Metalanguage as an Interdisciplinary Classifier for Mathematics and Computer Science

This repository contains all materials and code related to our capstone project, which investigates whether math and computer science texts can still be differentiated after removing domain-specific vocabulary. We aim to discover if a “metalanguage” remains that is strong enough for classification and clustering across various subfields.

---

Repository Structure

.
├── PhaseA
│   ├── PhaseA_Book.docx
│   └── PhaseA_Presentation.pptx
│   
├── PhaseB
│   ├── PhaseB_Book.docx
│   ├── PhaseB_Presentation.pptx
│   ├── User_Guide.docx
│   └── Maintenance_Guide.docx
│   
├── PDF_to_TXT_pipeline.ipynb
│   
└── classification_pipeline.ipynb
│   
└── README.md

PhaseA

• Book and Presentation describing our initial project proposal, literature review, and the scope of expected work.

PhaseB

• Book and Presentation detailing our final methodology, experiments, and findings.
• User Guide explaining how to run the pipelines and replicate our experiments.
• Maintenance Guide offering instructions for updating dependencies, troubleshooting, and extending the system.

Code

1. PDF_to_TXT_pipeline.ipynb

   • Scripts for converting PDFs to raw text using libraries such as PyMuPDF and Tesseract (for scanned PDFs).

2. classification_pipeline.ipynb

   • Preprocessing (lemmatization, domain-term removal, chunking).
   • Classification (XLNet fine-tuning on the preprocessed text).
   • Clustering (Doc2Vec embeddings fed into K-Means, PAM, DBSCAN, and Gaussian Mixture).

---

Project Goal

Many students find mathematics and computer science challenging, especially because each discipline has its own specialized terminology. In this project, we strip away domain-specific words (e.g., “matrix,” “stack,” “derivative”) to see if the remaining general language—“metalanguage”—still contains enough clues to classify the text by field or subfield. Our main goals are:

1. Evaluate classification performance using XLNet on text that lacks obvious “giveaway” terms.
2. Apply clustering algorithms to see if unsupervised methods replicate domain or subfield distinctions without key terminology.
3. Investigate whether mathematics and computer science exhibit a shared “metalanguage,” or if subfields (e.g., Set Theory vs. Abstract Algebra) retain enough unique style and structure to remain distinct once domain-specific words are removed.

---

Pipeline Overview

1. PDF Conversion
   • Extract text from both digital and scanned PDFs.
2. Text Preprocessing
   • Remove irrelevant pages (TOC, references), convert to lowercase, lemmatize, filter out domain-specific words.
3. Segmentation
   • Split the cleaned text into chunks.
4. Classification with XLNet
   • Fine-tune a transformer model on these chunks to see how well it distinguishes math vs. CS, as well as subfields.
5. Clustering
   • Convert chunks to Doc2Vec embeddings.
   • Run K-Means, PAM, DBSCAN, and GMM to see how unsupervised clusters align with known labels.

---

Main Findings

• Classification: XLNet achieved ~83% accuracy, showing that even after domain-specific terms are removed, each discipline still has distinctive linguistic cues.
• Clustering: Unsupervised methods tend to split the data into broad Math vs. CS clusters (e.g., K=2), but subfields often overlap, indicating shared language structures within each domain.
• Conclusion: A persistent “metalanguage” remains, yet distinguishing closely related subfields (like Calculus vs. Probability) is more challenging without specialized terms.

---

For detailed instructions, see the User Guide in the PhaseB folder.

---