Search

Zusammenfassung Unternehmensbewertungen in der Biotech-Branche, Pharmazie und Medizintechnik stellen eine anspruchsvolle Aufgabe dar, insbesondere bei Berücksichtigung der einzigartigen Risiken, denen Biotech-Startups beim Eintritt in neue Märkte ausgesetzt sind. Unternehmen, die auf globale Bewertungsdienstleistungen spezialisiert sind, kombinieren daher Bewertungsmodelle und Erfahrungen aus der Vergangenheit mit heterogenen Metriken und Indikatoren, die Einblicke in die Leistung eines Unternehmens geben. Dieser Beitrag veranschaulicht, wie automatisierte Wissensidentifikation, -extraktion und -integration genutzt werden können, um (i) zusätzliche Indikatoren zu ermitteln, die Einblicke in den Erfolg eines Unternehmens in der Produktentwicklung geben und um (ii) arbeitsintensive Datensammelprozesse zur Unternehmensbewertung zu unterstützen.

Automatic extraction of forum posts and metadata is a crucial but challenging task since forums do not expose their content in a standardized structure. Content extraction methods, therefore, often need customizations such as adaptations to page templates and improvements of their extraction code before they can be deployed to new forums. Most of the current solutions are also built for the more general case of content extraction from web pages and lack key features important for understanding forum content such as the identification of author metadata and information on the thread structure. This paper, therefore, presents a method that determines the XPath of forum posts, eliminating incorrect mergers and splits of the extracted posts that were common in systems from the previous generation. Based on the individual posts further metadata such as authors, forum URL and structure are extracted. We also introduce Harvest, a new open source toolkit that implements the presented methods and create a gold standard extracted from 52 different Web forums for evaluating our approach. A comprehensive evaluation reveals that Harvest clearly outperforms competing systems., IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT 2020), Accepted 27 October 2020

Inscriptis provides a library, command line client and Web service for converting HTML to plain text. Its development has been triggered by the need to obtain accurate text representations for knowledge extraction tasks that preserve the spatial alignment of text without drawing upon heavyweight, browser-based solutions such as Selenium. In contrast to related software packages, Inscriptis (i) provides a layout-aware conversion of HTML that more closely resembles the rendering obtained from standard Web browsers; and (ii) supports annotation rules, i.e., user-provided mappings that allow for annotating the extracted text based on structural and semantic information encoded in HTML tags and attributes. These unique features ensure that downstream knowledge extraction components can operate on accurate text representations, and may even use information on the semantics and structure of the original HTML document., Comment: Preprint of the published version, which includes all improvements made during the review process

A substantial number of international corporations have been affected by corruption. The research presented in this paper introduces the Integrity Risks Monitor, an analytics dashboard that applies Web Intelligence and Deep Learning to english and german-speaking documents for the task of (i) tracking and visualizing past corruption management gaps and their respective impacts, (ii) understanding present and past integrity issues, (iii) supporting companies in analyzing news media for identifying and mitigating integrity risks.Afterwards, we discuss the design, implementation, training and evaluation of classification components capable of identifying English documents covering the integrity topic of corruption. Domain experts created a gold standard dataset compiled from Anglo-American media coverage on corruption cases that has been used for training and evaluating the classifier. The experiments performed to evaluate the classifiers draw upon popular algorithms used for text classification such as Naive Bayes, Support Vector Machines (SVM) and Deep Learning architectures (LSTM, BiLSTM, CNN) that draw upon different word embeddings and document representations. They also demonstrate that although classical machine learning approaches such as Naive Bayes struggle with the diversity of the media coverage on corruption, state-of-the art Deep Learning models perform sufficiently well in the project's context.

Gold standard corpora and competitive evaluations play a key role in benchmarking named entity linking (NEL) performance and driving the development of more sophisticated NEL systems. The quality of the used corpora and the used evaluation metrics are crucial in this process. We, therefore, assess the quality of three popular evaluation corpora, identifying four major issues which affect these gold standards: (i) the use of different annotation styles, (ii) incorrect and missing annotations, (iii) Knowledge Base evolution, (iv) and differences in annotating co-occurrences. This paper addresses these issues by formalizing NEL annotations and corpus versioning which allows standardizing corpus creation, supports corpus evolution, and paves the way for the use of lenses to automatically transform between different corpus configurations. In addition, the use of clearly defined scoring rules and evaluation metrics ensures a better comparability of evaluation results.

