Knowledge Flow – kFLOW
|Discover the flow of event relations, beyond causality
|The kFLOW project addresses the lack of resources and methods for capturing fine-grained relations between events in text. Existing models focus mainly on temporal links, leaving causal and contextual relations underexplored. By combining ontological modeling, large-scale data creation, and extraction systems, kFLOW establishes the foundations for applications.
|Towards Fine-Grained Understanding of Event Relations
|The overarching goal of the kFLOW project was to enable machines to understand, represent, and reason about the semantic flow of events in natural language, with a focus on fine-grained and interpretable relations. This goal aligns with the broader ambition of bridging the gap between data-driven natural language processing and symbolic knowledge representation, in order to support applications like explainable fact-checking and structured narrative generation. The project was guided by the following scientific objectives, each of which corresponds to a key research question: Understanding how to model event-event relationships in a unified and expressive way → RQ1: How can we formally represent a diverse set of event-event relations, in a unified and machine-readable ontology? The project aimed to explore the ontological foundations of event semantics by reviewing and integrating concepts from linguistics, narratology, and the semantic web. Identifying how to construct high-quality resources for learning event relations → RQ2: How can we build a high-quality, large-scale annotated dataset of fine-grained event relations? A key objective was to develop an annotation methodology combining the expressive power of Large Language Models (LLMs), existing commonsense knowledge bases, and manual validation. Developing methods to extract fine-grained event-event relations from text → RQ3: What approaches can we use to reliably extract refined event relations from unstructured natural language? This objective goes beyond general relation extraction: it requires designing a system that is aligned with the exact set of semantic relations—including causality, enablement, prevention, intentionality—between pairs of events in text. The system must scale to real-world text genres. Assessing the added value of fine-grained event relations in real-world tasks → RQ4: How can these structured event relations be exploited in downstream applications? The project set out to demonstrate the utility of event relation modeling through two concrete use cases: narrative generation and explainable fact-checking. These serves both as validation and as drivers of methodological choices throughout the project.
|To meet its objectives, kFLOW combined methods from knowledge engineering, NLP, and machine learning.
- Ontological modeling: A survey of existing resources (EventKG, CIDOC CRM, discourse theories) revealed gaps in fine-grained relations. To address this, the project developed FARO, an OWL ontology formalizing 25 event-event relations (causal, temporal, comparative, mereological, modal). Logical axioms and constraints ensure consistency and reasoning.
- Dataset creation: A hybrid strategy produced a corpus of 500k+ annotated event pairs. Data was generated via LLM prompting (GPT-3), enriched with commonsense resources (ATOMIC), and partially manually validated. This ensured both scalability and quality.
- Extraction pipelines: Two complementary approaches were implemented: (I) A modular pipeline (event detection, coreference resolution, relation classification with RoBERTa). (II) An end-to-end model jointly predicting events and their relations.
Both systems achieved strong performance, with confidence scores provided for extracted relations.
- Applications and validation: Methods were tested in two demonstrators: a narrative generator creating coherent stories from event graphs, and a fact-checking prototype assessing causal consistency of claims.
|1. FARO: A Comprehensive "Dictionary" of Event Relationships
The project created FARO, a detailed framework that defines 25 types of relationships between events—such as cause and effect, intentions, time sequences, and contradictions. FARO is designed to work seamlessly with existing web standards, enabling machines to logically analyze these connections. All resources are freely available with full documentation: [http://purl.org/faro](http://purl.org/faro).
2. A Massive Dataset for Training AI Systems
To teach machines how to recognize these relationships, the team built a dataset of over 500,000 annotated sentences. These examples were:
- Automatically generated with AI (like GPT-3),
- Filtered for relevance using commonsense knowledge,
- Manually reviewed to ensure quality.
This dataset, the largest of its kind, is freely available to support AI training:
- Dataset: [https://github.com/ANR-kFLOW/Relation_extraction](https://github.com/ANR-kFLOW/Relation_extraction)
- Knowledge Graph: [https://github.com/ANR-kFLOW/knowledge-graph](https://github.com/ANR-kFLOW/knowledge-graph)
3. Smart Tools for Extracting Event Relationships
Two complementary methods were developed:
- A modular approach combining logical rules and advanced classification for clear, interpretable results.
- A neural model that simultaneously identifies events and their relationships, excelling at detecting cause-and-effect and intentional connections.
4. Real-World Applications for Verification and Storytelling
The project created two practical tools:
- A narrative generator that turns event data into coherent stories.
- A fact-checking tool that analyzes the logic between claims and evidence.
These applications demonstrate how a detailed understanding of event relationships improves the clarity, consistency, and transparency of AI systems.
|- Utilisation des grands modèles de langage (LLMs) pour l'augmentation de données
- Au début de kFLOW, les LLMs n'étaient pas largement adoptés pour la création de ressources
- GPT-3 a été utilisé pour générer des exemples annotés à grande échelle de relations entre événements
- Ces exemples ont été filtrés et validés, créant un jeu de données de plus de 500 000 paires d'événements
- Cette approche s'est avérée très efficace et est devenue centrale dans la méthodologie du projet
- LLMs pour la génération de connaissances
- Les LLMs ont été utilisés de manière innovante au-delà de l'augmentation de données
- Ils ont aidé à formuler des questions de compétence pour le développement et l'affinement des ontologies
- Ils ont démontré leur potentiel dans le soutien à l'ingénierie des connaissances et à la recherche basée sur les ontologies
- Extraction de relations spécifiques entre événements
- Concentration sur l'extraction de relations spécifiques et fines entre événements (par exemple, causalité, activation, prévention)
- Les systèmes ont été alignés avec l'ontologie FARO pour une extraction sémantiquement précise
- Avancée dans l'extraction de relations entre événements (ERE) en passant de catégories larges à une précision guidée par l'ontologie
- Génération de récits comme application clé
- La génération de récits est devenue un cas d'usage important
- Développement de méthodes pour transformer des graphes d'événements en récits textuels cohérents
- Montre comment les relations structurées entre événements améliorent la fluidité et la cohérence des récits
- Pose les bases pour des applications futures dans le storytelling, le patrimoine culturel et l'éducation
Human History is composed of a continuous flow of events. Each of them can impact subsequent events and contribute to the evolution of human knowledge. Knowledge Graphs try to encode the information about facts and events, often falling short when representing the temporal evolution of this knowledge and tracking cause-effect flows. kFLOW aims to propose strategies for representing, extracting, predicting and using the information about event relationships and knowledge evolution. For achieving these goals, a Knowledge Graph of interconnected events and facts will be realised. This graph will be populated and exploited through developing specialised strategies for data modelling, information extraction, link prediction, incorrect triple detection and automatic fact-checking.
Project coordination
Pasquale Lisena (EURECOM)
The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.
Partnership
EURECOM EURECOM
Help of the ANR 200,524 euros
Beginning and duration of the scientific project:
December 2021
- 36 Months
