AI symbolic for enhancing reasoning and trustworthiness of GPT

Problem description

The latest versions of GPT are achieving unprecedented capabilities for assisting humans in the completion of diverse and complex tasks. Despite their extraordinary performance, GPT models remain opaque and prone to hallucinations. Moreover, GPT models are still underperforming when completing complex reasoning processes. These problems are due to the lack of factual knowledge, because during training, GPT memorizes facts defined in the training data, but it is not always able to recall the correct facts and often experience hallucinations by generating statements that are factually incorrect [Pan 2024]. This issue severely affects the confidence of users about applying GPT for some complex tasks. To handle these shortcomings, the research community and the industry is exploring the possibility of integrating symbolic AI methods, which includes rule and reasoning engines and knowledge graphs. There is a consensus that symbolic systems and Large Language Models (LLMs), such as GPT, can be considered complementary to each other’s: while the former ones are human-interpretable, deterministic, and parsimonious in terms of data, the latter ones are inherently opaque, indeterministic, and data-hungry. At the same time, while symbolic approaches often require human experts to manually encode symbolic knowledge, LLMs, such as GPT, typically support some form of automatic learning from data [Calegari 2020].    

Goal of the thesis

The goal of this thesis is to explore the use of symbolic AI (e.g., knowledge graphs and reasoning engines) to enhance reasoning capabilities of LLMs, such as GPT, and reduce hallucinations and opaqueness, and improve trustworthiness. At the same time, it would be interesting to explore how LLMs can augment symbolic systems by providing natural language interfaces and reducing the cost of building large knowledge bases (aka knowledge graphs) that support these systems. This might require applying advanced prompt engineering techniques and fine-tuning.

Expected results and learning outcome

After completing the MSc thesis, the student should learn how to improve the performance of GPT using prompt engineering techniques, fine-tuning and symbolic systems such as knowledge graphs.

The expected results of the thesis might include the implementation of a prototype that integrates a GPT model with a symbolic AI system. This might include, for instance (non-exhaustive list): 

• A fine-tuned version of GPT trained with factual knowledge from existent knowledge graphs capable to execute natural language queries (or any complex task that requires some reasoning capabilities) over these knowledge graphs
• A RAG (retrieval augmented generation) configuration, where a knowledge graph database is used to validate and enhance the answers produced by GPT from natural language queries (or any complex task that requires some reasoning capabilities).

In addition, the student must validate the performance of the prototype using existent benchmarks or creating new ones tailored to the specific task to be solved.

Desired qualifications

Candidates should have a good understanding on deep learning techniques, data engineering, and semantic technologies. Moreover, it will be recommended some experience programming in Python with libraries for data processing (e.g., Pandas, SQLAlchemy, etc.), data analytics (NumPy, Scikit-learn, TensorFlow, PyTorch, etc.) and data visualisation (e.g., Matplotlib, Seaborne, etc.).

Some relevant courses at UiO: TEK5040, IN5550, IN3060, IN2090, IN5800 and IN3110.

References

1. Bhuyan, Bikram Pratim, et al. "Neuro-symbolic artificial intelligence: a survey." Neural Computing and Applications (2024): 1-36.
2. Calegari, Roberta, et al. ”On the integration of symbolic and sub-symbolic techniques for XAI: A survey.” Intelligenza Artificiale, IOS Press (2020).
3. Khorashadizadeh, Hanieh, et al. "Research Trends for the Interplay between Large Language Models and Knowledge Graphs." arXiv preprint arXiv:2406.08223 (2024).
4. Pan, Shirui, et al. "Unifying large language models and knowledge graphs: A roadmap." IEEE Transactions on Knowledge and Data Engineering (2024).
5. Wang, Xin, et al. "Large language model enhanced knowledge representation learning: A survey." arXiv preprint arXiv:2407.00936 (2024).