Link Prediction (LP) approaches based on Language Models (LMs) operate over the labels and descriptions of entities and relations in a Knowledge Graph (KG). Recent approaches have shown that incorporating a local graph neighborhood can improve the LP capabilities of LMs. These approaches usually sample a context from the neighborhood around a query triple randomly, thereby incorporating noise that might hinder the model in making correct predictions.

Federated learning (FL) facilitates shared training of machine learning models while maintaining data privacy. Unfortunately, it suffers from data imbalance among participating clients, causing the performance of the shared model to drop. To diminish the negative effects of unfavorable data-specific properties, both algorithm- and data-based approaches seek to make FL more resilient against them. In this regard, data-based approaches prove to be more versatile and require less domain knowledge to be applied efficiently. Hence, they seem particularly suitable for widespread application in various FL environments. Although data-based approaches such as local data sampling have been applied to FL in the past, previous research did not provide a systematic analysis of the potential and limitations of individual data sampling strategies to improve FL. To this end, we (1) identify relevant local data sampling strategies for FL, (2) identify data-specific properties that negatively affect FL performance, and (3) provide a benchmark of local data sampling strategies regarding their effect on model performance, convergence, and training time in synthetic, real-world, and large-scale FL environments. Moreover, we propose and rigorously test a novel method for data sampling in FL that locally optimizes the choice of sampling strategy prior to FL participation. Our results show that FL can benefit from applying local data sampling in terms of performance and convergence rate, especially when data imbalance is high or the number of clients and samples is low. Furthermore, our proposed sampling strategy offers the best trade-off between model performance and training time.

Kenneweg S, Deigmöller J, Cimiano P, Eggert J. Benchmarking the Ability of Large Language Models to Reason About Event Sets. In: Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management. SCITEPRESS - Science and Technology Publications; 2024: 74-82.

As the number of clinical trials carried out and published worldwide keeps growing, better tools for synthesizing the available knowledge become increasingly important. It still requires a significant effort and expertise to aggregate the evidence and results from different clinical trials, a task that is at the core of secondary or comparative studies, meta-analyses, and (living) systematic reviews. Our hypothesis is that the practical challenges involved in synthesizing evidence can be alleviated if the results of clinical trials would be published in a machine-readable format using a well-defined (semantic) vocabulary. Building on the C-TrO ontology that we developed in earlier work to support the aggregation of evidence from clinical trials as the main use case, in this paper we examine the question whether it is feasible for clinical researchers and medical practitioners to describe the results of clinical trials using the C-TrO ontology. For this purpose, we implemented a Web-based tool called CTrO-Editor that uses a form-based interaction paradigm to allow users to enter all the details regarding study population, arms, endpoints, observations and results of a clinical trial, and that exports the data in an RDF format. We describe the results of the evaluation of the CTrO-Editor with five medical students. Our preliminary results suggest that our paradigm for semantifying clinical trials is feasible, as the students could all successfully model a publication of their choice using our tool within a couple of hours.

There is a well-known lexical gap between content expressed in the form of natural language (NL) texts and content stored in an RDF knowledge base (KB). For tasks such as Information Extraction (IE), this gap needs to be bridged from NL to KB, so that facts extracted from text can be represented in RDF and can then be added to an RDF KB. For tasks such as Natural Language Generation, this gap needs to be bridged from KB to NL, so that facts stored in an RDF KB can be verbalized and read by humans. In this paper we propose LexExMachina, a new methodology that induces correspondences between lexical elements and KB elements by mining class-specific association rules. As an example of such an association rule, consider the rule that predicts that if the text about a person contains the token "Greek", then this person has the relation nationality to the entity Greece. Another rule predicts that if the text about a settlement contains the token "Greek", then this settlement has the relation country to the entity Greece. Such a rule can help in question answering, as it maps an adjective to the relevant KB terms, and it can help in information extraction from text. We propose and empirically investigate a set of 20 types of class-specific association rules together with different interestingness measures to rank them. We apply our method on a loosely-parallel text-data corpus that consists of data from DBpedia and texts from Wikipedia, and evaluate and provide empirical evidence for the utility of the rules for Question Answering.

Events play a prominent role in our lives, such that many social media documents describe or are related to some event. Organizing social media documents with respect to events thus seems a promising approach to better manage and organize the ever-increasing amount of content in social media applications. A challenge is to automatize this process so that incoming documents can be assigned to their corresponding event without any user intervention. We present a system that is able to classify a stream of social media data into a growing and evolving set of events. By doing this, we successfully address two key problems that arise in this context: i) scaling to the data sizes and rates encountered in social media applications, and ii) tackling the new event detection problem, i.e. the problem of determining whether an incoming data item belongs to a new or a known event. We successfully address these problems by i) including a candidate retrieval step that retrieves a set of event candidates that the incoming data point is likely to belong to and ii) by including a function trained using machine learning techniques to determine whether the incoming data item belongs to the top scoring candidate or rather to a new event. We show that our system addresses the above mentioned challenging issues successfully and that it outperforms other state-of-the-art approaches in terms of quality and scalability.

