Manipulation de Données Structurées et Interaction avec des Outils Externes grâce aux LLMs : L'Avenir de l'IA Générative dans les Systèmes d'Information ?

Laure SoulierINFORSID 2024 · OpenReview

Abstract

This paper explores the transition of Large Language Models (LLMs) from conversational interfaces to core functional components of enterprise Information Systems (IS). By focusing on the manipulation of structured data—such as SQL databases and knowledge graphs—and the orchestration of external APIs, it outlines a framework for "hybrid intelligence." This approach seeks to bridge the gap between rigid data architectures and fluid human interaction, moving beyond the "chatbot" paradigm to create systems capable of synthesizing complex, factually grounded information.

Introduction

The rapid ascent of Generative AI, spearheaded by Large Language Models like ChatGPT, has fundamentally altered the landscape of human-computer interaction. However, within the context of enterprise Information Systems (IS), the integration of these models remains largely superficial, often limited to front-end conversational layers. This research, presented by Dr. Laure Soulier at INFORSID 2024, argues for a deeper integration. The goal is to leverage the generalization and adaptation capabilities of LLMs to transform how structured data is valued and utilized within organizations.

The Challenge: Bridging the Gap in Data Valuation

The primary hurdle in modern Information Systems is the disconnect between how data is stored and how it is consumed. Enterprise intelligence typically resides in rigid, structured formats: SQL databases, knowledge graphs, and specialized APIs. While these structures ensure precision, they are often inaccessible to non-technical stakeholders without intermediary layers.

The research addresses the "valuation" of this structured data. The challenge lies in creating a functional layer where LLMs act as a bridge, capable of understanding the underlying logic of structured sources and translating them into human-readable insights without sacrificing the factual integrity of the source data.

Methodological Framework

The proposed framework for integrating LLMs into Information Systems rests on three methodological pillars:

  1. Mechanistic Exploration: This involves analyzing the "inner workings" of LLMs to understand how they solve increasingly complex tasks. By understanding these internal processes, researchers can better predict and control how models interact with structured logic.
  2. Structured Data Valuation (Data-to-Text): This pillar focuses on the translation of raw database or graph information into natural language. By utilizing LLMs for data-to-text tasks, systems can provide human-readable summaries of complex datasets, making the information more accessible and actionable.
  3. Tool Use and API Interaction: Perhaps the most transformative aspect is the implementation of LLMs as autonomous agents. In this capacity, the LLM is not just a generator of text but a controller that can call external functions or APIs. This allows the system to gather and synthesize information that is not present in its static training data, effectively expanding its knowledge base in real-time.

Key Insights: Toward Hybrid Intelligence

The research highlights several critical shifts in how we should perceive LLMs in a business context:

  • Beyond Mimicry: The true potential of LLMs lies in their ability to generalize and adapt to new tasks, rather than their ability to mimic human speech.
  • The IS Integration Frontier: There is a necessary shift from viewing LLMs as standalone tools to seeing them as integral components of the IS architecture. This allows for a synergy where the LLM provides the linguistic interface while the Information System provides the factual grounding.
  • Mitigating Hallucinations: By anchoring LLMs in structured data and external APIs—a concept referred to as hybrid intelligence—systems can overcome the inherent limitations of pure generative models, such as the tendency to produce factually incorrect "hallucinations."

Future Directions and Limitations

While the potential is vast, the integration of LLMs into Information Systems faces technical constraints that require further exploration. Future research directions include:

  • Reliability in Data-to-Text: Ensuring that synthesized information remains 100% faithful to the underlying structured data is paramount for enterprise adoption.
  • Cross-Domain Application: The principles of LLM-driven data manipulation are already being applied to fields like robotics, suggesting that these models will eventually bridge the gap between digital information systems and physical-world applications.
  • Information Access: Deepening the synergy between LLMs and traditional Machine Learning to improve how users discover and interact with complex information ecosystems.

Summary updated

BibTeX
@inproceedings{soulier2024manipulation,
  author = {Laure Soulier},
  title = {{Manipulation de Données Structurées et Interaction avec des Outils Externes grâce aux LLMs : L'Avenir de l'IA Générative dans les Systèmes d'Information ?}},
  booktitle = {INFORSID},
  year = {2024},
  url = {https://openreview.net/forum?id=xHXNhaONMc},
}
Figure