Extracting and analyzing affective knowledge from social media in a structured manner is a challenging task. Decision makers require insights into the public perception of a company's products and services, as a strategic feedback channel to guide communication campaigns, and as an early warning system to quickly react in the case of unforeseen events. The approach presented in this article goes beyond bipolar metrics of sentiment. It combines factual and affective knowledge extracted from rich public knowledge bases to analyze emotions expressed toward specific entities (targets) in social media. The authors obtain common and common-sense domain knowledge from DBpedia and ConceptNet to identify potential sentiment targets. They employ affective knowledge about emotional categories available from SenticNet to assess how those targets and their aspects (such as specific product features) are perceived in social media. An evaluation shows the usefulness and correctness of the extracted domain knowledge, which is used in a proof-of-concept data analytics application to investigate the perception of car brands on social media in the period between September and November 2015.

Knowledge-rich Information Extraction (IE) methods aspire towards combining classical IE with background knowledge obtained from third-party resources. Linked Open Data repositories that encode billions of machine readable facts from sources such as Wikipedia play a pivotal role in this development. The recent growth of Linked Data adoption for Information Extraction tasks has shed light on many data quality issues in these data sources that seriously challenge their usefulness such as completeness, timeliness and semantic correctness. Information Extraction methods are, therefore, faced with problems such as name variance and type confusability. If multiple linked data sources are used in parallel, additional concerns regarding link stability and entity mappings emerge. This paper develops methods for integrating Linked Data into Named Entity Linking methods and addresses challenges in regard to mining knowledge from Linked Data, mitigating data quality issues, and adapting algorithms to leverage this knowledge. Finally, we apply these methods to Recognyze, a graph-based Named Entity Linking (NEL) system, and provide a comprehensive evaluation which compares its performance to other well-known NEL systems, demonstrating the impact of the suggested methods on its own entity linking performance.

The news media landscape tends to focus on long-running narratives. Correctly processing new information, therefore, requires considering multiple lenses when analyzing media content. Traditionally it would have been considered sufficient to extract the topicsor entities contained in a text in order to classify it, but today it is important to also look at more sophisticated annotations related to fine grained geolocation, events, stories and the relations between them. In order to leverage such lenses we propose a new corpus that offers a diverse set of annotations over texts collected from multiple media sources. We also showcase the framework used for creating the corpus, as well as how the information from the various lenses can be used in order to support different use cases in the EU project InVID for verifying the veracity of online video.

Given the intense attention that environmental topics such as climate change attract in news and social media coverage, scientists and communication professionals want to know how different stakeholders perceive observable threats and policy options, how specific media channels react to new insights, and how journalists present scientific knowledge to the public. This paper investigates the potential of semantic technologies to address these questions. After summarizing methods to extract and disambiguate context information, we present visualization techniques to explore the lexical, geospatial, and relational context of topics and entities referenced in these repositories. The examples stem from the Media Watch on Climate Change , the Climate Resilience Toolkit and the NOAA Media Watch —three applications that aggregate environmental resources from a wide range of online sources. These systems not only show the value of providing comprehensive information to the public, but also have helped to develop a novel communication success metric that goes beyond bipolar assessments of sentiment.

Advances in research areas such as named entity linking and sentiment analysis have triggered the emergence of knowledge-intensive information extraction methods that combine classical information extraction with background knowledge from the Web. Despite data quality concerns, linked data sources such as DBpedia, GeoNames and Wikidata which encode facts in a standardized structured format are particularly attractive for such applications.This paper addresses the problem of data quality by introducing a framework that elaborates on linked data quality issues relevant to different stages of the background knowledge acquisition process, their impact on information extraction performance and applicable mitigation strategies. Applying this framework to named entity linking and data enrichment demonstrates the potential of the introduced mitigation strategies to lessen the impact of different kinds of data quality problems. An industrial use case that aims at the automatic generation of image metadata from image descriptions illustrates the successful deployment of knowledge-intensive information extraction in real-world applications and constraints introduced by data quality concerns.