As an increasing amount of RDF data is published as Linked Data, intuitive ways of accessing this data become more and more important. Question answering approaches have been proposed as a good compromise between intuitiveness and expressivity. Most question answering systems translate questions into triples which are matched against the RDF data to retrieve an answer, typically relying on some similarity metric. However, in many cases, triples do not represent a faithful representation of the semantic structure of the natural language question, with the result that more expressive queries can not be answered. To circumvent this problem, we present a novel approach that relies on a parse of the question to produce a SPARQL template that directly mirrors the internal structure of the question. This template is then instantiated using statistical entity identification and predicate detection. We show that this approach is competitive and discuss cases of questions that can be answered with our approach but not with competing approaches.

Keyword queries enjoy widespread usage as they represent an intuitive way of specifying information needs. Recently, answering keyword queries on graph-structured data has emerged as an important research topic. The prevalent approaches build on dedicated indexing techniques as well as search algorithms aiming at finding substructures that connect the data elements matching the keywords. In this paper, we introduce a novel keyword search paradigm for graph-structured data, focusing in particular on the RDF data model. Instead of computing answers directly as in previous approaches, we first compute queries from the keywords, allowing the user to choose the appropriate query, and finally, process the query using the underlying database engine. Thereby, the full range of database optimization techniques can be leveraged for query processing. For the computation of queries, we propose a novel algorithm for the exploration of top-k matching subgraphs. While related techniques search the best answer trees, our algorithm is guaranteed to compute all k subgraphs with lowest costs, including cyclic graphs. By performing exploration only on a summary data structure derived from the data graph, we achieve promising performance improvements compared to other approaches.

Increased availability of mobile computing, such as personal digital assistants (PDAs), creates the potential for constant and intelligent access to up-to-date, integrated and detailed information from the Web, regardless of one's actual geographical position.Intelligent question-answering requires the representation of knowledge from various domains, such as the navigational and discourse context of the user, potential user questions, the information provided by Web services and so on, for example in the form of ontologies.Within the context of the SmartWeb project, we have developed a number of domain-specific ontologies that are relevant for mobile and intelligent user interfaces to open-domain question-answering and information services on the Web.To integrate the various domain-specific ontologies, we have developed a foundational ontology, the SmartSUMO ontology, on the basis of the DOLCE and SUMO ontologies.This allows us to combine all the developed ontologies into a single SmartWeb Integrated Ontology 1 (SWIntO) having a common modeling basis with conceptual clarity and the provision of ontology design patterns for modeling consistency.In this paper, we present SWIntO, describe the design choices we made in its construction, illustrate the use of the ontology through a number of applications, and discuss some of the lessons learned from our experiences.1 http://www.smartweb-project.org 2 SmartWeb is a large international

In this poster we present an approach to query answering over knowledge sources that makes use of different ontology management components within an application scenario of the BT Digital Library. The novelty of the approach lies in the combination of different semantic technologies providing a clear benefit for the application scenario considered.

The realization of the Semantic Web is constrained by a knowledge acquisition bottleneck, i.e. the problem of how to add RDF mark-up to the millions of ordinary web pages that already exist. Information Extraction (IE) has been proposed as a solution to the annotation bottleneck. In the task based evaluation reported here, we compared the performance of users without access to annotation, users working with annotations which had been produced from manually constructed knowledge bases, and users working with annotations augmented using IE. We looked at retrieval performance, overlap between retrieved items and the two sets of annotations, and usage of annotation options. Automatically generated annotations were found to add value to the browsing experience in the scenario investigated.

Without the proliferation of formal semantic annotations, the Semantic Web is certainly doomed to failure. In earlier work we presented a new paradigm to avoid this: the 'Self Annotating Web', in which globally available knowledge is used to annotate resources such as web pages. In particular, we presented a concrete method instantiating this paradigm, called PANKOW (Pattern-based ANnotation through Knowledge On the Web). In PANKOW, a named entity to be annotated is put into several linguistic patterns that convey competing semantic meanings. The patterns that are matched most often on the Web indicate the meaning of the named entity --- leading to automatic or semi-automatic annotation.In this paper we present C-PANKOW (Context-driven PANKOW), which alleviates several shortcomings of PANKOW. First, by downloading abstracts and processing them off-line, we avoid the generation of large number of linguistic patterns and correspondingly large number of Google queries.Second, by linguistically analyzing and normalizing the downloaded abstracts, we increase the coverage of our pattern matching mechanism and overcome several limitations of the earlier pattern generation process. Third, we use the annotation context in order to distinguish the significance of a pattern match for the given annotation task. Our experiments show that C-PANKOW inherits all the advantages of PANKOW (no training required etc.), but in addition it is far more efficient and effective.

The success of the Semantic Web depends on the availability of ontologies as well as on the proliferation of web pages annotated with metadata conforming to these ontologies. Thus, a crucial question is where to acquire these metadata from. In this paper wepropose PANKOW (Pattern-based Annotation through Knowledge on theWeb), a method which employs an unsupervised, pattern-based approach to categorize instances with regard to an ontology. The approach is evaluated against the manual annotations of two human subjects. The approach is implemented in OntoMat, an annotation tool for the Semantic Web and shows very promising results.