AKASE - Argumentation Knowledge-Graphs for Advanced Search Engines

OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report Table of Contents 1 Project Overview.......................................................................................................................................................5 2 Argument Attribute Identification........................................................................................................................6 2.1 Issue Generation..............................................................................................................................................6 2.1.1 Topics, Issues, and Claims...................................................................................................................6 2.1.2 Data Collection...................................................................................................................................... 7 2.1.3 Embeddings and Clustering............................................................................................................... 7 2.1.4 LLM-based Issue Summarization and Extraction.........................................................................8 2.1.5 Novel Issue Generation.......................................................................................................................9 2.2 Argument Mining.............................................................................................................................................9 2.3 Argument Characterization........................................................................................................................ 10 2.4 Argument Assessment.................................................................................................................................10 2.4.1 Critical Question Generation........................................................................................................... 10 2.4.2 Fallacy Detection.................................................................................................................................11 3 Argument Graph Construction.............................................................................................................................13 4 Integration into Search Engines..........................................................................................................................16 4.1 Argument Presentation – Search...............................................................................................................16 4.2 Argument Presentation – Deliberation....................................................................................................17 5 Conclusion................................................................................................................................................................20 6 References..........................................................................................................................21 1 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report Preliminaries i. Project Info Project number 101070014 Project acronym ows.eu Project name OpenWebSearch.eu –AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines Call HORIZON-CL4-2021-HUMAN-01 Topic HORIZON-CL4-2021-HUMAN-01-05 Type of action HORIZON-RIA Responsible unit DG CNECT Project starting date / Duration 01/09/2022 Project reporting period 1 Project Coordinator Prof. Dr. Michael Granitzer, University of Passau ii. Project Partners Acronym Partner UG University of Groningen iii. Deliverable Info Due Date / Delivery Date 28/10/2025 Deliverable Lead Khalid Al-Khatib Deliverable type Report, R Dissemination level SEN Document Status / Version V2 Work-package / Lead Partner NN 2 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report iv. Deliverable Summary This document describes AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines within the OpenWebSearch.eu project funded by the EC under GA 101070014 in the Horizon Europe framework. It reports on the main outcomes of AKASE, including a large-scale issue space and web argument mining pipeline, a valueand quality-enriched Argumentation Knowledge Graph (AKG) and its large-scale deployment, an argument-aware search prototype with interpretable reranking and RAG summaries, and the ArgsBase multi-agent deliberation platform. The deliverable also summarizes several accepted and in-progress papers at international conferences, together with the accompanying code, data, and tools. v. List of Deliverables Publications Kutepova, M., & Al Khatib, K. (2025, November). Hybrid Intelligence for Logical Fallacy Detection. In Proceedings of the Fourth Workshop on Bridging Human-Computer Interaction and Natural Language Processing (HCI+ NLP) (pp. 197-208). Turkstra, F., Nabhani, S., & Al Khatib, K. (2025, July). TriLLaMa at CQs-Gen 2025: A Two-Stage LLM-Based System for Critical Question Generation. In Proceedings of the 12th Argument mining Workshop (pp. 349-357). Upcoming submissions Turkstra, F., & Al Khatib, K. Deliberation-Based Multi-Agent Approach for Fallacy Detection. Submission deadline: January 5, 2026. Turkstra, F., Nabhani, S., & Al Khatib, K. ArgsBase: A Multi-Agent Interface for Structured Human–AI Deliberation. Submission deadline: December 1, 2025. Related publications Musi, E., Kökciyan, N., Al Khatib, K., Ceolin, D., Dietz, E., Gutekunst, K. M., ... & Wachsmuth, H. (2025, July). Toward reasonable parrots: Why large language models should argue with us by design. In Proceedings of the 12th Argument mining Workshop (pp. 24-31). Heinrich, M., Al Khatib, K., & Stein, B. (2025, July). Multi-Class versus Means-End: Assessing Classification Approaches for Argument Patterns. In Proceedings of the 12th Argument mining Workshop (pp. 195-204). Conferences The 63rd Annual Meeting of the Association for Computational Linguistics. Vienna, Austria (July 27 – August 1, 2025). Oral presentation about Musi et al. (2025, July) and poster presentation about Turkstra et al. (2025, July). The 35th Meeting of Computational Linguistics in The Netherlands. Leuven, Belgium (September 12, 2025). Oral presentation about hybrid argumentation and poster presentation about the upcoming submission: Deliberation-Based Multi-Agent Approach for Fallacy Detection. 