Over the last years, the amount of data published as Linked Data on the Web has grown enormously. In spite of the high availability of Linked Data, organizations still encounter an accessibility challenge while consuming it. This is mostly due to the large size of some of the datasets published as Linked Data. The core observation behind this work is that a subset of these datasets suffices to address the needs of most organizations. In this paper, we introduce Torpedo, an approach for efficiently selecting and extracting relevant subsets from RDF datasets. In particular, Torpedo adds optimization techniques to reduce seek operations costs as well as the support of multi-join graph patterns and SPARQL FILTERs that enable to perform a more granular data selection. We compare the performance of our approach with existing solutions on nine different queries against four datasets. Our results show that our approach is highly scalable and is up to 26% faster than the current state-of-the-art RDF dataset slicing approach.

This paper investigates the linking of sentiments to their respective targets, a sub-task of fine-grained sentiment analysis. Many different features have been proposed for this task, but often without a formal evaluation. We employ a recursive feature elimination approach to identify features that optimize predictive performance. Our experimental evaluation draws upon two corpora of product reviews and news articles annotated with sentiments and their targets. We introduce competitive baselines, outline the performance of the proposed approach, and report the most useful features for sentiment target linking. The results help to better understand how sentiment-target relations are expressed in the syntactic structure of natural language, and how this information can be used to build systems for fine-grained sentiment analysis.

A context-aware approach based on machine learning and lexical analysis identifies ambiguous terms and stores them in contextualized sentiment lexicons, which ground the terms to concepts corresponding to their polarity.

Organizations require tools that can assess their online reputations as well as the impact of their marketing and public outreach activities. The Media Watch on Climate Change is a Web intelligence and online collaboration platform that addresses this requirement. It aggregates large archives of digital content from multiple stakeholder groups and enables the co-creation and visualization of evolving knowledge archives. Here, the authors introduce the base platform and a context-aware document editor as an add-on that supports concurrent authoring by multiple users. While documents are being edited, semantic methods analyze them on the fly to recommend related content. The system computes positive or negative sentiment automatically to provide a better understanding of third-party perceptions. The editor is part of an interactive dashboard that uses trend charts and map projections to show how often and where relevant information is published, and to provide a real-time account of concepts that stakeholders associate with a topic.

Understanding stakeholder perceptions and assessing the impact of campaigns are key questions of communication experts. Web intelligence platforms help to answer such questions, provided that they are scalable enough to analyze and visualize information flows from volatile online sources in real time. This paper presents a distributed architecture for aggregating Web content repositories from Web sites and social media streams, memory-efficient methods to extract factual and affective knowledge, and interactive visualization techniques to explore the extracted knowledge. The presented examples stem from the Media Watch on Climate Change, a public Web portal that aggregates environmental content from a range of online sources.

Policy makers and environmental organizations have a keen interest in awareness building and the evolution of stakeholder opinions on environmental issues. Mere polarity detection, as provided by many existing methods, does not suffice to understand the emergence of collective awareness. Methods for extracting affective knowledge should be able to pinpoint opinion targets within a thread. Opinion target extraction provides a more accurate and fine-grained identification of opinions expressed in online media. This paper compares two different approaches for identifying potential opinion targets and applies them to comments from the YouTube video sharing platform. The first approach is based on statistical keyword analysis in conjunction with sentiment classification on the sentence level. The second approach uses dependency parsing to pinpoint the target of an opinionated term. A case study based on YouTube postings applies the developed methods and measures their ability to handle noisy input data from social media streams.

One key property of the Semantic Web is its support for interoperability. Recent research in this area focuses on the integration of multiple data sources to facilitate tasks such as ontology learning, user query expansion and context recognition. The growing popularity of such machups and the rising number of Web APIs supporting links between heterogeneous data providers asks for intelligent methods to spare remote resources and minimize delays imposed by queries to external data sources. This paper suggests a cost and utility model for optimizing such queries by leveraging optimal stopping theory from business economics: applications are modeled as decision makers that look for optimal answer sets. Queries to remote resources cause additional cost but retrieve valuable information which improves the estimation of the answer set's utility. Optimal stopping optimizes the trade-off between query cost and answer utility yielding optimal query strategies for remote resources. These strategies are compared to conventional approaches in an extensive evaluation based on real world response times taken from seven popular Web services.

