Loss Of Semantic Information Research Articles

Semantics-preserving optimisation of mapping multi-column key constraints for RDB to RDF transformation

The relational database (RDB) to resource description framework (RDF) transformation is a major semantic information extraction method because most web data are managed by RDBs. Existing automatic RDB-to-RDF transformation methods generate RDF data without losing the semantics of original relational data. However, two major problems have been observed during the mapping of multi-column key constraints: repetitive data generation and semantic information loss. In this article, we propose an improved RDB-to-RDF transformation method that ensures mapping without the aforementioned problems. Optimised rules are defined to generate an accurate semantic data structure for a multi-column key constraint and to reduce repetitive constraint data. Experimental results show that the proposed method achieves better accuracy in transforming multi-column key constraints and generates compact semantic results without repetitive data.

Md-Net: Multi-scale Dilated Convolution Network for CT Images Segmentation

Accurate CT image segmentation is of great importance to the clinical diagnosis. Due to the high similarity of gray values in CT image, the segmented areas are easily affected by their surroundings, which leads to the loss of semantic information. In this paper, we propose a multi-scale dilated convolution network (Md-Net) for CT image segmentation with superior segmentation performance compared with state-of-the-art methods. Specifically, our Md-Net utilizes the dilated convolutions with different sizes to form feature pyramids for extracting the semantic information. Moreover, we use a weighted Diceloss to accelerate the convergence in training process. Meanwhile, the bilinear interpolation and multiple convolutions are taken to reduce the computational cost. Experiment results show that our proposed Md-Net outperforms the representative medical image segmentation methods, including Unet, Unet++, MaskRcnn and CE-Net, in terms of sensitivity, accuracy and area under curve both on lung dataset and Bladder dataset.

Leveraging Smart Devices for Scene Text Preserved Image Stylization: A Deep Gaming Approach

With the rapid advancement of real-time rendering quality in digital art, the use of computer vision techniques has become popular in producing aesthetically pleasing stylized images of natural scenes. However, the processing time to produce a stylized image is high, especially in resource-constrained environments, such as smart devices. Most stylized imaging approaches are unable to preserve the fine details of the natural scene images, such as texts, symbols, and logos in the painted image, which leads to the loss of significant semantic information. In this article, we propose a fast image stylization framework (digital oil painting) using incremental histogramming that preserves the text content of natural scenes while efficiently painting it in resource-constrained environments. We design a stable multiplayer stochastic game with deep networks to classify regions into text or nontext using deep networks. We propose a multiscale fully convolutional character level text detector (xEASTLite) to detect the presence of a text character or part of a character with high accuracy. We have used different publicly available datasets in smart devices to illustrate the efficacy of the framework.

Deep Category-Level and Regularized Hashing With Global Semantic Similarity Learning.

The hashing technique has been extensively used in large-scale image retrieval applications due to its low storage and fast computing speed. Most existing deep hashing approaches cannot fully consider the global semantic similarity and category-level semantic information, which result in the insufficient utilization of the global semantic similarity for hash codes learning and the semantic information loss of hash codes. To tackle these issues, we propose a novel deep hashing approach with triplet labels, namely, deep category-level and regularized hashing (DCRH), to leverage the global semantic similarity of deep feature and category-level semantic information to enhance the semantic similarity of hash codes. There are four contributions in this article. First, we design a novel global semantic similarity constraint about the deep feature to make the anchor deep feature more similar to the positive deep feature than to the negative deep feature. Second, we leverage label information to enhance category-level semantics of hash codes for hash codes learning. Third, we develop a new triplet construction module to select good image triplets for effective hash functions learning. Finally, we propose a new triplet regularized loss (Reg-L) term, which can force binary-like codes to approximate binary codes and eventually minimize the information loss between binary-like codes and binary codes. Extensive experimental results in three image retrieval benchmark datasets show that the proposed DCRH approach achieves superior performance over other state-of-the-art hashing approaches.