3 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report The 2025 Conference on Empirical Methods in Natural Language Processing. Suzhou, China (November 4 – 9, 2025). Oral presentation about Kutepova & Al Khatib (2025, November). Workshops Hybrid Argumentation and Responsible AI. Leiden, the Netherlands (March 31 – April 4, 2025). Khalid Al Khatib was one of the scientific organizers of this workshop. Building Trustworthy and Interactive Recommender Systems through Argumentation. Kamiyamaguchi, Japan (January 19 – 22, 2026). Tools Akase: A search engine powered by our argumentation knowledge graph. Argsbase: A multi-agent interface for structured human–AI deliberation. Awards JTS Early Career Researcher Prize, for the development of the multi-agent deliberation interface described in section 4.2 and in the upcoming submission: ArgsBase: A Multi-Agent Interface for Structured Human–AI Deliberation. The code, data, papers and tools created during this project are made publicly available.1 1 https://github.com/fturkstra-rug/akase 4 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report 1 Project Overview The proliferation of online content has led to an unprecedented abundance of information on complex and often controversial topics—ranging from climate change and artificial intelligence to public policy and health. While this abundance creates opportunities for more informed decision-making, it also brings major challenges: arguments are scattered across diverse sources, expressed in both structured and unstructured forms, and differ widely in clarity, quality, and perspective. This project addresses these challenges by developing a computational framework for extracting, organizing, and presenting argumentative content from the web in a coherent, scalable, and actionable way. At the core of this framework is the Argumentation Knowledge Graph (AKG), which structures argumentation at multiple levels of abstraction. The AKG connects topics, issues, and claims, capturing not only the logical relations between argumentative units but also their underlying human values. By combining state-of-the-art argument mining, clustering, and LLM-based summarization, it enables large-scale analysis, visualization, and reasoning over public argumentation. The graph supports a wide range of use cases. In search, it allows users to retrieve documents based on the structure, coherence, and balance of arguments—not just keyword matches. In deliberation, it powers multi-agent human–AI dialogue, fostering critical thinking, perspective-taking, and reflective reasoning through interactive discussion. The societal relevance of this work lies in its potential to enhance public understanding, support informed decision-making, and promote responsible discourse in areas where arguments strongly influence policy, public opinion, and ethical considerations. This report is organized as follows. Section 2 describes the methods for argument attribute identification, issue generation, and argument mining from both structured debating datasets and web-based sources. Section 3 details the process of graph construction, including node population, relation extraction, and graph expansion. Section 4 presents the integration of the graph into search interfaces and the development of ‘ArgsBase’, our multi-agent deliberation platform. 5 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report 2 Argument Attribute Identification This section describes the methods used to identify, extract, and represent key argumentative attributes that form the foundation of the Argumentation Knowledge Graph (AKG). Argument attributes include issues, claims, values, and relations between argumentative units. Together, they provide the structure required to represent and reason about public discourse in a coherent and interpretable way. We begin with Issue Generation, which organizes raw argumentative content into hierarchical levels (topics, issues, and claims). We then detail Argument Mining, the process of extracting argumentative structures from unstructured web data. Next, we discuss Argument Characterization, where arguments are enriched with human value annotations. Finally, we describe Argument Assessment, focusing on evaluating argument quality through critical question generation and fallacy detection. 2.1 Issue Generation The first step in building the Argumentation Knowledge Graph is identifying what people are arguing about. This involves moving from broad thematic areas to specific, debatable issues, and finally to concrete argumentative statements. To represent this structure computationally, we distinguish between topics, issues, and claims. 