This paper presents a method to integrate external knowledge sources such as DBpedia and OpenCyc into an ontology learning system that automatically suggests labels for unknown relations in domain ontologies based on large corpora of unstructured text. The method extracts and aggregates verb vectors from semantic relations identified in the corpus. It composes a knowledge base which consists of (i) verb centroids for known relations between domain concepts, (ii) mappings between concept pairs and the types of known relations, and (iii) ontological knowledge retrieved from external sources. Applying semantic inference and validation to this knowledge base improves the quality of suggested relation labels. A formal evaluation compares the accuracy and average ranking precision of this hybrid method with the performance of methods that solely rely on corpus data and those that are only based on reasoning and external data sources.

The advantages and positive effects of multiple coordinated views on search performance have been documented in several studies. This paper describes the implementation of multiple coordinated views within the Media Watch on Climate Change, a domain-specific news aggregation portal available at www.ecoresearch.net/climate that combines a portfolio of semantic services with a visual information exploration and retrieval interface. The system builds contextualized information spaces by enriching the content repository with geospatial, semantic and temporal annotations, and by applying semi-automated ontology learning to create a controlled vocabulary for structuring the stored information. Portlets visualize the different dimensions of the contextualized information spaces, providing the user with multiple views on the latest news media coverage. Context information facilitates access to complex datasets and helps users navigate large repositories of Web documents. Currently, the system synchronizes information landscapes, domain ontologies, geographic maps, tag clouds and just-in-time information retrieval agents that suggest similar topics and nearby locations.

This paper presents the U.S. Election 2004 Web Monitor, a public Web portal that captured trends in political media coverage before and after the 2004 U.S. presidential election. Developed by the authors of this article, the webLyzard suite of Web mining tools provided the required functionality to aggregate and analyze about a half-million documents in weekly intervals. The study paid particular attention to the editorial slant, which is defined as the quantity and tone of a Web site's coverage as influenced by its editorial position. The observable attention and attitude toward the candidates served as proxies of editorial slant. The system identified attention by determining the frequency of candidate references and measured attitude towards the candidate by looking for positive and negative expressions that co-occur with these references. Keywords and perceptual maps summarized the most important topics associated with the candidates, placing special emphasis on environmental issues.

To identify trends and assign metadata elements such as location and sentiment to the correct entities, Web intelligence applications require methods for linking named entities and revealing relations between organizations, persons and products. For this purpose we introduce Recognyze, a named entity linking component that uses background knowledge obtained from linked data repositories. This paper outlines the underlying methods, provides insights into the migration of proprietary knowledge sources to linked enterprise data, and discusses the lessons learned from adapting linked data for named entity linking. A large dataset obtained from Orell Fussli, the largest Swiss business information provider, serves as the main showcase. This dataset includes more than nine million triples on companies, their contact information, management, products and brands. We identify major challenges towards applying this data for named entity linking and conduct a comprehensive evaluation based on several news corpora to illustrate how Recognyze helps address them, and how it improves the performance of named entity linking components drawing upon linked data rather than machine learning techniques.

Knowledge capture approaches in the age of massive Web data require robust and scalable mechanisms to acquire, consolidate and pre-process large amounts of heterogeneous data, both unstructured and structured. This paper addresses this requirement by introducing the Extensible Web Retrieval Toolkit (eWRT), a modular Python API for retrieving social data from Web sources such as Delicious, Flickr, Yahoo! and Wikipedia. eWRT has been released as an open source library under GNU GPLv3. It includes classes for caching and data management, and provides low-level text processing capabilities including language detection, phonetic string similarity measures, and string normalization.

The webLyzard media monitoring and Web intelligence platform (www.webLyzard.com) presented in this paper is a generic tool for assessing the strategic positioning of an organization and the effectiveness of its communication strategies. The platform captures and aggregates large archives of digital content from multiple stakeholder groups. Each week it processes millions of documents and user comments from news media, blogs, Web 2.0 platforms such as Facebook, Twitter and YouTube, the Web sites of companies and NGOs, and other sources. An interactive dashboard with trend charts and complex map projections shows how often and where information is published. It also provides a real-time account of topics that stakeholders associate with an organization. Positive or negative sentiment is computed automatically, which reflects the impact of public relations and marketing campaigns.

