevant implementation scheme along with experimental results that prove the performance enhancements achieved against the classic client/server model. Copyright © 2014 John Wiley & Sons, Ltd.

The vast amounts of data generated, exchanged and consumed on a daily basis by contemporary networks and devices renders their analysis a cumbersome procedure with inherent difficulties. On the one hand, the need for efficient Machine Learning algorithms and tools that scale on large datasets is continuously growing. On the other, parallel or distributed solutions have proven to conceal many pitfalls. The MapReduce programming model has quickly emerged as the de facto model for executing simple algorithmic tasks over huge volumes of data, since it is simple, highly abstract and efficient. However, due to its unidirectional communication model and the inherent lack of support for iterative execution, few Machine Learning algorithms can easily be implemented on MapReduce. In this paper, we present a classification rule discovery algorithm, namely RuleMR, which despite its iterative nature, can capitalize on MapReduce. In order to construct quality rules in less iterations, the algorithm exploits the distributed nature of MapReduce to explore only the promising areas in the search space. We conduct a series of experimental evaluations which indicate that the proposed approach not only scales well with respect to the size of the training dataset, but also, in many cases, the resulting model is comparable to many well known algorithms in matters of accuracy.

Legislators, designers of legal information systems, as well as citizens face often problems due to the interdependence of the laws and the growing number of references needed to interpret them. Quantifying this complexity is not an easy task. In this paper, we introduce the “Legislation Network” as a novel approach to address related problems. We have collected an extensive data set of a more than 60-year old legislation corpus, as published in the Official Journal of the European Union, and we further analysed it as a complex network, thus gaining insight into its topological structure. Results are quite promising, showing that our approach can lead towards an enhanced explanation in respect to the structure and evolution of legislation properties.

We propose a biological-inspired algorithmic procedure for improving the precision level in respect to a web user's query. This algorithm is capable of tracing relevant information in web graphs, imitating the mission of foraging ants. Our proposal is an Ant Colony Optimization (ACO) modification, however its differentiation lies in the fact that no predefined graph is employed, since the search graph is dynamically constructed during the algorithm execution. Despite some weaknesses in terms of time execution, it was evaluated that the effectiveness of the algorithm was quite satisfying in large web graph sets.

In this paper a statistical approach for estimating the evolution of categorized web page populations in web directories is proposed. The proposal is based on the capture-recapture method used in wildlife biological studies and it is modified accordi

In the context of this paper, a meta-search third-party Web result ranking mechanism is proposed, which is capable of adapting over the continuous changes that occur on the Web, providing in parallel personalized information acquisition considering the user's navigation behaviour. The designed meta search engine service rates, re-organises and combines the results acquired from search services for a specific user information resource request in accordance with a weighted combination of a performance related factor (tightly related to the ranking of the Web result as given by the search engine service) and a reliability related factor (corresponding to the user's satisfaction stemming from the specific Web result that he/she browses), while the performance of each search engine with respect to adequately adapting to the Web evolution is taken into account. Assuming a group of users falling within the same category with respect to the information/resource needs, the Web result reliability rating system is collaborative in the sense that it considers both first-hand information (acquired from the user's past experiences with the search engine services) and second-hand information (corresponding to other users' experiences with search engine services), while the matching degrees of the users' profiles have been taken into account. A set of results indicative of the efficiency of our proposed scheme is provided. Transparency is achieved for both personalization and Web evolution adaptation mechanisms, requiring virtually none effort from the user's part.

License plate recognition (LPR) algorithms in images or videos are generally composed of the following three processing steps: 1) extraction of a license plate region; 2) segmentation of the plate characters; and 3) recognition of each character. This task is quite challenging due to the diversity of plate formats and the nonuniform outdoor illumination conditions during image acquisition. Therefore, most approaches work only under restricted conditions such as fixed illumination, limited vehicle speed, designated routes, and stationary backgrounds. Numerous techniques have been developed for LPR in still images or video sequences, and the purpose of this paper is to categorize and assess them. Issues such as processing time, computational power, and recognition rate are also addressed, when available. Finally, this paper offers to researchers a link to a public image database to define a common reference point for LPR algorithmic assessment.

This paper proposes an innovative method of monitoring the ability of Internet search services to maintain new and up-to-date results in their directories. The method is based on capture-recapture experiments performed in real wildlife biological studies. After the necessary modifications needed for conducting similar experiments on the Web, a sixteen-week period of capture-recapture experiments was made over the caches of Google and MSN. During this period the birth as well as the survival rates of Google's and MSN's cached results were estimated based on the Pollock's robust design method. The paper provides the implementation details of the applied model along with its initial assessment. The experimental results showed that this approach can be applied as a method for measuring the ability of Internet search services in terms of maintaining new and up-to-date results in their caches.

In this paper, we describe a capture-recapture experiment conducted on Google's and MSN's cached directories. The anticipated outcome of this work was to monitor evolution rates in these web search services as well as measure their ability to index and maintain fresh and up-to-date results in their cached directories.

This paper proposes a statistical approach for estimating the evolution of Web pages in directories. The proposal is based on the capture-recapture method used in wildlife biological studies in an animal, bird or fish populations, and it is modified according to the necessary assumptions and amendments for applying the experiments in a search engine directory. During these experiments, Web pages are considered as animals and the specific types of Web pages as particular species of animals whose abundance, birth, death and survival rates are estimated. The population is open, meaning that new Web pages are submitted to the search engine directory, while others are removed from the directory indexes, resembling to emigration/immigration processes in nature. The role of the biologist who recognizes the species under study and records their history is assigned to a Web page classifier, which is trained under the open directory's (DMOZ project) taxonomy. The classifier is a three layer probabilistic neural network capable of identifying and categorizing Web pages, on the basis of information filtering. A virtual experiment is simulated based on the classifier performance over real Web pages, while the results are quite promising

In this paper, a new algorithm for vehicle license plate identification is proposed, on the basis of a novel adaptive image segmentation technique (sliding concentric windows) and connected component analysis in conjunction with a character recognition neural network. The algorithm was tested with 1334 natural-scene gray-level vehicle images of different backgrounds and ambient illumination. The camera focused in the plate, while the angle of view and the distance from the vehicle varied according to the experimental setup. The license plates properly segmented were 1287 over 1334 input images (96.5%). The optical character recognition system is a two-layer probabilistic neural network (PNN) with topology 108-180-36, whose performance for entire plate recognition reached 89.1%. The PNN is trained to identify alphanumeric characters from car license plates based on data obtained from algorithmic image processing. Combining the above two rates, the overall rate of success for the license-plate-recognition algorithm is 86.0%. A review in the related literature presented in this paper reveals that better performance (90% up to 95%) has been reported, when limitations in distance, angle of view, illumination conditions are set, and background complexity is low.