Hierarchical Graph Transformer-Based Deep Learning Model for Large-Scale Multi-Label Text Classification

Traditional methods of multi-label text classification, particularly deep learning, have achieved remarkable results. However, most of these methods use word2vec technology to represent sequential text information, while ignoring the logic and internal hierarchy of the text itself. Although these approaches can learn the hypothetical hierarchy and logic of the text, it is unexplained. In addition, the traditional approach treats labels as independent individuals and ignores the relationships between them, which not only does not reflect reality but also causes significant loss of semantic information. In this paper, we propose a novel Hierarchical Graph Transformer based deep learning model for large-scale multi-label text classification. We first model the text into a graph structure that can embody the different semantics of the text and the connections between them. We then use a multi-layer transformer structure with a multi-head attention mechanism at the word, sentence, and graph levels to fully capture the features of the text and observe the importance of the separate parts. Finally, we use the hierarchical relationship of the labels to generate the representation of the labels, and design a weighted loss function based on the semantic distances of the labels. Extensive experiments conducted on three benchmark datasets demonstrated that the proposed model can realistically capture the hierarchy and logic of text and improve performance compared with the state-of-the-art methods.

Hierarchical Taxonomy-Aware and Attentional Graph Capsule RCNNs for Large-Scale Multi-Label Text Classification

CNNs, RNNs, GCNs, and CapsNets have shown significant insights in representation learning and are widely used in various text mining tasks such as large-scale multi-label text classification. Most existing deep models for multi-label text classification consider either the non-consecutive and long-distance semantics or the sequential semantics. However, how to coherently take them into account is still far from studied. In addition, most existing methods treat output labels as independent medoids, ignoring the hierarchical relationships among them, which leads to a substantial loss of useful semantic information. In this paper, we propose a novel hierarchical taxonomy-aware and attentional graph capsule recurrent CNNs framework for large-scale multi-label text classification. Specifically, we first propose to model each document as a word order preserved graph-of-words and normalize it as a corresponding word matrix representation preserving both the non-consecutive, long-distance and local sequential semantics. Then the word matrix is input to the proposed attentional graph capsule recurrent CNNs for effectively learning the semantic features. To leverage the hierarchical relations among the class labels, we propose a hierarchical taxonomy embedding method to learn their representations, and define a novel weighted margin loss by incorporating the label representation similarity. Extensive evaluations on three datasets show that our model significantly improves the performance of large-scale multi-label text classification by comparing with state-of-the-art approaches.

A text semantic topic discovery method based on the conditional co-occurrence degree

The topic discovery method, as an effective tool for semantic mining and a key means to extract new features from original text, plays an important role in the field of text mining and knowledge discovery. To solve the problems encountered in traditional topic models, such as the loss of semantic information and the ambiguity of topic concepts, as well as the crossover and coverage among topics, we propose a semantic topic discovery method based on the conditional co-occurrence degree (CCOD_STDM). First, every document is split into multiple subdocuments according to the semantic structure of the document and the independence decision rules. Second, combinatorial words with strong semantic relevance are extracted based on the conditional co-occurrence degree within the subdocuments. Based on these combinatorial words, new subdocuments are formed by feature expansion and content reconstruction. Third, “topic-word” distributions and “document-topic” distributions of new subdocuments are obtained by topic modeling with Gibbs sampling. Finally, “document-topic” distributions of the original documents are obtained by merging new subdocuments’ “document-topic” distributions with specific strategies. The numerical experiments are compared with six topic models and two evaluation methods on seven kinds of public corpora, and the experimental results verify the superiority of CCOD_STDM and its efficiency in topic discovery. More importantly, a case study illustrates that the combinatorial words can effectively avoid the polysemy problem and can facilitate the condensation and summary of topics.

Incorporating Topological Representation in 3D City Models

3D city models are being extensively used in applications such as evacuation scenarios and energy consumption estimation. The main standard for 3D city models is the CityGML data model which can be encoded through the CityJSON data format. CityGML and CityJSON use polygonal modelling in order to represent geometries. True topological data structures have proven to be more computationally efficient for geometric analysis compared to polygonal modelling. In a previous study, we have introduced a method to topologically reconstruct CityGML models while maintaining the semantic information of the dataset, based solely on the combinatorial map (C-Map) data structure. As a result of the limitations of C-Map’s semantic representation mechanism, the resulting datasets could suffer either from semantic information loss or the redundant repetition of them. In this article, we propose a solution for a more efficient representation of geometry, topology and semantics by incorporating the C-Map data structure into the CityGML data model and implementing a CityJSON extension to encode the C-Map data. In addition, we provide an algorithm for the topological reconstruction of CityJSON datasets to append them according to this extension. Finally, we apply our methodology to three open datasets in order to validate our approach when applied to real-world data. Our results show that the proposed CityJSON extension can represent all geometric information of a city model in a lossless way, providing additional topological information for the objects of the model.