Incrementally computed information landscapes are an effective means to visualize longitudinal changes in large document repositories. Resembling tectonic processes in the natural world, dynamic rendering reflects both long-term trends and short-term fluctuations in such repositories. To visualize the rise and decay of topics, the mapping algorithm elevates and lowers related sets of concentric contour lines. Addressing the growing number of documents to be processed by state-of-the-art knowledge discovery applications, we introduce an incremental, scalable approach for generating such landscapes. The processing pipeline includes a number of sequential tasks, from crawling, filtering and pre-processing Web content to projecting, labeling and rendering the aggregated information. Incremental processing steps are localized in the projection stage consisting of document clustering, cluster force-directed placement and fast document positioning. We evaluate the proposed framework by contrasting layout qualities of incremental versus non-incremental versions. Documents for the experiments stem from the blog sample of the Media Watch on Climate Change (www.ecoresearch.net/climate). Experimental results indicate that our incremental computation approach is capable of accurately generating dynamic information landscapes.

Sentiment detection analyzes the positive or negative polarity of text. The field has received considerable attention in recent years, since it plays an important role in providing means to assess user opinions regarding an organization's products, services, or actions. Approaches towards sentiment detection include machine learning techniques as well as computationally less expensive methods. Both approaches rely on the use of language-specific sentiment lexicons, which are lists of sentiment terms with their corresponding sentiment value. The effort involved in creating, customizing, and extending sentiment lexicons is considerable, particularly if less common languages and domains are targeted without access to appropriate language resources. This paper proposes a semi-automatic approach for the creation of sentiment lexicons which assigns sentiment values to sentiment terms via crowd-sourcing. Furthermore, it introduces a bootstrapping process operating on unlabeled domain documents to extend the created lexicons, and to customize them according to the particular use case. This process considers sentiment terms as well as sentiment indicators occurring in the discourse surrounding a articular topic. Such indicators are associated with a positive or negative context in a particular domain, but might have a neutral connotation in other domains. A formal evaluation shows that bootstrapping considerably improves the method's recall. Automatically created lexicons yield a performance comparable to professionally created language resources such as the General Inquirer.

By providing interoperability and shared meaning across actors and domains, lightweight domain ontologies are a cornerstone technology of the Semantic Web. This chapter investigates evidence sources for ontology learning and describes a generic and extensible approach to ontology learning that combines such evidence sources to extract domain concepts, identify relations between the ontology’s concepts, and detect relation labels automatically. An implementation illustrates the presented ontology learning and relation labeling framework and serves as the basis for discussing possible pitfalls in ontology learning. Afterwards, three use cases demonstrate the usefulness of the presented framework and its application to real-world problems.

Recent research shows the potential of utilizing data collected through Web 2.0 applications to capture domain evolution. Relying on external data sources, however, often introduces delays due to the time spent retrieving data from these sources. The method introduced in this paper streamlines the data acquisition process by applying optimal stopping theory. An extensive evaluation demonstrates how such an optimization improves the processing speed of an ontology refinement component which uses Delicious to refine ontologies constructed from unstructured textual data while having no significant impact on the quality of the refinement process. Domain experts compare the results retrieved from optimal stopping with data obtained from standardized techniques to assess the effect of optimal stopping on data quality and the created domain ontology.

Recent research shows the potential of utilizing data collected through Web 2.0 applications to capture changes in a domain's terminology. This paper presents an approach to augment corpus-based ontology learning by considering terms from collaborative tagging systems, social networking platforms, and micro-blogging services. The proposed framework collects information on the domain's terminology from domain documents and a seed ontology in a triple store. Data from social sources such as Delicious, Flickr, Technorati and Twitter provide an outside view of the domain and help incorporate external knowledge into the ontology learning process. The neural network technique of spreading activation is used to identify relevant new concepts, and to determine their positions in the extended ontology. Evaluating the method with two measures (PMI and expert judgements) demonstrates the significant benefits of social evidence sources for ontology learning.

