Hanna Abi Akl

Current position

I am a PhD student at INRIA as part of the WIMMICS team. My research topic is focused on neuro-symbolic and knowledge-driven approaches for developing reasoning capabilities in LLM-based systems.

PhD Topic summary: Latest advances in Artificial Intelligence (AI) rely on Large Language Model (LLM) systems that demonstrated their capabilities in learning and solving numerous tasks that involve processing and understanding natural text. However, these models still exhibit weaknesses such as their failure to demonstrate sound reasoning, their risk of repeating previously learned information, their strong tendency to hallucinate, and their black-box configuration that make it hard to interpret and explain their results to users. These weaknesses are driving factors behind the reluctance to adopt such systems in many domains that involve decision-making that relies on clear, justifiable, and provable results for users. In the face of these limitations, Neuro-Symbolic (NeSy) approaches have gained increasing popularity as credible solutions to the problems of LLM-based systems. The main appeal behind NeSy systems is the combination of the dual symbolic and neural AI approaches. The promise of these systems is to instill a best-of-both-world approach that taps into the strengths of both types of AI. The symbolic approach brings the logical, rule-based methodology that allows control, traceability and interpretation of results. The connectionist or neural paradigm retains the ability of processing massive amounts of data and the pattern-based matching method that made deep learning a success in solving many problems. Another advantage of neural networks is their ability to handle complex, non-linear data. Logic Neural Networks (LNN) and Logic Tensor Networks (LTN) are some examples of NeSy systems that combine forms of symbolic logic (e.g., First-Order Logic (FOL)) and discrete representations of data (e.g., tensors) with a neural network. On the other hand, the organisation and orchestration of knowledge on the web is governed by the Semantic Web through different data structures such as Knowledge Graphs (KG). These tools play a pivotal role in integrating reasoning in LLM systems. The added value of combining LLMs with knowledge-based methods and structures is to instill real reasoning mechanisms on top of approximative, purely statistical predictions. Knowledge-based methods are starting to be applied in some domains at dataset level (e.g., creating factual datasets with fact justification to train language models to better explain facts) and model level (e.g., using LLMs to explain the relation between different arguments in a discussion). New evaluation metrics have also been introduced in the literature to measure the reasoning capabilities of deep learning systems like graph neural networks on factual bases (e.g., domain ontologies). These methodologies are part of the continuing work on evaluating how well LLMs learn prior knowledge (e.g., knowledge graph data) on the one hand, and trying to improve the effectiveness of knowledge learning and retention on the other. The PhD thesis proposes to tackle the problems related to reasoning in LLM-based systems by studying both NeSy and knowledge-based approaches to develop a more robust reasoning framework. The research will focus on the types of knowledge infused in LLMs, the learning paradigms of LLM-based systems, and the role of the architecture of these models. The work aims to propose alternatives to natural text input data by iterating on Web knowledge representation formats, as well as methods of tracing knowledge acquisition by LLMs to validate the learning and retention of predefined key knowledge during the course of the thesis. The research will also include a review of current popular LLM architectures to propose robust solutions involving symbolic structures to better capture and represent knowledge, as well as adequate paradigms to train them.

I also serve as the Chief Academic Officer at Data ScienceTech Institute and member of the Wimmics research team at INRIA. My research lies at the intersection of logic, language and computation and the application of these areas to knowledge engineering. I am particularly interested in neuro-symbolic AI systems that combine classical symbolic frameworks, neural networks, knowledge graphs and semantic web technology to better understand the mapping of language to symbolic systems and produce better knowledge representation and reasoning methods. My work leverages both theoretical and applied perspectives in interdisciplinary domains (e.g., finance, pharmaceutical, education).

Previous positions

I was a data scientist and machine learning engineer at Yseop in Paris (France) from 2019 to 2023.

I worked as a data integration engineer at Data Consult in Beirut (Lebanon) from 2015 to 2018.

Interests

Neuro-symbolic AI - Natural Language Processing - Knowledge Graphs - Ontologies - Semantic Web