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.
