Multilingual sentiment in NLP - Model Pipeline Trace

Model Pipeline - Multilingual sentiment

This pipeline reads text data in multiple languages, cleans and converts it into numbers, trains a model to understand positive or negative feelings, and then predicts sentiment for new sentences.

Data Flow - 7 Stages

1Data in

5000 rows x 2 columns→Raw text data with columns: 'text' (sentences in English, Spanish, French) and 'label' (0=negative, 1=positive)→5000 rows x 2 columns

"text": 'I love this product', "label": 1

↓

2Preprocessing

5000 rows x 2 columns→Lowercase, remove punctuation, and tokenize text→5000 rows x 2 columns

"text": ['i', 'love', 'this', 'product'], "label": 1

↓

3Feature Engineering

5000 rows x 2 columns→Convert tokens to multilingual word embeddings (300 features per sentence)→5000 rows x 300 features

[0.12, -0.05, ..., 0.33] (embedding vector), label: 1

↓

4Train/Test Split

5000 rows x 300 features→Split data into 4000 training and 1000 testing samples→4000 rows x 300 features (train), 1000 rows x 300 features (test)

Train sample embedding vector with label 1

↓

5Model Trains

4000 rows x 300 features→Train a simple neural network classifier→Trained model

Model learns to map embeddings to sentiment labels

↓

6Metrics Improve

Validation data 1000 rows x 300 features→Evaluate model accuracy and loss improving over epochs→Accuracy and loss values per epoch

Epoch 10: loss=0.25, accuracy=0.90

↓

7Prediction

New sentence embedding vector (300 features)→Model predicts sentiment probability→Probability scores for negative and positive classes

[0.1, 0.9] means 90% positive sentiment

Training Trace - Epoch by Epoch

Loss
0.7 |****
0.6 |*** 
0.5 |**  
0.4 |*   
0.3 |    
0.2 |    
     1 2 3 4 5 6 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.65	0.60	Model starts learning basic patterns
2	0.50	0.72	Accuracy improves as model adjusts weights
3	0.40	0.80	Model captures multilingual sentiment features
4	0.32	0.85	Loss decreases steadily, accuracy rises
5	0.28	0.88	Model converging with good performance
6	0.25	0.90	Final epoch with best validation accuracy

Prediction Trace - 3 Layers

Layer 1: Input embedding

Layer 2: Neural network hidden layer (ReLU activation)

Layer 3: Output layer (Softmax)

Model Quiz - 3 Questions

What happens to the data shape after converting text to embeddings?

A Rows reduce by half, columns stay the same

B Rows stay the same, columns change from text to 300 features

C Rows increase, columns reduce to 1

D Rows and columns both stay the same

Key Insight

This visualization shows how multilingual text is turned into numbers that a model can learn from. The model improves by reducing mistakes (loss) and increasing correct guesses (accuracy). Softmax outputs give clear probabilities for each sentiment class.