A semantic-based inference control algorithm for OWL repository privacy protection

Web Ontology Language (OWL) designed to represent the class and relationships between classes is bringing a technological revolution in the Semantic Web. More and more knowledge is stored in the form of OWL. OWL repository refers to the systems that can store, query and infer OWL data. In OWL repository, information that users are forbidden to access (sensitive triples) can be inferred from information that the user have access to (non-sensitive triples) through inference rules, which brings about a security problem that some users’ sensitive information may be leaked. To protect sensitive triples, some researches have been done on the security problem, which mainly focus on removing some non-sensitive triples with low access frequency to achieve inference control. However, access frequency of a triple is unstable and it can not fully represent the importance of semantic information. In this paper, an inference control algorithm based on Semantic Value and access frequency is proposed. We introduced the concept of Degree of triples and Semantic Value, which can fully represent the importance of the semantic information represented by the triple. Moreover, we did experiments from the perspective of protecting sensitive triples of every user. Experimental results show that this algorithm can effectively prevent illegal inference and minimize the loss of semantic information.

GIS AND BIM INTEGRATION AT DATA LEVEL: A REVIEW

Abstract. City Geography Markup Language (CityGML) and Industry Foundation Class (IFC) are the two most popular data exchange format for the integration of Geographic Information System (GIS) and Building Information Modelling (BIM) respectively and has been identified by many researchers as an auspicious means of data interoperability between the two domains but with challenges on the compatibility between them. The main issue is the data loss in the process of information transformation. The success of integrating these two domains (GIS and BIM objects) is a great achievement toward solving problems in Architecture, Engineering and Construction (AEC), Facility Management (FM), Disaster Management (DM) sectors. Nevertheless, as we all know GIS and BIM are different fields used by different professionals using different software packages, used for different purposes, it is definitely face with many challenges including data interoperability, mismatch and loss of semantic information are bound to occur during the process of integration. In order to comprehend the two domains and their data models of CityGML and IFC. This paper review existing models on GIS and BIM developed by different researchers, the complementarity and compatibility of GIS and BIM on the previous integration techniques were also reviewed and finally, the paper review the integration of GIS and BIM at the data level aimed at solving different problems surrounding it by considering the transformation of coordinates at geometric level from CityGML to IFC, in order to achieve flow of information between GIS and BIM.

Computing semantic similarity based on novel models of semantic representation using Wikipedia

Computing Semantic Similarity (SS) between concepts is one of the most critical issues in many domains such as Natural Language Processing and Artificial Intelligence. Over the years, several SS measurement methods have been proposed by exploiting different knowledge resources. Wikipedia provides a large domain-independent encyclopedic repository and a semantic network for computing SS between concepts. Traditional feature-based measures rely on linear combinations of different properties with two main limitations, the insufficient information and the loss of semantic information. In this paper, we propose several hybrid SS measurement approaches by using the Information Content (IC) and features of concepts, which avoid the limitations introduced above. Considering integrating discrete properties into one component, we present two models of semantic representation, called CORM and CARM. Then, we compute SS based on these models and take the IC of categories as a supplement of SS measurement. The evaluation, based on several widely used benchmarks and a benchmark developed by ourselves, sustains the intuitions with respect to human judgments. In summary, our approaches are more efficient in determining SS between concepts and have a better human correlation than previous methods such as Word2Vec and NASARI.

Retrograde Amnesia for Episodic and Semantic Memories in Amnestic Mild Cognitive Impairment.

Retrograde amnesia (RA), which includes loss of memory for past personal events (autobiographical RA) and for acquired knowledge (semantic RA), has been largely documented in patients with amnestic mild cognitive impairment (aMCI). However, previous studies have produced controversial results particularly concerning the temporal extent of memory impairment. Here we investigated whether, with the onset of hippocampal pathology, age of memory acquisition and retrieval frequency play different roles in modulating the progressive loss of semantic and episodic contents of retrograde memory respectively. For this purpose, aMCI patients and healthy controls were tested for the ability to recall semantic and autobiographical information related to famous public events as a function of both age of acquisition and retrieval frequency. In aMCI patients, we found that the impairment in recollecting past personal incidents was modulated by the combined action of memory age and retrieval frequency, because older and more frequently retrieved episodes are less susceptible to loss than more recent and less frequently retrieved ones. On the other side, we found that the loss of semantic information depended only on memory age, because the remoteness of the trace allows for better preservation of the memory. Our results provide evidence that the loss of the two components of retrograde memory is regulated by different mechanisms. This supports the view that diverse neural mechanisms are involved in episodic and semantic memory trace storage and retrieval, as postulated by the Multiple Trace Theory.