Digital ecosystems typically involve a large number of participants from different sectors who generate rapidly growing archives of unstructured text. Measuring the frequency of certain terms to determine the popularity of a topic is comparably straightforward. Detecting sentiment expressed in user-generated electronic content is more challenging, especially in the case of digital ecosystems comprising heterogeneous sets of multilingual documents. This paper describes the use of language-specific grammar patterns and multilingual tagged dictionaries to detect sentiment in German and English document repositories. Digital ecosystems may contain millions of frequently updated documents, requiring sentiment detection methods that maximize throughput. The ideal combination of high-throughput techniques and more accurate (but slower) approaches depends on the specific requirements of an application. To accommodate a wide range of possible applications, this paper presents (i) an adaptive method, balancing accuracy and scalability of multilingual textual sources, (ii) a generic approach for generating language- specific grammar patterns and multilingual tagged dictionaries, and (iii) an extensive evaluation verifying the method's performance based on Amazon product reviews and user evaluations from Sentiment Quiz, a “game with a purpose” that invites users of the Facebook social networking platform to assess the sentiment of individual sentences.

This paper presents a method for integrating DBpedia data into an ontology learning system that automatically suggests labels for relations in domain ontologies based on large corpora of unstructured text. The method extracts and aggregates verb vectors for semantic relations identified in the corpus. It composes a knowledge base which consists of (i) centroids for known relations between domain concepts, (ii) mappings between concept pairs and the types of known relations, and (iii) ontological knowledge retrieved from DBpedia.Refining similarities between the verb centroids of labeled and unlabeled relations by means of including domain and range constraints applying DBpedia data yields relation type suggestions. A formal evaluation compares the accuracy and average ranking performance of this hybrid method with the performance of methods that solely rely on corpus data and those that are only based on reasoning and external data sources.

Analysts are often interested in how sentiment towards an organization, a product or a particular technology changes over time. Popular methods that process unstructured textual material to automatically detect sentiment based on tagged dictionaries are not capable of fulfilling this task, even when coupled with part-of speech tagging, a standard component of most text processing toolkits that distinguishes grammatical categories such as article, noun, verb, and adverb. Small corpus size, ambiguity and subtle incremental change of tonal expressions between different versions of a document complicate sentiment detection. Parsing grammatical structures, by contrast, outperforms dictionary-based approaches in terms of reliability, but usually suffers from poor scalability due to its computational complexity. This work provides an over view of different dictionary- and machine-learning-based sentiment detection methods and evaluates them on several Web corpora.After identifying the shortcomings of these methods, the paper proposes an approach based on automatically building Tagged Linguistic Unit (TLU) databases to overcome the restrictions of dictionaries with a limited set of tagged tokens.

The rapid increase in the amount of available information from various online sources poses new challenges for programs that endeavor to process these sources automatically and identify the most relevant material for a given application.

The advantages and positive effects of tightly coupled interface components on search performance have been documented in several studies. This paper focuses on the implementation of tightly coupled views within the Media Watch on Climate Change, an interactive Web portal (www.ecoresearch.net/climate) combining a portfolio of semantic services with a visual exploration and information retrieval interface. The portal builds contextualized information spaces by (a) enriching the content repository with spatial, semantic, and temporal annotations, and (b) applying semi-automated ontology learning to the repository yielding a controlled vocabulary that helps structuring the stored information. Different portlets visualize aspects of the contextualized information spaces, providing the user with multiple views on the available information. Currently, synchronized semantic maps, domain ontologies, geographic maps, tag clouds and real-time information retrieval agents suggesting similar topics and nearby locations, provide users with important context information, facilitate the access to complex datasets, and help users navigate large collections of Web documents.

In the case of tourism applications, it is particularly evident that geography is emerging as a fundamental principle for structuring Web resources. Recent improvements in semantic and geographic Web technology, often referred to as the Geospatial Web, acknowledge the relevance of adding location metadata to existing databases and accessing the vast amounts of information stored in these databases via geospatial services. This paper outlines the acquisition of geospatial context information, describes usage scenarios and real-world applications in the tourism industry, and presents an automated software tool for annotating large collections of Web documents automatically. The quality of this tool is tested based upon Web pages from the Austrian National Tourism Organization. Initial results are encouraging and help define a roadmap for further improving the automated tagging of tourism resources.

