| # sentiment_analysis_bert_multilingual |
| |
| ## Overview |
| This model is a fine-tuned version of the Multilingual BERT (mBERT) base model. It is designed to classify the sentiment of text across 100+ languages into three categories: Negative, Neutral, and Positive. |
| |
| ## Model Architecture |
| The model utilizes the standard BERT-base architecture: |
| - **Layers**: 12 Transformer blocks |
| - **Hidden Size**: 768 |
| - **Attention Heads**: 12 |
| - **Parameters**: ~177M |
| It includes a sequence classification head on top of the hidden state of the `[CLS]` token. |
| |
| ## Intended Use |
| - Social media monitoring for global brands. |
| - Customer feedback analysis in multilingual support tickets. |
| - Market research across different geographical regions. |
| |
| ## Limitations |
| - **Context Window**: Limited to 512 tokens; longer texts will be truncated. |
| - **Sarcasm**: May struggle with highly idiomatic or sarcastic expressions in low-resource languages. |
| - **Bias**: Subject to biases present in the Wikipedia and BookCorpus datasets used for pre-training. |