Semantic-based topic detection using Markov decision processes

In the field of text mining, topic modeling and detection are fundamental problems in public opinion monitoring, information retrieval, social media analysis, and other activities. Document clustering has been used for topic detection at the document level. Probabilistic topic models treat topics as a distribution over the term space, but this approach overlooks the semantic information hidden in the topic. Thus, representing topics without loss of semantic information as well as detecting the optimal topic is a challenging task. In this study, we built topics using a network called a topic graph, where the topics were represented as concept nodes and their semantic relationships using WordNet. Next, we extracted each topic from the topic graph to obtain a corpus by community discovery. In order to find the optimal topic to describe the related corpus, we defined a topic pruning process, which was used for topic detection. We then performed topic pruning using Markov decision processes, which transformed topic detection into a dynamic programming problem. Experimental results produced using a newsgroup corpus and a science literature corpus showed that our method obtained almost the same precision and recall as baseline models such as latent Dirichlet allocation and KeyGraph. In addition, our method performed better than the probabilistic topic model in terms of its explanatory power and the runtime was lower compared with all three baseline methods, while it can also be optimized to adapt the corpus better by using topic pruning.

A Cluster based Approach with N-grams at Word Level for Document Classification

A breakneck progress of computers and web makes it easier to collect and store large amount of information in the form of text; e.g., reviews, forum postings, blogs, web pages, news articles, email messages. In text mining, growing size of text datasets and high dimensionality associated with natural language is great challenge which makes it difficult to classify documents in various categories and sub-categories. This paper focuses on cluster based document classification technique so that data inside each cluster shares some common trait. The common approach for document clustering problem is bag of words model (BOW), where words are considered as features. But some semantic information is always lost as only words are considered. Thus we aim at using vector-space model based on N-grams at word level which helps to reduce the loss of semantic information. The problem of high dimensionality is solved with feature selection technique by applying threshold on feature values of vector space model. The vector space is mapped into a modified one with latent semantic analysis (LSA). Clustering of documents is done using k-means algorithm. Experiments are performed on Stack Exchange data set of some categories. R is used as text mining tool for implementation purpose. Experiment results show that tri-grams give better clustering results than words and bi-grams.

Categorising texts more accurately with field association terms

Popular text classification algorithms such as Naive Bayes, kNN, Centroid-based classifiers and support vector machines SVM are based on supervised machine learning. They normally use classical text representation technique consisting of a 'bag of words' as features. This representation leads to the inclusion of unimportant features, and the loss of important semantic relationships and inflection information, resulting in accuracy reduction. To address this problem, we propose a new text classification methodology based on field association terms - a set of terms that identify specific document fields. The methodology is compared against Naive Bayes, kNN, Centroid-based classifier and SVM on a close dataset of 3180 documents from Wikipedia dumps and open dataset of 9449 documents from Reuters RCV1 Corpus, 20-Newsgroup and 4-Universities datasets. The new method outperformed the other algorithms with a precision of 97% as compared with Centroid-based 85%, Naive Bayes 78%, kNN 48% and SVM 42%.

Web Service Clustering using a Hybrid Term-Similarity Measure with Ontology Learning

Clustering Web services into functionally similar clusters is a very efficient approach to service discovery. A principal issue for clustering is computing the semantic similarity between services. Current approaches use similarity-distance measurement methods such as keyword, information-retrieval or ontology based methods. These approaches have problems that include discovering semantic characteristics, loss of semantic information and a shortage of high-quality ontologies. In this paper, the authors present a method that first adopts ontology learning to generate ontologies via the hidden semantic patterns existing within complex terms. If calculating similarity using the generated ontology fails, it then applies an information-retrieval-based method. Another important issue is identifying the most suitable cluster representative. This paper proposes an approach to identifying the cluster center by combining service similarity with term frequency–inverse document frequency values of service names. Experimental results show that our term-similarity approach outperforms comparable existing approaches. They also demonstrate the positive effects of our cluster-center identification approach.