2.1.1 Topics, Issues, and Claims In constructing the Argumentation Knowledge Graph, we differentiate between three hierarchical levels of abstraction in argumentative discourse: ● Topics are broad and often divisive themes, such as climate change, artificial intelligence, or immigration policy. ● Issues are narrower sub-problems within these topics, such as government carbon tax policies or AI regulation in hiring. ● Claims are specific argumentative statements, for instance, Governments should implement carbon taxes to reduce emissions. Topics serve as general thematic anchors but are typically too abstract to capture the nuances of debate. Claims, while rich in argumentative detail, are often too granular to serve as graph-level pivots due to their diversity and polarity. Issues, in contrast, strike a useful balance: they are specific enough to enable clustering and comparison but abstract enough to link multiple related claims. This makes issues the most practical unit for constructing the AKG from large-scale web data. 6 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report 2.1.2 Data Collection To populate the initial set of issues, we collected a large corpus of structured argument data from multiple online debating platforms. These platforms contain curated pro/con discussions on a wide range of public controversies, which makes them valuable sources for high-quality argument extraction. The main data sources and counts are shown below: Debate Data 27393 Kialo 1578 Debatebase 683 iSideWith 250 Britannica 101 Total 29965 Table 1. Argument count per data source. Total is after cleaning and deduplication. Each record was parsed to extract issues, arguments, and topics (if explicitly present). In addition, we assigned each entry to one of 16 thematic domains (e.g., Politics & Government, Society & Culture, Science & Technology, Health & Ethics). This semantic categorization forms an upper layer in the knowledge hierarchy, enabling cross-domain analysis and the construction of domain-specific subgraphs. Figure 1: Overview of the pipeline for collecting data from online debating platforms. 2.1.3 Embeddings and Clustering Many of the debating platforms contained overlapping or paraphrased issues. To eliminate redundancy and merge semantically similar issues, we applied a multi-step clustering pipeline combining sentence embeddings, dimensionality reduction, and density-based clustering. Sentence Embedding Generation: Each issue was embedded using gte-Qwen1.5-7B-instruct, currently the top-ranked model on the MTEB (English v2) clustering benchmark. The embeddings are 4096-dimensional floating-point vectors that capture rich semantic features suitable for downstream clustering. 7 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report Dimensionality Reduction: Because high-dimensional embeddings are computationally expensive to cluster, we systematically compared several dimensionality reduction techniques from Pandove et al. (2018), including Principal Component Analysis (PCA), Independent Component Analysis (ICA), Locally Linear Embedding (LLE), Isometric Feature Mapping (Isomap), Multidimensional Scaling (MDS), Uniform Manifold Approximation and Projection (UMAP), and T-distributed Stochastic Neighbor Embedding (t-SNE). These methods were tested on a subsample of 1,000 embeddings using multiple target dimensions (5, 10, 25, 50, 100, 339, 482). UMAP consistently achieved the best clustering performance, yielding the highest silhouette score (0.2265). The 50-dimensional configuration offered the optimal trade-off between cluster cohesion, interpretability, and computational efficiency. Other methods either produced overly coarse groupings or exhibited excessive noise. Clustering: After dimensionality reduction, we applied Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), which is well-suited for large, high-dimensional, and unevenly distributed data. It has the advantage of not requiring a predefined number of clusters and can handle variable cluster densities. With a minimum cluster size of two, HDBSCAN produced 3,003 clusters and 8,904 unclustered points (noise). Qualitative inspection of random samples showed that clusters generally corresponded to coherent and well-formed “issues” that could serve as building blocks for the knowledge graph. 2.1.4 LLM-based Issue Summarization and Extraction Once clustering was complete, each cluster was processed with a LLM to extract a generalized issue statement summarizing its central theme. The goal was to abstract from specific formulations toward a unified, domain-independent description suitable for graph-level representation. The LLM prompt was designed to: 1. Identify the shared issue or theme across all sentences in a cluster. 2. Produce a concise and neutral issue formulation. 3. Merge duplicate or semantically equivalent issues. We evaluated several models, including Claude 3.7 Sonnet, Llama 3.3 70B Instruct, Mistral Large, Phi-4, and DeepSeek R1. A manual evaluation of five clusters and 50 noise samples revealed that Claude 3.7 Sonnet (with reasoning) produced the most coherent and precise issue summaries, albeit with higher computational cost. Llama 3.3 70B provided a strong cost-effective alternative, especially in batch generation. Few-shot prompting (two cluster examples for clusters and ten examples for noise samples) consistently outperformed one-shot and zero-shot prompting, improving contextual consistency and factual grounding. After summarization and merging, the original 29,965 issues were consolidated into 16,020 unique issues. Each issue retained its domain metadata, enabling the generation of domain-specific subgraphs within the larger AKG. 8 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report approach ensures that connections between documents reflect semantic proximity and thematic relevance while maintaining linear scalability and navigability across the entire graph. Following this expansion, the resulting graph contains: ● 51,623 nodes, representing unique argumentative entities (issues, claims, or premises), and ● 470,340 edges, representing argumentative or similarity-based relations. Approximately 91% of all nodes belong to the largest connected component, indicating a highly cohesive and traversable structure. This integrated graph forms the backbone of the Argumentation Knowledge Graph, enabling: ● Cross-document reasoning, by linking related arguments across sources. ● Evidence propagation, by tracing supportive or opposing claims through interconnected domains. ● Exploration of argumentative perspectives, allowing users to navigate from broad topics to specific claims and their justifications. Together, these features make the AKG a scalable and semantically rich resource for large-scale analysis, visualization, and reasoning over public argumentation. 