LLM-based web agents have the potential to automate long-running web tasks, such as searching for products in multiple e-shops and subsequently ordering the cheapest products that meet the user's needs. Benchmarks for evaluating web agents either require agents to perform tasks online using the live Web or offline using simulated environments, the latter allowing for the exact reproduction of the experimental setup. While DeepShop and ShoppingComp provide online benchmarks that require agents to perform challenging shopping tasks, existing offline benchmarks such as WebShop, WebArena, and Mind2Web cover only comparatively simple e-commerce tasks performed against a single shop containing product data from a single source. What is missing is an e-commerce benchmark that simulates multiple shops containing heterogeneous product data and requires agents to perform complex retrieval tasks. We fill this gap by introducing WebMall, the first offline multi-shop benchmark for evaluating web agents on challenging comparison shopping tasks. WebMall consists of four simulated shops populated with product data extracted from the Common Crawl. The WebMall tasks range from specific product searches and price comparisons to advanced searches for complementary or substitute products, as well as checkout processes. We validate WebMall using eight agents that differ in observation space, availability of short-term memory, and the employed LLM. The validation highlights the difficulty of the benchmark, with the best-performing agents achieving task completion rates below 65% in the task categories cheapest product search and vague product search.

LLM-based agents are increasingly used to automate web tasks such as product search, offer comparison, and order placement. Current research explores different interfaces through which these agents interact with websites, including traditional HTML browsing, retrieval-augmented generation (RAG) over pre-crawled content, communication via Web APIs using the Model Context Protocol (MCP), and natural-language querying through the NLWeb interface. Yet no systematic comparison of the effectiveness and efficiency of these interfaces on identical challenging task sets exists. To address this gap, we introduce a testbed consisting of four simulated e-shops, each offering its products via HTML, MCP, and NLWeb interfaces. For each interface (HTML, RAG, MCP, and NLWeb), we develop specialized agents that perform the same sets of tasks, ranging from simple product searches and price comparisons to complex queries for complementary or substitute products and checkout processes. We evaluate the agents using GPT-5 and GPT-5-mini. Our evaluation shows that RAG, MCP, and NLWeb agents outperform HTML browsing agents by 11 percentage points in task completion while requiring 2–5 times fewer tokens on search-oriented tasks. The GPT-5 RAG agent achieves the highest task completion rate (0.79) while maintaining moderate token consumption.

An increasing number of data providers have adopted shared numbering schemes such as GTIN, ISBN, DUNS, or ORCID numbers for identifying entities in the respective domain. This means for data integration that shared identifiers are often available for a subset of the entity descriptions to be integrated while such identifiers are not available for others. The challenge in these settings is to learn a matcher for entity descriptions without identifiers using the entity descriptions containing identifiers as training data. The task can be approached by learning a binary classifier which distinguishes pairs of entity descriptions for the same real-world entity from descriptions of different entities. The task can also be modeled as a multi-class classification problem by learning classifiers for identifying descriptions of individual entities. We present a dual-objective training method for BERT, called JointBERT, which combines binary matching and multi-class classification, forcing the model to predict the entity identifier for each entity description in a training pair in addition to the match/non-match decision. Our evaluation across five entity matching benchmark datasets shows that dual-objective training can increase the matching performance for seen products by 1% to 5% F1 compared to single-objective Transformer-based methods, given that enough training data is available for both objectives. In order to gain a deeper understanding of the strengths and weaknesses of the proposed method, we compare JointBERT to several other BERT-based matching methods as well as baseline systems along a set of specific matching challenges. This evaluation shows that JointBERT, given enough training data for both objectives, outperforms the other methods on tasks involving seen products, while it underperforms for unseen products. Using a combination of LIME explanations and domain-specific word classes, we analyze the matching decisions of the different deep learning models and conclude that BERT-based models are better at focusing on relevant word classes compared to RNN-based models.

Entity matching is a central task in data integration which has been researched for decades. Over this time, a wide range of benchmark tasks for evaluating entity matching methods has been developed. This resource paper systematically complements, profiles, and compares 21 entity matching benchmark tasks. In order to better understand the specific challenges associated with different tasks, we define a set of profiling dimensions which capture central aspects of the matching tasks. Using these dimensions, we create groups of benchmark tasks having similar characteristics. Afterwards, we assess the difficulty of the tasks in each group by computing baseline evaluation results using standard feature engineering together with two common classification methods. In order to enable the exact reproducibility of evaluation results, matching tasks need to contain exactly defined sets of matching and non-matching record pairs, as well as a fixed development and test split. As this is not the case for some widely-used benchmark tasks, we complement these tasks with fixed sets of non-matching pairs, as well as fixed splits, and provide the resulting development and test sets for public download. By profiling and complementing the benchmark tasks, we support researchers to select challenging as well as diverse tasks and to compare matching systems on clearly defined grounds.