Ontology evolution is an intrinsic phenomenon of any knowledge-intensive system, which can be addressed either implicitly or explicitly. This paper describes an explicit, data-driven approach to capture and visualize ontology evolution by semi-automatically extending small seed ontologies. This process captures ontology changes reflected in large document collections. The visualization of these changes helps characterize the evolution process, and distinguish core, extended and peripheral relations between concepts.

The identification and labeling of non-hierarchical relations are among the most challenging tasks in ontology learning. This paper describes an approach for suggesting ontology relationship types to domain experts based on implicitly learned relations from a domain corpus. The learning process extracts verb- vectors from sentences containing domain concepts. It computes centroids for known relationship types and stores them in the knowledge base. Vectors of unknown relationships are compared to the stored centroids using the cosine similarity metric. The system then suggests the relationship type of the most similar centroid. Domain experts evaluate these suggestions to refine the knowledge base and constantly improve the component's accuracy. Using four sample ontologies on "energy sources", this paper demonstrates how link type suggestion aids the ontology design process. It also provides a statistical analysis on the accuracy and average ranking performance of batch learning versus online learning.

Semantic Web technologies in general and ontology-based approaches in particular are considered the foundation for the next generation of information services. While ontologies enable software agents to exchange knowledge and information in a standardized, intelligent manner, describing today\(\check{\rm S}\)s vast amount of information in terms of ontological knowledge remains a challenge.

Semantic Web technologies in general and ontologybased approaches in particular are considered the foundation for the next generation of information services. While ontologies enable software agents to exchange knowledge and information in a standardised, intelligent manner, describing todays vast amount of information in terms of ontological knowledge and to track the evolution of such ontologies remains a challenge. In this paper we describe Web2.0 crawling for ontology evolution. The World Wide Web, or Web for short, is due, its evolutionary properties and social network characteristics a perfect fitting data source to evolve an ontology. The decentralised structure of the Internet, the huge amount of data and upcoming Web2.0 technologies arise several challenges for a crawling system. In this paper we present a distributed crawling system with standard browser integration. The proposed system is a high performance, sitescript based noise reducing crawler, which loads standard browser equivalent content from Web2.0 resources. Furthermore we describe the integration of this spider into our ontology evolution framework.

Linking named entities to structured knowledge sources paves the way for state-of-the-art Web intelligence applications which assign sentiment to the correct entities, identify trends, and reveal relations between organizations, persons and products. For this purpose this paper introduces Recognyze, a named entity linking component that uses background knowledge obtained from linked data repositories, and outlines the process of transforming heterogeneous data silos within an organization into a linked enterprise data repository which draws upon popular linked open data vocabularies to foster interoperability with public data sets. The presented examples use comprehensive real-world data sets from Orell Füssli Business Information, Switzerland's largest business information provider. The linked data repository created from these data sets comprises more than nine million triples on companies, the companies' contact information, key people, products and brands. We identify the major challenges of tapping into such sources for named entity linking, and describe required data pre-processing techniques to use and integrate such data sets, with a special focus on disambiguation and ranking algorithms. Finally, we conduct a comprehensive evaluation based on business news from the New Journal of Zurich and AWP Financial News to illustrate how these techniques improve the performance of the Recognyze named entity linking component.

This paper presents technical foundation, roadmap and initial results of the IDIOM project (Information Diffusion across Interactive Online Media). Information spreads rapidly across Web sites, Web logs and online forums. IDIOM tracks this process and compares it to direct communication through electronic mail and instant messaging. Linguists define "idiom" as an expression whose meaning is different from the literal meanings of its component words. Similarly, the study of information diffusion promises insights that cannot be inferred from individual network elements. Similar projects often focus on particular media, or neglect important aspects of the human language. IDIOM addresses these gaps to reveal fundamental mechanisms of information diffusion across media with distinct interactive characteristics.

