Narrow Results

Complex network studies, as an interdisciplinary framework, span a large variety of subjects including social media. In social networks, several mechanisms generate miscellaneous structures like friendship networks, mention networks, tag networks, etc. Focusing on tag networks (namely, hashtags in twitter), we made a two-layer analysis of tag networks from a massive dataset of Twitter entries. The first layer is constructed by converting the co-occurrences of these tags in a single entry (tweet) into links, while the second layer is constructed converting the semantic relations of the tags into links. We observed that the universal properties of the real networks like small-world property, clustering and power-law distributions in various network parameters are also evident in the multilayer network of hashtags. Moreover, we outlined that co-occurrences of hashtags in tweets are mostly coupled with semantic relations, whereas a small number of semantically unrelated, therefore random links reduce node separation and network diameter in the co-occurrence network layer. Together with the degree distributions, the power-law consistencies of degree difference, edge weight and cosine similarity distributions in both layers are also appealing forms of Zipf’s law evident in nature.

Defect detection is a crucial technology that is extensively employed in the manufacturing industry to monitor and ensure the quality of output. Deep learning models have shown remarkable potential for defect detection. However, the success of these models heavily relies on voluminous training data. Collecting substantial amounts of defect data is challenging in practical settings, and the tedious process of pixel-level defect annotation further complicates the task. Among the common defects encountered in manufacturing, scratches are particularly significant. To address these challenges, this study proposes a two-phase generative adversarial network (GAN) approach for synthesizing defect images and generating semi-automatic pixel-wise labels for anomaly detection. The first phase primarily focuses on synthesizing images, while the second phase involves the pixel-wise labeling of the images. The synthesized paired images generated by the GANs serve as input to the semantic network. Notably, the proposed methodology requires only a few real defect samples for training and a small amount of annotated data, making it practical and computationally efficient for implementation in the manufacturing industry. Experimental results indicate the effectiveness of the proposed deep-learning solution in defect detection, specifically in identifying scratches on various textured and patterned surfaces. A notably high detection accuracy is achieved, validating the potential of the approach in real-world manufacturing scenarios.

Any system for natural language processing must be based on a lexicon. Once a model has been defined, there is the problem of acquiring and inserting words. This task is tedious for a human operator; on the one hand he must not forget any of the words, and on the other the acquisition of a new concept requires the input of a number of parameters.

In view of these difficulties, research work has been undertaken in order to integrate pre-existing “paper” dictionaries. Nevertheless, these are not faultless, and are often incomplete when processing a very specialized technical field. We have therefore searched to mitigate these problems by automating the enrichment of an already partially integrated lexicon.

We work in a technical field on which we have gathered different sorts of texts: written texts, specialist interviews, technical reports, etc. These documents are stored in an object oriented database, and form part of a wider project, called REX (“Retour d’EXpérience” in French, or “Feedback of Experience” in English).

Our system, called ANA, reads the documents, analyses them, and deduces new knowledge, allowing the enrichment of the lexicon. The group of words already integrated into the lexicon form the “Bootstrap” of the discovery process of new words: it collects the instances of the different concepts thought to be interesting, in order to gather the semantic information. A special module makes it possible to avoid an explosion of the size of the database. It is responsible for forgetting certain instances and maintaining the database in such a way that the order in which the texts are introduced bears no influence.

Evolutionary computation has been successfully applied in a variety of problem domains and applications. In this paper we discuss the use of a specific form of evolutionary computation known as Cultural Algorithms to improve the efficiency of the subsumption algorithm in semantic networks. We identify two complementary methods of using Cultural Algorithms to solve the problem of re-engineering large-scale dynamic semantic networks in order to optimize the efficiency of subsumption: top-down and bottom-up.

The top-down re-engineering approach improves subsumption efficiency by reducing the number of attributes that need to be compared for every node without impacting the results. We demonstrate that a Cultural Algorithm approach can be used to identify these defining attributes that are most significant for node retrieval. These results are then utilized within an existing vehicle assembly process planning application that utilizes a semantic network based knowledge base to improve the performance and reduce complexity of the network. It is shown that the results obtained by Cultural Algorithms are at least as good, and in most cases better, than those obtained by the human developers. The advantage of Cultural Algorithms is especially pronounced for those classes in the network that are more complex.

