|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- ar |
|
|
tags: |
|
|
- iraqi-dialect |
|
|
- pragmatics |
|
|
- nlp |
|
|
- social-reasoning |
|
|
- cultural-alignment |
|
|
--- |
|
|
|
|
|
# Project Nabu: A Model-Agnostic Pragmatic Layer |
|
|
|
|
|
**Author:** Abdullah Hawas (Independent Researcher, Iraq) |
|
|
**Paper Title:** Project Nabu: A Model-Agnostic Pragmatic Layer for Social Intent Understanding in Arabic Dialects |
|
|
|
|
|
## 1. Abstract |
|
|
[cite_start]Most natural language processing (NLP) systems rely on surface-level sentiment cues, which leads to systematic failures when processing language in high-context cultures[cite: 7]. [cite_start]**Project Nabu** introduces a model-agnostic pragmatic layer designed to sit on top of any pretrained language model (like BERT or MARBERT), allowing inference of social intent beyond traditional sentiment analysis[cite: 8]. |
|
|
|
|
|
[cite_start]We use **Iraqi Arabic** as a stress-test case due to its dense hierarchical signaling[cite: 9]. [cite_start]Our evaluation on the **ICLE dataset (4,000 annotated sentences)** demonstrates that the Nabu Layer can suppress literal sentiment cues when they conflict with pragmatic intent[cite: 10]. |
|
|
|
|
|
## 2. The Problem: "The Pragmatic Gap" |
|
|
Standard sentiment pipelines often misclassify utterances in hierarchical settings. For example, exaggerated praise or apparent sympathy often serves strategic goals like: |
|
|
* Deference signaling (Respect) |
|
|
* Request softening |
|
|
* [cite_start]Status negotiation [cite: 18] |
|
|
|
|
|
Current models see "Good job" as **Positive**, while Nabu analyzes if it is **Sarcastic** or **Flattery**. |
|
|
|
|
|
## 3. Methodology & Architecture |
|
|
[cite_start]The Nabu Layer operates on the embeddings of a pretrained base model without retraining the base model itself[cite: 37]. |
|
|
|
|
|
### Architecture Design |
|
|
The layer extracts pragmatic features based on: |
|
|
1. **Hierarchical role indicators** |
|
|
2. **Pragmatic trigger density** |
|
|
3. [cite_start]**Status comparison patterns** [cite: 40-42] |
|
|
|
|
|
*(Note: See Figure 1 in the attached Paper PDF for the full diagram)* |
|
|
|
|
|
## 4. Evaluation & Results |
|
|
We evaluated the model on **Test Case II: Sentiment Paradox (Status Inflation)**. |
|
|
[cite_start]Example: *"By God, Professor, frankly you are oppressed in this position, you should be a minister not a manager."* [cite: 78] |
|
|
|
|
|
| Metric | Standard Sentiment | **Nabu Layer** | |
|
|
| :--- | :--- | :--- | |
|
|
| **Interpretation** | Negative (Sadness) | **Strategic Flattery (Hypocrisy)** | |
|
|
| **Confidence** | N/A | [cite_start]**73.04%** [cite: 81] | |
|
|
|
|
|
**Overall Performance:** |
|
|
[cite_start]The Nabu framework achieved an average accuracy of **89%** on the Iraqi Arabic test set, compared to **54%** for standard sentiment classifiers[cite: 94]. |
|
|
|
|
|
## 5. Technical Usage |
|
|
To use the Nabu Layer (Weights coming soon): |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForSequenceClassification, AutoTokenizer |
|
|
|
|
|
# Load the Nabu-Trained Layer |
|
|
model_name = "ay933/Nabu-Iraqi" |
|
|
|
|
|
model = AutoModelForSequenceClassification.from_pretrained(model_name) |
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name) |
