# 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.