README.md

---
language:
- en
- tl
tags:
- sarcasm-detection
- mock-politeness
- multi-task-learning
- code-mixed
license: mit
base_model:
- FacebookAI/xlm-roberta-base
---

# XLM-RoBERTa with Multi-Task Learning for Sarcasm and Mock Politeness Detection

## Model Description
This project fine-tunes **XLM-RoBERTa** for detecting **sarcasm** and **mock politeness** in **Filipino (English, Tagalog, or code-mixed (Taglish))** faculty evaluation texts.  

Two models are included:
- **MTL model** → sarcasm detection (main task) + mock politeness detection (auxiliary task)  
- **STL model** → sarcasm detection only  

The models are packaged into a **desktop app (Tkinter + Python)** for easy testing.

---

## Intended Uses & Limitations

### Intended Use
- Demonstrating multi-task learning in NLP  
- Exploring sarcasm and politeness detection in Taglish text  
- Academic/research purposes only  

### Limitations
- Trained on a domain-specific dataset (faculty evaluations)  
- May not generalize well outside Taglish or academic settings  
- Predictions are not guaranteed to be accurate for all contexts  

---

## How to Use

1. Download the **XLM-R folder** from this repository.  
2. Inside the folder, locate and open:  XLM-R/XLM-R.exe
3. Use the GUI to input text or upload a `.csv` file (see included `INPUT_SAMPLE.csv`).  
4. The app will output predictions for sarcasm (and mock politeness if using MTL).  

*(No coding required — the `.exe` is standalone on Windows.)*

---

## Training Data
- Collected faculty evaluation texts written in **Filipino**  (English, Tagalog, or code-mixed (Taglish)) 
- Annotated for sarcasm and mock politeness  

---

## Evaluation
- Compared **Single-Task (STL)** vs **Multi-Task (MTL)**  
- Metrics: accuracy, precision, recall, F1