| # 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. |