15 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report 4 Integration into Search Engines To make the Argumentation Knowledge Graph (AKG) accessible and useful to end users, we integrated it into two complementary systems: (1) a search-based interface for retrieval, interpretation, and comparison of arguments, and (2) a multi-agent deliberation platform (ArgsBase) that supports structured human–AI dialogue. Together, these interfaces demonstrate the practical applications of the AKG for exploring argumentative content, assessing reasoning quality, and engaging in critical deliberation. 4.1 Argument Presentation – Search The integration of the AKG into a search interface enables users to go beyond traditional keyword retrieval and access argumentatively enriched results. The system not only retrieves documents but also evaluates their argumentative quality, coherence, and balance, presenting transparent and interpretable explanations to users. The search workflow consists of four main stages: 1. Query Input: The user submits a query via a React-based front-end, which provides an intuitive and responsive interface for interaction. 2. Retrieval: The query is sent to the back-end, which retrieves candidate documents using BM25 ranking on both the document title and main content. 3. Reranking: Retrieved documents are reranked using subgraph-based argumentative features that measure justification, coherence, and balance. 4. Presentation: The final ranked list is displayed with (a) a ranking explanation for transparency, (b) a retrieval-augmented generation (RAG) summary for a concise overview, and (c) related issues inferred from the AKG for contextual exploration. Each document in the index is represented as an argumentative subgraph, from which three interpretable metrics are computed for reranking: ● Justification: estimates how well claims are supported, by at least k other nodes. ○ min(#premises / #claims / k, 1) ● Coherence: measures structural connectivity, each node should have on average k edges. ○ min(#edges / #nodes / k, 1) ● Balance: assessing argumentative diversity, there should be both attack and support edges. ○ 1 - |#support - #attack| / (#support + #attack) These scores allow the ranking algorithm to prioritize well-structured, justified, and balanced argumentation over documents retrieved purely by keyword similarity. This provides users not only with relevant results but also with transparent reasoning about why those results are prioritized. 16 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report For each query, the system also employs a retrieval-augmented generation (RAG) component to produce a concise textual summary synthesizing the top-ranked results. The combination of subgraph-based reranking and LLM-generated summaries supports both breadth-oriented exploration (via retrieval) and depth-oriented understanding (via generation). Future work on the search component includes optimizing latency, expanding domain coverage, and refining reranking through multi-level graph features. Additional extensions include the introduction of semantic filters, interactive visualizations, and full graph-augmented search, enabling users to traverse cross-document argumentative relations derived from the global AKG. Figure 3: Overview of the multi-agent, deliberative process. 4.2 Argument Presentation – Deliberation Beyond search and retrieval, we also explored interactive, deliberative engagement with the AKG through ArgsBase, a multi-agent platform designed for structured human–AI dialogue. ArgsBase enables users to deliberate collaboratively with multiple large language models (LLMs), guided by a moderator and supported by a real-time analyzer. Unlike conventional chat interfaces, ArgsBase treats LLMs as reasoning partners rather than information providers, fostering reflective, multi-perspective, and epistemically responsible deliberation. The platform consists of four main components: 1. Moderator: Coordinates the discussion by managing turn-taking, defining the topic scope, enforcing participation rules, prompting clarifications, and maintaining coherence throughout the dialogue. 17 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report 2. Deliberator Agents: Multiple LLM agents contribute claims, justifications, counterarguments, and refinements. They dynamically respond to one another and to the human user, ensuring diversity of reasoning and balanced exploration of perspectives. 3. Human User: Participates actively in the discussion while the moderator ensures engagement without imposing procedural overhead. The human can introduce arguments, request clarifications, or ask for synthesis at any point. 