Formula omitted]-Dependency for privacy-preserving XML data publishing

An ever increasing amount of medical data such as electronic health records, is being collected, stored, shared and managed in large online health information systems and electronic medical record systems (EMR) (Williams et al., 2001; Virtanen, 2009; Huang and Liou, 2007) [1–3]. From such rich collections, data is often published in the form of census and statistical data sets for the purpose of knowledge sharing and enabling medical research. This brings with it an increasing need for protecting individual people privacy, and it becomes an issue of great importance especially when information about patients is exposed to the public.While the concept of data privacy has been comprehensively studied for relational data, models and algorithms addressing the distinct differences and complex structure of XML data are yet to be explored. Currently, the common compromise method is to convert private XML data into relational data for publication. This ad hoc approach results in significant loss of useful semantic information previously carried in the private XML data. Health data often has very complex structure, which is best expressed in XML. In fact, XML is the standard format for exchanging (e.g. HL7 version 31http://www.hl7.org.1) and publishing health information. Lack of means to deal directly with data in XML format is inevitably a serious drawback.In this paper we propose a novel privacy protection model for XML, and an algorithm for implementing this model. We provide general rules, both for transforming a private XML schema into a published XML schema, and for mapping private XML data to the new privacy-protected published XML data. In addition, we propose a new privacy property, δ-dependency, which can be applied to both relational and XML data, and that takes into consideration the hierarchical nature of sensitive data (as opposed to “quasi-identifiers”). Lastly, we provide an implementation of our model, algorithm and privacy property, and perform an experimental analysis, to demonstrate the proposed privacy scheme in practical application.

Improving Efficiency of Relational Classification Technique Based on Relational Database Using Contribution of Tables

Classification is one of the most popular data mining tasks with a wide range of applications. Many algorithms have been proposed to build an accurate classifier. These algorithms work for a single Table as an input but in real-world applications most of the data relies on multiple Tables. As converting data from multiple relations into single flat relation usually causes many problems, development of classification across multiple database relations becomes important. There have been many approaches for classification, such as neural networks and support vector machines. However, they can only be applied to data in single flat relations. It is counterproductive to convert multi-relational data into single flat Tables because such conversion may lead to the generation of huge relation and loss of essential semantic information. In this work, we propose algorithm for multi-relational classification (MRC) which uses weighted voting technique to combine classifiers to get class label based on the contribution of Tables. This will classify the instance accurately and efficiently.

Near-lossless semantic video summarization and its applications to video analysis

The ever increasing volume of video content on the Web has created profound challenges for developing efficient indexing and search techniques to manage video data. Conventional techniques such as video compression and summarization strive for the two commonly conflicting goals of low storage and high visual and semantic fidelity. With the goal of balancing both video compression and summarization, this article presents a novel approach, called Near-Lossless Semantic Summarization (NLSS), to summarize a video stream with the least high-level semantic information loss by using an extremely small piece of metadata. The summary consists of compressed image and audio streams, as well as the metadata for temporal structure and motion information. Although at a very low compression rate (around 1/40 of H.264 baseline, where traditional compression techniques can hardly preserve an acceptable visual fidelity), the proposed NLSS still can be applied to many video-oriented tasks, such as visualization, indexing and browsing, duplicate detection, concept detection, and so on. We evaluate the NLSS on TRECVID and other video collections, and demonstrate that it is a powerful tool for significantly reducing storage consumption, while keeping high-level semantic fidelity.

Emotional response and visual attention to non-photorealistic images

Non-photorealistic rendering (NPR) algorithms are used to produce stylized images, and have generally been evaluated on the aesthetic qualities of the resulting images. NPR-produced images have been used for aesthetic and practical reasons in media intended to produce an emotional reaction in a consumer (e.g., computer games, films, advertisements, and web sites); however, it is not understood how non-photorealistic rendering affects the emotion portrayed in an image. We conducted a study of subjective emotional response and visual attention to five common NPR approaches, two blurring techniques, and the original image with 42 participants, and found that the NPR algorithms dampened participants' emotional responses in terms of arousal (activation) and valence (pleasure). Gaze data revealed that NPR rendering of images might reduce emotional response to an image by producing confusion, creating distracting visual artifacts, causing the loss of meaningful semantic information, or causing users to lose interest in the resulting image.