The identification and labelling of non-hierarchical relations are among the most challenging tasks in ontology learning. This paper describes a bottom-up approach for automatically suggesting ontology link types. The presented method extracts verb-vectors from semantic relations identified in the domain corpus, aggregates them by computing centroids for known relation types, and stores the centroids in a central knowledge base. Comparing verb-vectors extracted from unknown relations with the stored centroids yields link type suggestions. Domain experts evaluate these suggestions, refining the knowledge base and constantly improving the component's accuracy. A final evaluation provides a detailed statistical analysis of the introduced approach.

Information spreads rapidly across websites and other online media. The IDIOM research project analyses this process by identifying redundant content elements, mapping them to ontology concepts, and tracking their temporal and geographic distribution. Linguists define 'idiom' as an expression whose meaning is different from the literal meanings of its component words. Similarly, investigating information diffusion promises insights that cannot be inferred from individual network elements. Previous research often focused on particular media, or neglected important aspects of the human language. IDIOM addresses these gaps to reveal fundamental mechanisms of information diffusion across media with distinct interactive characteristics.

Rigorous evaluations and analyses of evaluation results are key towards improving Named Entity Linking systems. Nevertheless, most current evaluation tools are focused on benchmarking and comparative evaluations. Therefore, they only provide aggregated statistics such as precision, recall and F1-measure to assess system performance and no means for conducting detailed analyses up to the level of individual annotations. This paper addresses the need for transparent benchmarking and fine-grained error analysis by introducing Orbis, an extensible framework that supports drill-down analysis, multiple annotation tasks and resource versioning. Orbis complements approaches like those deployed through the GERBIL and TAC KBP tools and helps developers to better understand and address shortcomings in their Named Entity Linking tools. We present three uses cases in order to demonstrate the usefulness of Orbis for both research and production systems: (i) improving Named Entity Linking tools; (ii) detecting gold standard errors; and (iii) performing Named Entity Linking evaluations with multiple versions of the included resources.

Opinion holder and opinion target extraction are among the most popular and challenging problems tackled by opinion mining researchers, recognizing the significant business value of such components and their importance for applications such as media monitoring and Web intelligence. This paper describes an approach that combines opinion target extraction with aspect extraction using syntactic patterns. It expands previous work limited by sentence boundaries and includes a heuristic for anaphora resolution to identify targets across sentences. Furthermore, it demonstrates the application of concepts known from research on open information extraction to the identification of relevant opinion aspects. Qualitative analyses performed on a corpus of 100,000 Amazon product reviews show that the approach is promising. The extracted opinion targets and aspects are useful for enriching common knowledge resources and opinion mining ontologies, and support practitioners and researchers to identify opinions in document collections.

Annotation styles express guidelines that direct human annotators in what rules to follow when creating gold standard annotations of text corpora. These guidelines not only shape the gold standards they help create, but also influence the training and evaluation of Named Entity Linking (NEL) tools, since different annotation styles correspond to divergent views on the entities present in the same texts. Such divergence is particularly present in texts from the media domain that contain references to creative works. In this work we present a corpus of 1000 annotated documents selected from the media domain. Each document is presented with multiple gold standard annotations representing various annotation styles. This corpus is used to evaluate a series of Named Entity Linking tools in order to understand the impact of the differences in annotation styles on the reported accuracy when processing highly ambiguous entities such as names of creative works. Relaxed annotation guidelines that include overlap styles lead to better results across all tools.

Dependency parsing is considered a key technology for improving information extraction tasks. Research indicates that dependency parsers spend more than 95% of their total runtime on feature computations. Based on this insight, this paper investigates the potential of improving parsing throughput by designing feature representations which are optimized for combining single features to more complex feature templates and by optimizing parser constraints. Applying these techniques to MDParser increased its throughput four fold, yielding Syntactic Parser, a dependency parser that outperforms comparable approaches by factor 25 to 400.

Ontology evolution is an intrinsic phenomenon of any knowledge-intensive system, which can be addressed either implicitly or explicitly. This paper describes an approach to capture and visualize implicit data-driven ontology evolution using ontologies semi-automatically generated by extending small seed ontologies. This process captures ontology changes reflected in large document collections. Visualizing of these changes helps characterize the evolution process, and distinguish core, extended and peripheral relations between concepts. Finally, the paper presents an example of ontology evolution by monitoring and analyzing online media coverage on "energy sources" over a period of ten months.