4. Analyzer: Operates in real time to summarize the discussion, identify key agreements and open questions, and produce an argument map visualizing the evolving relationships between claims. The user interface emphasizes clarity and engagement. It features a main dialogue panel for the discussion, a user input area, and a side panel displaying dynamic summaries and the argument map. This visualization helps users trace reasoning patterns and follow the progression of debate in an interpretable way. A formative user study conducted with computational linguistics students evaluated usability and perceived impact. Participants found the system intuitive to use, helpful for exploring multiple perspectives, and effective in visualizing argument structures. Notably, they preferred multi-agent interaction over single-LLM responses due to its greater diversity and structured guidance. Observed limitations included occasional conversational instability, instances of superficial agreement or disagreement between agents, and challenges balancing turn length with content richness. Future evaluations will focus on goal-oriented deliberations to assess practical benefits for decision-making and learning outcomes. ArgsBase demonstrates that structured multi-agent deliberation can significantly enhance critical thinking, perspective-taking, and transparent reasoning. Its applications span research, education, and decision-support domains, illustrating how the AKG can serve as both a computational and an epistemic resource for human–AI collaboration. 18 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report Figure 4: Overview of the multi-agent, deliberative process, Figure 5: Screenshot of the deliberation chat interface with the main discussion panel on the left, and the real-time analyzer panels on the right. 19 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report 5 Conclusion This project addressed a comprehensive set of tasks in computational argumentation—spanning issue extraction, argument mining, relation classification, human value annotation, critical question generation, fallacy detection, and graph construction. Across these nine components, we developed and evaluated tailored computational methods that integrate machine learning, large language model (LLM) reasoning, and large-scale data processing to ensure efficiency, scalability, and accuracy. The outcome is a high-quality Argumentation Knowledge Graph (AKG) that organizes and connects argumentative insights from tens of thousands of documents. The AKG enables users to explore arguments at multiple levels of abstraction—linking topics, issues, claims, and values—while assessing argument quality and uncovering the underlying motivations that shape public reasoning. Beyond offline analysis, we translated these computational advances into interactive applications. The search interface supports transparent, argument-aware document retrieval and ranking through interpretable argumentative features. The ‘ArgsBase’ deliberation platform extends this functionality into multi-agent, human–AI interaction, fostering critical thinking, perspective-taking, and responsible deliberation through structured dialogue and real-time visualization. Together, these contributions demonstrate the feasibility and value of large-scale, computationally mediated argumentation. By uniting structured knowledge representation with interactive, user-centered tools, this work provides a foundation for supporting transparent discourse, responsible reasoning, and informed societal debate. Future work will focus on extending the AKG with dynamic issue generation, multi-modal argument mining, and human-in-the-loop evaluation, aiming to further align computational reasoning with human deliberative practices. 20 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01 OpenWebSearch.eu Third-party Project AKASE – Argumentation Knowledge-Graphs for Advanced Search Engines – Final Report 6 References Figueras, B. C., Agerri, R., Heredia, M., Bengoetxea, J., Cabrio, E., & Villata, S. (2025, July). Overview of the Critical Questions Generation Shared Task. In Proceedings of the 12th Argument mining Workshop (pp. 243-257). Kiesel, J., Çöltekin, Ç., Heinrich, M., Fröbe, M., Alshomary, M., Longueville, B. D., Erjavec, T., Handke, N., Kopp, M., Ljubešić, N., Meden, K., Mirzakhmedova, N., Morkevičius, V., Reitis-Münstermann, T., Scharfbillig, M., Stefanovitch, N., Wachsmuth, H., Potthast, M., & Stein, B. (2024). Overview of Touché 2024: Argumentation systems. In L. Goeuriot, P. Mulhem, G. Quénot, D. Schwab, L. Soulier, G. M. D. Nunzio, P. Galuščáková, A. G. S. de Herrera, G. Faggioli, & N. Ferro (Eds.), Experimental IR Meets Multilinguality, Multimodality, and Interaction: Proceedings of the Fifteenth International Conference of the CLEF Association (CLEF 2024). Springer. Lawrence, J., & Reed, C. (2019). Argument mining: A survey. Computational linguistics, 45(4), 765-818. Legkas, S., Christodoulou, C., Zidianakis, M., Koutrintzes, D., Petasis, G., & Dagioglou, M. (2024). Hierocles of alexandria at touché: multi-task & multi-head custom architecture with transformer-based models for human value detection. In Working Notes of the Conference and Labs of the Evaluation Forum (CLEF 2024). CEUR Workshop Proceedings, CEUR-WS. org. Mancini, E., Ruggeri, F., Colamonaco, S., Zecca, A., Marro, S., & Torroni, P. (2024, August). MAMKit: A Comprehensive Multimodal Argument Mining Toolkit. In Proceedings of the 11th Workshop on Argument Mining (ArgMining 2024) (pp. 69-82). Pandove, D., Goel, S., & Rani, R. (2018). Systematic review of clustering high-dimensional and large datasets. ACM Transactions on Knowledge Discovery from Data (TKDD), 12(2), 1-68. Stab, C., Miller, T., & Gurevych, I. (2018). Cross-topic argument mining from heterogeneous sources using attention-based neural networks. arXiv preprint arXiv:1802.05758. 21 | Project Openwebsearch.eu (GA.101070014) — HORIZON-CL4-2021-HUMAN-01