Cross-domain knowledge bases such as YAGO, DBpedia, or the Google Knowledge Graph are being used as background knowledge within an increasing range of applications including web search, data integration, natural language understanding, and question answering. The usefulness of a knowledge base for these applications depends on its completeness. Relational HTML tables from the Web cover a wide range of topics and describe very specific long tail entities, such as small villages, less-known football players, or obscure songs. \nThis systems and applications paper explores the potential of web table data for the task of completing cross-domain knowledge bases with descriptions of formerly unknown entities. We present the first system that handles all steps that are necessary for this task: schema matching, row clustering, entity creation, and new detection. The evaluation of the system using a manually labeled gold standard shows that it can construct formerly unknown instances and their descriptions from table data with an average F1 score of 0.80. In a second experiment, we apply the system to a large corpus of web tables extracted from the Common Crawl. This experiment allows us to get an overall impression of the potential of web tables for augmenting knowledge bases with long tail entities. The experiment shows that we can augment the DBpedia knowledge base with descriptions of 14 thousand new football players as well as 187 thousand new songs. The accuracy of the facts describing these instances is 0.90.

Comparison shopping portals integrate product offers from large numbers of e-shops in order to support consumers in their buying decisions. Product offers often consist of a title and a free-text product description, both describing product attributes that are considered relevant by the specific vendor. In addition, product offers might contain structured or semi-structured product specifications in the form of HTML tables and HTML lists. As product specifications often cover more product attributes than free-text descriptions, being able to extract attribute-value pairs from these specifications is a critical prerequisite for achieving good results in tasks such as product matching, product categorisation, faceted product search, and product recommendation.

HTML tables on web pages ("web tables") cover a wide variety of topics. Data from web tables can thus be useful for tasks such as knowledge base completion or ad hoc table extension. Before table data can be used for these tasks, the tables must be matched to the respective knowledge base or base table. The challenges of web table matching are the high heterogeneity and the small size of the tables. Though it is known that the majority of web tables are very small, the gold standards that are used to compare web table matching systems mostly consist of larger tables. In this experimental paper, we evaluate T2K Match, a web table to knowledge base matching system, and COMA, a standard schema matching tool, using a sample of web tables that is more realistic than the gold standards that were previously used. We find that both systems fail to produce correct results for many of the very small tables in the sample. As a remedy, we propose to stitch (combine) the tables from each web site into larger ones and match these enlarged tables to the knowledge base or base table afterwards. For this stitching process, we evaluate different schema matching methods in combination with holistic correspondence refinement. Limiting the stitching procedure to web tables from the same web site decreases the heterogeneity and allows us to stitch tables with very high precision. Our experiments show that applying table stitching before running the actual matching method improves the matching results by 0.38 in F1-measure for T2K Match and by 0.14 for COMA. Also, stitching the tables allows us to reduce the amount of tables in our corpus from 5 million original web tables to as few as 100,000 stitched tables.

Relational tables collected from HTML pages ("web tables") are used for a variety of tasks including table extension, knowledge base completion, and data transformation. Most of the existing algorithms for these tasks assume that the data in the tables has the form of binary relations, i.e., relates a single entity to a value or to another entity. Our exploration of a large public corpus of web tables, however, shows that web tables contain a large fraction of non-binary relations which will likely be misinterpreted by the state-of-the-art algorithms. In this paper, we propose a categorisation scheme for web table columns which distinguishes the different types of relations that appear in tables on the Web and may help to design algorithms which better deal with these different types. Designing an automated classifier that can distinguish between different types of relations is non-trivial, because web tables are relatively small, contain a high level of noise, and often miss partial key values. In order to be able to perform this distinction, we propose a set of features which goes beyond probabilistic functional dependencies by using the union of multiple tables from the same web site and from different web sites to overcome the problem that single web tables are too small for the reliable calculation of functional dependencies.

A subset of the HTML tables on the Web contains relational data. The data in these tables covers a multitude of topics and is thus very useful for complementing or validating cross-domain knowledge bases, such as DBpedia, YAGO, or the Google Knowledge Graph. A large fraction of the data in these knowledge bases is time-dependent, meaning that the correctness of an attribute value depends on a point in time. Fusing data from web tables in order to determine correct values for time-dependent attributes is challenging as most web tables do not contain timestamp information. A possibility to deal with this sparsity is to exploit timestamps which appear in different locations on the web page around the table. But as these timestamps might not apply to the web table value in question, this approach introduces noise. This paper investigates the extent to which the performance of data fusion strategies that rely on voting, PageRank, and Knowledge-Based-Trust can be improved by incorporating noisy and sparse timestamp information. For this, we present a machine-learning-based approach which considers different types of noisy timestamps in the data fusion process, and experiment with propagating timestamp information between web tables in order to overcome sparsity. We evaluate the data fusion strategies using a large public corpus of web tables and a public gold standard of time-dependent attribute values. We find that our methods effectively choose and weigh timestamp information per attribute and reduce sparsity using propagation. By incorporating timestamp information into data fusion strategies that previously did not exploit temporal meta information, we are able to increase F1-measure on average by 5%.

