SLICE team : Software and knowledge sciences

The development and evolution of computer systems and information systems are becoming increasingly complex. One of the major factors contributing to this complexity is the growing integration of artificial intelligence components into these systems.

Depending on the nature of the system—AI-powered, AI-enabled, AI-assisted, or AI-driven—artificial intelligence can be either an optional component or a structuring and indispensable element. The different configurations thus reflect a real variability in design and architecture: software systems are not designed, integrated, or maintained according to identical principles.

The objective of this research theme is to analyze and propose architectural models, from both a data and software perspective, as well as models for representing and extracting knowledge specifically adapted to these systems. These approaches include agentic architectures, which place the autonomy and decision-making capacity of agents at the heart of the system.

Software safety and security are among the most essential characteristics of software systems and information systems. Given the increasing complexity of these systems and their execution environments, evaluating and guaranteeing these two characteristics is becoming an increasingly difficult task.

The objective of this research theme is to propose methods for extracting data from software (static analysis, observability, dynamic analysis, etc.), the web (e.g., from various security threat databases), and operational environments. It also involves designing appropriate, federated architectures for this data, which is an essential first step in developing models and analysis processes (heuristics, metaheuristics, symbolic AI, machine learning, deep learning, LLM, etc.).

The latter aim to automatically extract knowledge that is useful for evaluating and guaranteeing software safety properties—in particular through various types of software testing—as well as software security, including vulnerability identification, intrusion detection, remediation, and architectural analysis.

 

This area of the team’s work focuses on exploring and exploiting complex data, particularly when it has multiple dimensions, fuzzy properties, and rich relational structures. In this context, the team develops analysis and modeling methods aimed at extracting relevant knowledge that can be interpreted and used by experts. A significant part of this research is based on symbolic approaches, such as formal concept analysis (construction of conceptual structures and rules of implication or association), or the extraction of frequent patterns and exceptions, which make it possible to highlight regularities, dependencies, and latent structures within the data. The team is also interested in how this knowledge can be retrieved, particularly through the use of large language models (LLMs), in order to facilitate the interpretation of results, dialogue with business experts, and decision support. This research is applied to several fields, including recommendation systems, software data analysis, health data, and agro-ecological data.

In this area, the team’s work will focus on the interface between explainable artificial intelligence and software engineering, with the aim of better understanding, analyzing, and improving the reliability of code and computer models. Particular attention will be paid to the semantics conveyed by identifiers (names of variables, functions, classes, model elements), which are considered essential clues for interpreting design intent and business logic. The flow of code is also viewed as a narrative structure, in which a process can be described as a coherent sequence of events, actions, and dependencies. In this context, the team will focus on approaches that combine the reasoning and generation capabilities of LLMs with knowledge graphs, particularly those derived from resources such as JeuxDeMots, in order to produce analyses that are more contextualized, traceable, and interpretable by experts. This work aims to enhance the explainability of code analysis tools and provide more robust support for model and software control, verification, architecture, and validation activities.