|
|
--- |
|
|
license: cc-by-4.0 |
|
|
datasets: |
|
|
- DSL-13-SRMAP/Telugu-Dataset |
|
|
language: |
|
|
- te |
|
|
tags: |
|
|
- sentiment-analysis |
|
|
- text-classification |
|
|
- telugu |
|
|
- indic-languages |
|
|
- muril |
|
|
- rationale-supervision |
|
|
- explainable-ai |
|
|
base_model: google/muril-base-cased |
|
|
pipeline_tag: text-classification |
|
|
metrics: |
|
|
- accuracy |
|
|
- f1 |
|
|
- auroc |
|
|
--- |
|
|
|
|
|
# MuRIL_WR |
|
|
|
|
|
## Model Description |
|
|
|
|
|
**MuRIL_WR** is a Telugu sentiment classification model built on **MuRIL (Multilingual Representations for Indian Languages)**, a Transformer-based BERT model specifically designed for **Indian languages**, including Telugu and English. |
|
|
|
|
|
MuRIL is pretrained on a **large and diverse corpus of Indian language text**, including web data, religious scriptures, and news content. In contrast to general multilingual models such as mBERT and XLM-R, MuRIL is better suited to capture **Telugu morphology, syntax, and linguistic structure**. |
|
|
|
|
|
The suffix **WR** denotes **With Rationale supervision**. This model is fine-tuned using both **sentiment labels and human-annotated rationales**, enabling improved alignment between model predictions and human-identified evidence. |
|
|
|
|
|
--- |
|
|
|
|
|
## Pretraining Details |
|
|
|
|
|
- **Pretraining corpus:** Indian language text from web sources, religious texts, and news data |
|
|
- **Training objectives:** |
|
|
- Masked Language Modeling (MLM) |
|
|
- Translation Language Modeling (TLM) |
|
|
- **Language coverage:** 17+ Indian languages, including Telugu and English |
|
|
|
|
|
--- |
|
|
|
|
|
## Training Data |
|
|
|
|
|
- **Fine-tuning dataset:** Telugu-Dataset |
|
|
- **Task:** Sentiment classification |
|
|
- **Supervision type:** Label + rationale supervision |
|
|
- **Rationales:** Token-level human-annotated evidence spans |
|
|
|
|
|
--- |
|
|
|
|
|
## Rationale Supervision |
|
|
|
|
|
During fine-tuning, **human-provided rationales** guide model learning. Alongside the standard classification loss, an **auxiliary rationale loss** encourages the model’s attention or explanation scores to align with annotated rationale tokens. |
|
|
|
|
|
This approach improves: |
|
|
|
|
|
- Interpretability of sentiment predictions |
|
|
- Alignment between model explanations and human judgment |
|
|
- Plausibility of generated explanations |
|
|
|
|
|
--- |
|
|
|
|
|
## Intended Use |
|
|
|
|
|
This model is intended for: |
|
|
|
|
|
- Explainable Telugu sentiment classification |
|
|
- Rationale-supervised learning experiments |
|
|
- Indian-language explainability research |
|
|
- Comparative evaluation against label-only (WOR) baselines |
|
|
|
|
|
MuRIL_WR is particularly effective for **informal, conversational, and social media Telugu text**, where rationale supervision further enhances explanation quality. |
|
|
|
|
|
--- |
|
|
|
|
|
## Performance Characteristics |
|
|
|
|
|
Compared to label-only training, rationale supervision typically improves **explanation plausibility** while maintaining competitive sentiment classification performance. |
|
|
|
|
|
### Strengths |
|
|
|
|
|
- Strong Telugu-specific linguistic modeling |
|
|
- Human-aligned explanations via rationale supervision |
|
|
- Suitable for explainable AI benchmarking in Indian languages |
|
|
|
|
|
### Limitations |
|
|
|
|
|
- Requires human-annotated rationales, increasing annotation effort |
|
|
- Pretraining data bias toward informal text may affect formal Telugu tasks |
|
|
- Classification gains over WOR may be modest |
|
|
|
|
|
--- |
|
|
|
|
|
## Use in Explainability Evaluation |
|
|
|
|
|
**MuRIL_WR** is well-suited for evaluation with explanation frameworks such as FERRET, enabling: |
|
|
|
|
|
- **Faithfulness evaluation:** How well explanations support model predictions |
|
|
- **Plausibility evaluation:** How closely explanations align with human rationales |
|
|
|
|
|
--- |
|
|
|
|
|
## References |
|
|
|
|
|
- Khanuja et al., 2021 |
|
|
- Joshi, 2022 |
|
|
- Das et al., 2022 |
|
|
- Rajalakshmi et al., 2023 |