Cross-domain knowledge bases such as DBpedia, YAGO, or the Google Knowledge Graph have gained increasing attention over the last years and are starting to be deployed within various use cases. However, the content of such knowledge bases is far from being complete, far from always being correct, and suffers from deprecation (i.e. population numbers become outdated after some time). Hence, there are efforts to leverage various types of Web data to complement, update and extend such knowledge bases. A source of Web data that potentially provides a very wide coverage are millions of relational HTML tables that are found on the Web. The existing work on using data from Web tables to augment cross-domain knowledge bases reports only aggregated performance numbers. The actual content of the Web tables and the topical areas of the knowledge bases that can be complemented using the tables remain unclear. In this paper, we match a large, publicly available Web table corpus to the DBpedia knowledge base. Based on the matching results, we profile the potential of Web tables for augmenting different parts of cross-domain knowledge bases and report detailed statistics about classes, properties, and instances for which missing values can be filled using Web table data as evidence. In order to estimate the potential quality of the new values, we empirically examine the Local Closed World Assumption and use it to determine the maximal number of correct facts that an ideal data fusion strategy could generate. Using this as ground truth, we compare three data fusion strategies and conclude that knowledge-based trust outperforms PageRank- and voting-based fusion.

The lecture discusses the concept of Search Joins. A Search Join is a join operation which extends a local table with ad-ditional attributes based on the large corpus of structured data that is published on the Web in various formats. A Search Join takes as input a local table, a corpus of het-erogeneous Web tables, and a description of the attributes that should be added to the local table. The challenges that Search Joins need to handle are threefold: 1. Determine the set of the top-k Web tables which are beneficial candi-dates for the join operation; 2. Join the local table with the top-k candidate tables given no external knowledge about key attributes; 3. Merge corresponding attributes and fuse attribute values in order to return a concise result table con-taining high-quality data. Search Joins are useful in various application scenarios. They allow for example a local table about cities to be ex-tended with an attribute containing the average temperature of each city for manual inspection [5]. They also allow tables to be extended with large sets of additional attributes as a basis for data mining, for instance to identify factors that might explain why the inhabitants of one city claim to be happier than the inhabitants of another [7]. Existing work on extending local tables with additional at-tributes from the Web mainly focused on corpora of HTML tables extracted from Web crawls [3][4][8][9]. The recent increase in the adoption of Linked Data publishing [2], Mi-crodata and RDFa annotations [1] as well as the growth of public data repositories such as datahub.io and data.gov.uk make a wide range of larger tables available on the Web and enable Search Joins to exploit these more comprehensive data sets. In the lecture, I will draw a theoretical framework for Search Joins and will survey the state of the art methods employed by Search Join systems to handle the challenges outlined above. Afterward, I will highlight how the recent developments in the context of Linked Data, RDFa and Microdata publishing, public data repositories, as well as

Linked Data on the Web is either created from structured data sources (such as relational databases), from semi-structured sources (such as Wikipedia), or from unstructured sources (such as text). In the latter two cases, the generated Linked Data will likely be noisy and incomplete. In this paper, we present two algorithms that exploit statistical distributions of properties and types for enhancing the quality of incomplete and noisy Linked Data sets: SDType adds missing type statements, and SDValidate identifies faulty statements. Neither of the algorithms uses external knowledge, i.e., they operate only on the data itself. We evaluate the algorithms on the DBpedia and NELL knowledge bases, showing that they are both accurate as well as scalable. Both algorithms have been used for building the DBpedia 3.9 release: With SDType, 3.4 million missing type statements have been added, while using SDValidate, 13,000 erroneous RDF statements have been removed from the knowledge base.

A central problem in data integration and data cleansing is to find entities in different data sources that describe the same real-world object. Many existing methods for identifying such entities rely on explicit linkage rules which specify the conditions that entities must fulfill in order to be considered to describe the same real-world object. In this paper, we present the GenLink algorithm for learning expressive linkage rules from a set of existing reference links using genetic programming. The algorithm is capable of generating linkage rules which select discriminative properties for comparison, apply chains of data transformations to normalize property values, choose appropriate distance measures and thresholds and combine the results of multiple comparisons using non-linear aggregation functions. Our experiments show that the GenLink algorithm outperforms the state-of-the-art genetic programming approach to learning linkage rules recently presented by Carvalho et. al. and is capable of learning linkage rules which achieve a similar accuracy as human written rules for the same problem.

