Abstract
Argumentation mining seeks structured representations of argumentative discourse, yet state-of-the-art neural approaches struggle with the uncertainty and combinatorial constraints inherent to these tasks. This work investigates Probabilistic Answer Set Programming (PASP) as a unifying framework that combines the declarative modeling power of ASP with probabilistic semantics. We replace traditional integer linear programming coherence modules with a neuro-symbolic PASP pipeline and study how knowledge compilation can turn PASP programs into differentiable circuits that support scalable learning and inference.
By targeting both max-entropy and credal semantics, the thesis shows how probabilistic circuits enable tractable query answering, gradient-based optimization, and integration with modern deep-learning toolchains such as PyTorch and JAX. The resulting approach broadens the expressiveness of neuro-symbolic argument mining while retaining interpretability.
Project Materials
-
Thesis manuscript sources
Chapters, bibliography, and figures used by the IME-USP thesis template.
-
IME-USP thesis template
Customised LaTeX class files and styles for building the final dissertation (`tese.tex`, `tese.pdf`).
-
Defense slides
Beamer presentation (`slides.tex`, `slides.pdf`) detailing the motivation, methodology, and findings.
-
Poster
Beamer poster layout (`Poster.tex`) summarizing the knowledge-compilation approach for PASP.
Repository Structure
.
├── docs/ # GitHub Pages site (this page)
│ └── index.html
└── text/
├── content/ # Thesis chapters, bibliography, and figures
├── slides/ # Defense presentation (LaTeX + PDF)
├── poster/ # Poster source files
├── proposal/ # Initial research proposal
└── template-ime/ # IME-USP thesis template + compiled thesis
Additional resources, build scripts, and auxiliary files are included in their respective directories. For usage instructions, consult the README in each subdirectory or the template guide.