Czeslaw Jedrzejek

Usage of the Jess Engine, Rules and Ontology to Query a Relational Database

We present a prototypical implementation of a library tool, the Semantic Data Library (SDL), which integrates the Jess (Java Expert System Shell) engine, rules and ontology to query a relational database. The tool extends functionalities of previous OWL2Jess with SWRL implementations and takes full advantage of the Jess engine, by separating forward and backward reasoning. The optimization of integration of all these technologies is an advancement over previous tools. We discuss the complexity of the query algorithm. As a demonstration of capability of the SDL library, we execute queries using crime ontology which is being developed in the Polish PPBW project.

Extended rules in knowledge-based data access

We present a method for an efficient knowledge-based access to relational data. Knowledge is represented as a set of rules (basic rules) and describes a source data at concept (ontological) level. Forward chaining in the integrated system is performed with extended rules, which are obtained by a goal- and dependency-directed transformation of the basic rules. The novel feature of our method is generality - every rule is generated so that includes all possible binding of the head predicates, and variable dependencies, while in many implementations of the magic method the succession of bindings depends on a query. We demonstrate a query answering algorithm and our prototypical implementation of the system coupled with the Jess engine. The results of performance evaluation are presented and compared to the results described in our previous works.

Towards ontology of fraudulent disbursement

The task to ontologically model the knowledge concerning the selected class of economic crimes is considered; particularly we focus on fraudulent disbursement. The ontology has a layered structure with the foundational ontology (constructive descriptions and situations) on the structures top and the application ontology at the structures bottom. The application level entities were manually separated from the motivating crime scenarios, having a domain- and a task-based parts. Domain-based ontology contains descriptions of attributes and relations of the domain while the task-based part, designed to support the knowledge extraction from databases, is implemented via rules that are used to extract data about documents and their attributes, transactions, engaged people actions and their legal qualifications.

Semantic Extraction with Use of Frames

This work describes an information extraction methodology which uses shallow parsing. We present detailed information on the extraction process, data structures used within that process as well as the evaluation of the described method. The extraction is fully automatic. Instead of machine learning it uses predefined frame templates and vocabulary stored within a domain ontology with elements related to frame templates. The architecture of the information extractor is modular and the main extraction module is capable of processing various languages when lexicalization for these languages is provided.

RuQAR: Reasoning Framework for OWL 2 RL Ontologies

This paper addresses the first release of the Rule-based Query Answering and Reasoning framework (RuQAR). The tool provides the ABox reasoning and query answering with OWL 2 RL ontologies executed by forward chaining rule reasoners. We describe current implementation and an experimental evaluation of RuQAR by performing reasoning on the number of benchmark ontologies. Additionally, we compare obtained results with inferences provided by HermiT and Pellet. The evaluation shows that we can perform the ABox reasoning with considerably better performance than DL-based reasoners.

Collaborative Filtering Based on Bi-Relational Data Representation

Abstract Widely-referenced approaches to collaborative filtering (CF) are based on the use of an input matrix that represents each user profile as a vector in a space of items and each item as a vector in a space of users. When the behavioral input data have the form of (userX, likes, itemY) and (userX, dislikes, itemY) triples one has to propose a representation of the user feedback data that is more suitable for the use of propositional data than the ordinary user-item ratings matrix. We propose to use an element-fact matrix, in which columns represent RDF-like behavioral data triples and rows represent users, items, and relations. By following such a triple-based approach to the bi-relational behavioral data representation we are able to improve the quality of collaborative filtering. One of the key findings of the research presented in this paper is that the proposed bi-relational behavioral data representation, while combined with reflective matrix processing, significantly outperforms state-of-the-art collaborative filtering methods based on the use of a ‘standard’ user-item matrix.

Perceptually transparent audio compression based on a variable bit rate AAC coder

The paper presents an implementation of a perceptually transparent variable bit rate (VBR) audio coder that complies with the MPEG-4 advanced audio coding (AAC) standard. The coder is based on the MPEG-4 reference software source code of CBR coder. With small modifications in the reference source code a coder guaranteeing perceptual transparency is achieved at compression ratios considerably better than that of the reference CBR coder. Comparison tests indicate that files generated by the perceptually transparent VBR coder applied to a standard set of test items are considerably smaller (in average 24%) than the smallest CBR files of the same quality.

Rule-Based Reasoning System for OWL 2 RL Ontologies

In this paper we present a method of transforming OWL 2 ontologies into a set of rules which can be used in a forward chaining rule engine. We use HermiT reasoner to perform the TBox reasoning and to produce classified form of an ontology. The ontology is automatically transformed into a set of Abstract Syntax of Rules and Facts. Then, it can be transformed into any forward chaining reasoning engine. We present an implementation of our method using two engines: Jess and Drools. We evaluate our approach by performing the ABox reasoning on the number of benchmark ontologies. Additionally, we compare obtained results with inferences provided by the HermiT reasoner. The evaluation shows that we can perform the ABox reasoning with considerably better performance than HermiT. We describe the details of our approach as well as future research and development.

Calculating Optimal Queries from the Query Relevance File

Query Expansion could bring very significant improvement of baseline results of Information Retrieval process. This has been known for many years, but very detailed results on annotated sets provide richer insight on the preferred added word space. In this work we introduce novel expanded term adequacy measures related to term frequency and inverse document frequency in relevant and non-relevant groups of documents. Term evaluation scores are derived using two term characteristics: inverse term representativeness and term usability. We generate the Optimal Queries based on the documents contents and the Qrels files of data used in the Text Retrieval Conference 2016 – Clinical Decision Support track (TREC-CDS 2016). The improvement can be up to a factor of 2 depending on the evaluation measure. Potentially, the method can be improved by increasing the learning set and applied to retrieval of documents in biomedical contests.

Malware Detection Using Black-Box Neural Method.

Because of the great loss and damage caused by malwares, malware detection has become a central issue of computer security. It has to be fast and very accurate. To develop suitable methods on needs very good quality benchmarks. One such benchmark is the Microsoft Kaggle malware challenge system run in 2015. Since then over 50 papers were published on this system. The best result were achieved with complex feature engineering. In this work we analyze the black-box neural method and what is novel analyze its results against the Microsoft Kaggle malware challenge benchmark. It is tempting to use convolution neural networks for malware analysis following the great success with analysis of images. Even the use of balanced classes and drop-out convergence does not beat XGBoost with feature engineering, although some room for improvement exists. The situation is similar to that for language analysis. The language is much more hierarchical than image, and apparently malware is too. The malware analysis still awaits optimal neural network architecture.