Enriching knowledge bases with multimedia information makes it possible to complement textual descriptions with visual and audio information. Such complementary information can help users to understand the meaning of assertions, and in general improve the user experience with the knowledge base. In this paper we address the problem of how to enrich ontology instances with candidate images retrieved from existing Web search engines. DBpedia has evolved into a major hub in the Linked Data cloud, interconnecting millions of entities organized under a consistent ontology. Our approach taps into the Wikipedia corpus to gather context information for DBpedia instances and takes advantage of image tagging information when this is available to calculate semantic relatedness between instances and candidate images. We performed experiments with focus on the particularly challenging problem of highly ambiguous names. Both methods presented in this work outperformed the baseline. Our best method leveraged context words from Wikipedia, tags from Flickr and type information from DBpedia to achieve an average precision of 80%.

In 2005, Michael Franklin, Alon Halevy, and David Maier coined the term dataspaces as a new abstraction and target architecture for data management. In 2006, Tim Berners-Lee introduced the Linked Data principles a set of best practices for publishing and interlinking structured data on the Web in accordance with the general architecture of the Web. These two lines of thinking have come together and are realized in the Web of Linked Data - a global public dataspace. The Web of Linked Data was kick started by the W3C Linking Open Data community effort in 2007. Today, the Web of Linked Data consists of hundreds of datasets published by universities, companies, government and public sector bodies, as well as by individual Web enthusiasts. The content of the Web of Linked Data is diverse in nature, comprising data about people, organizations, products, geographic locations, books, scientific publications, films, music, television and radio programs, genes, proteins, drugs and clinical trials, online communities, statistical data, census results, and reviews. In addition to publishing and interlinking datasets, there is intensive work on Linked Data browsers, Web of Linked Data search engines and other applications that consume Linked Data from the Web. In his talk, Christian Bizer will introduce the principle ideas behind Linked Data and relate them to the dataspace architecture. Afterwards, he will give an overview of the identity resolution and pay-as-you-go data integration techniques that are currently used in the context of the Web of Linked Data and will explain the state-of-the-art in applications that consume Linked Data from the Web.

The term “Linked Data” refers to a set of best practices for publishing and connecting structured data on the Web. These best practices have been adopted by an increasing number of data providers over the last three years, leading to the creation of a global data space containing billions of assertions— the Web of Data. In this article, the authors present the concept and technical principles of Linked Data, and situate these within the broader context of related technological developments. They describe progress to date in publishing Linked Data on the Web, review applications that have been developed to exploit the Web of Data, and map out a research agenda for the Linked Data community as it moves forward.

The SPARQL Query Language for RDF and the SPARQL Protocol for RDF are implemented by a growing number of storage systems and are used within enterprise and open Web settings. As SPARQL is taken up by the community, there is a growing need for benchmarks to compare the performance of storage systems that expose SPARQL endpoints via the SPARQL protocol. Such systems include native RDF stores as well as systems that rewrite SPARQL queries to SQL queries against non-RDF relational databases. This article introduces the Berlin SPARQL Benchmark (BSBM) for comparing the performance of native RDF stores with the performance of SPARQL-to-SQL rewriters across architectures. The benchmark is built around an e-commerce use case in which a set of products is offered by different vendors and consumers have posted reviews about products. The benchmark query mix emulates the search and navigation pattern of a consumer looking for a product. The article discusses the design of the BSBM benchmark and presents the results of a benchmark experiment comparing the performance of four popular RDF stores (Sesame, Virtuoso, Jena TDB, and Jena SDB) with the performance of two SPARQL-to-SQL rewriters (D2R Server and Virtuoso RDF Views) as well as the performance of two relational database management systems (MySQL and Virtuoso RDBMS).

The Web is increasingly understood as a global information space consisting not just of linked documents, but also of Linked Data. More than just a vision, the resulting Web of Data has been brought into being by the maturing of the Semantic Web technology stack, and by the publication of an increasing number of datasets according to the principles of Linked Data.

The Semantic Web consists of many RDF graphs nameable by URIs. This paper extends the syntax and semantics of RDF to cover such Named Graphs. This enables RDF statements that describe graphs, which is beneficial in many Semantic Web application areas. As a case study, we explore the application area of Semantic Web publishing: Named Graphs allow publishers to communicate assertional intent, and to sign their graphs; information consumers can evaluate specific graphs using task-specific trust policies, and act on information from those Named Graphs that they accept. Graphs are trusted depending on: their content; information about the graph; and the task the user is performing. The extension of RDF to Named Graphs provides a formally defined framework to be a foundation for the Semantic Web trust layer.