Event-Centric Reasoning for Interpreting Everyday Narratives – ERIANA

When a human listens to or reads a conversation in natural language, they begin by identifying the entities involved, the relationships between these entities, and how events are formed and interconnected. They then use general knowledge and common sense to interpret what is happening and predict future events. This reasoning process relies on a combination of observation, logic, and implicit knowledge.

The ERIANA project aims to replicate this process in artificial intelligence. Our approach seeks to capture and formalize the implicit knowledge contained in a text, structuring it within a framework that facilitates efficient reasoning. To achieve this, we combine symbolic artificial intelligence (ensuring reasoning transparency) and digital artificial intelligence (optimizing model efficiency and adaptability).

Methodology

Our approach is based on three complementary components:

1. Knowledge Representation and Structuring: We design a framework that encodes general knowledge related to entities, relationships, and events in a narrative. This involves developing an ontological space that models concepts and their interactions, associating each element with semantic and logical properties.

2. Event Learning and Extraction from Narratives

We develop symbolic representations of narratives, structuring narrative information by considering temporal, causal, and logical relationships between facts. These representations are enriched using deep learning models capable of inferring implicit knowledge from large text corpora. We implement two complementary approaches: i) A numerical approach, based on large language models (LLMs), which learn from vast text datasets, and ii) A neuro-symbolic approach, integrating logical rules to structure and enhance AI reasoning.

3. Implementation of an Advanced Reasoning Layer

Unlike traditional models that rely on simple rule-based sequences, we develop a dynamic reasoning system that accounts for context and interactions between events, improving the interpretation of narratives.

Each model developed within these three components undergoes empirical validation to assess its effectiveness and impact on narrative comprehension.

The list of articles related to the project is available in open access on HAL and the project website, where the associated resources can also be found.

With the rise of large language models (LLMs), attention has focused on their fine-tuning and utilization for complex linguistic tasks. However, precisely modeling entities, concepts, and relationships as static vectors remains a major challenge. Our work explores multiple approaches to learning semantic vector representations, which are essential for plausible reasoning. We have developed three learning strategies, leading to several publications and the release of freely accessible resources.

Natural language serves as an alternative to formal logics for representing and utilizing knowledge. Recently, interest has grown in extracting knowledge from LLMs and leveraging it through prompting approaches. We developed a similarity-based algorithm that automatically optimizes prompts by adjusting their parameters based on the representation of knowledge.

However, reasoning in natural language presents a challenge, as traditional automated reasoning methods do not apply directly. To address this, we designed transformer-based language models capable of reasoning over knowledge expressed in natural language. Additionally, we reformulated knowledge base completion as a natural language inference task, using LLMs to enrich knowledge bases.

Using LLMs for reasoning also raises major concerns regarding unintended biases inherited from training data. To address this issue, we introduced REFINE-LM, a debiasing method based on reinforcement learning. REFINE-LM corrects biases related to gender, ethnicity, religion, and nationality without requiring human annotations or costly fine-tuning. Unlike existing approaches, REFINE-LM effectively reduces biases while preserving model performance, using a lightweight model trained on word probability distributions. A demonstration of REFINE-LM was presented to the public at the European Conference on Artificial Intelligence and is available here: biasinai.github.io/refinedemo/.

Finally, the research conducted within ERIANA is also applied in various real-world domains, such as document processing in medical and legal fields, as well as in the discovery of new materials in chemistry.

Despite the advancements made within the ERIANA project, artificial intelligence systems still struggle to capture the fundamental conceptual structures that humans naturally use to understand and interact with their environment. Our research has achieved promising results in automated reasoning on narratives, with applications in various fields, such as document processing in medical and legal contexts, knowledge extraction in chemistry, and bias reduction in language models through the REFINE-LM method.

However, one key observation remains: despite these advances, current systems still struggle to structure their understanding of the world as fluidly and intuitively as humans do. To address this challenge, we are exploring a new research framework, inspired by embodied cognition theory and applied to intelligent agent systems. This framework is based on modeling recurring sensorimotor experience patterns that shape human cognition. The goal is to adapt LLMs to translate natural language descriptions into formal representations, grounded in sensorimotor patterns, to develop a neurosymbolic system that anchors agents’ understanding in fundamental conceptual structures.

This approach paves the way for greater interpretability and efficiency in AI systems while enabling more intuitive human-agent interactions through a shared understanding of fundamental concepts. Beyond the initially envisioned applications in ERIANA, this research represents a major step toward AI models that can reason in a way more closely aligned with human reasoning, thereby enhancing their adaptability and explainability.

Making sense of everyday narratives is a highly challenging task, which requires a deep understanding of the language, and abundant background knowledge about the world. While human readers can rely on high-level reasoning to draw conclusions, current NLP models largely lack this ability. Commonsense knowledge plays a crucial role in interpreting everyday narratives, which generally "pop-up" as supplementary assumptions and expectations and can be used "on-the-fly" when drawing high-level (context-level) reasoning about events. The majority of existing approaches of language understanding, especially those relying on end-to-end neural network models, mainly focus on performing low-level forms of reasoning at the sentence level to achieve tasks. However, if we want to move forward we need high-level reasoning abilities that combine commonsense knowledge in a principled way.

This project aims to develop an nterpretable, modular and neuro-symbolic approach of high-level reasoning about everyday narratives. Reasoning about narratives is event-centric, which intrinsically relies on event affordances (aspects or expectations), and the interactions (e.g. causality or temporal) between events. I will first develop a neuro-symbolic representation framework of narratives that permits encoding events along with their affordances, and interactions between them in an interpretable way. Based on such framework, I will then develop deep generative models leading to discover commonsense knowledge that can be seen as symbolic rules. Once learned, those rules will be combined transparently, using Mixture-of-Experts-like models to implement high-level reasoning methods, similarly to how rules are chained in logical-based formalism. By engaging with practice throughout the project, several methods, benchmarks and applications will be developed.