The goal of bottom-up approach is to classify the input concepts into new clusters that are most efficient for subsumption and classification. While the resultant subsumption efficiency for the bottom-up approach exceeds that for the top-down approach, it does so by removing structural relationships that made the network understandable to human observers. Like a Rete network in expert systems, it is a compilation of only those relationships that impact subsumption. A direct comparison of the two approaches shows that bottom-up semantic network re-engineering creates a semantic network that is approximately 5 times more efficient than the top-down approach in terms of the cost of subsumption. In conclusion, we will discuss these results and show that some knowledge that is useful to the system users is lost during the bottom-up re-engineering process and that the best approach for re-engineering a semantic network requires a combination of both of these approaches.

Semantic networks as a means for knowledge representation and manipulation are used in many artificial intelligence applications. A number of computer architectures, that have been reported for semantic network processing, are presented in this paper. A novel set of evaluation criteria for such semantic network architectures has been developed. Semantic network processing as well as architectural issues are considered in such evaluation criteria. A study of how the reported architectures meet the requirements of each criterion is presented. This set of evaluation criteria is useful for future designs of machines for semantic networks because of its comprehensive range of issues on semantic networks and architectures.

Concepts in a certain domain of science are linked via intrinsic connections reflecting the structure of knowledge. To get a qualitative insight and a quantitative description of this structure, we perform empirical analysis and modeling of the network of scientific concepts in the domain of physics. To this end, we use a collection of manuscripts submitted to the e-print repository arXiv and the vocabulary of scientific concepts collected via the ScienceWISE.info platform and construct a network of scientific concepts based on their co-occurrences in publications. The resulting complex network possesses a number of specific features (high node density, dissortativity, structural correlations, skewed node degree distribution) that cannot be understood as a result of simple growth by several commonly used network models. We show that the model based on a simultaneous account of two factors, growth by blocks and preferential selection, gives an explanation of empirically observed properties of the concepts network.

Extracting semantic information from multiple natural language sources and combining that information into a single unified resource is an important and fundamental goal for natural language processing. Large scale resources of this kind can be useful for a wide variety of tasks including question answering, word sense disambiguation and knowledge discovery. A single resource representing the information in multiple documents can provide significantly more semantic information than is available from the documents considered independently.

The ASKNet system utilises existing NLP tools and resources, together with spreading activation based techniques, to automatically extract semantic information from a large number of English texts, and combines that information into a large scale semantic network. The initial emphasis of the ASKNet system is on wide-coverage, robustness and speed of construction. In this paper we show how a network consisting of over 1.5 million nodes and 3.5 million edges, more than twice as large as any network currently available, can be created in less than 3 days. Evaluation of large-scale semantic networks is a difficult problem. In order to evaluate ASKNet we have developed a novel evaluation metric based on the notion of a network "core" and employed human evaluators to determine the precision of various components of that core. We have applied this evaluation to networks created from randomly chosen articles used by DUC (Document Understanding Conference). The results are highly promising: almost 80% precision in the semantic core of the networks.

The construction of suitable and scalable representations of semantic knowledge is a core challenge in Semantic Computing. Manually created resources such as WordNet have been shown to be useful for many AI and NLP tasks, but they are inherently restricted in their coverage and scalability. In addition, they have been challenged by simple distributional models on very large corpora, questioning the advantage of structured knowledge representations.

We present a framework for building large-scale semantic networks automatically from plain text and Wikipedia articles using only linguistic analysis tools. Our constructed resources cover up to 2 million concepts and were built in less than 6 days. Using the task of measuring semantic relatedness, we show that we achieve results comparable to the best WordNet based methods as well as the best distributional methods while using a corpus of a size several magnitudes smaller. In addition, we show that we can outperform both types of methods by combining the results of our two network variants. Initial experiments on noun compound paraphrasing show similar results, underlining the quality as well as the flexibility of our constructed resources.

In this paper we describe an approach to manage multimedia data in the framework of an advanced information and knowledge management system based on the terminological representation model Back. The guide-lines of the approach are based on using Back to provide conceptual descriptions of multimedia data, and on providing a multimedia object multimedia supporting all functionalities related to the physical organization. The usage of Back as conceptual tool for multimedia data has several advantages, such as adequate expressive power and reasoning capabilities. The multimedia object manager has been designed following an object-oriented approach. The main advantage is to provide extensibility, thus allowing new types of media to dynamically added to the system.