Multilingual sentiment in NLP - ML Experiment: Train & Evaluate

Experiment - Multilingual sentiment

Problem:You want to build a model that can understand if a sentence is positive or negative in multiple languages.

Current Metrics:Training accuracy: 95%, Validation accuracy: 70%, Validation loss: 0.85

Issue:The model is overfitting. It performs very well on training data but poorly on validation data, especially for languages with fewer examples.

Your Task

Reduce overfitting so that validation accuracy improves to at least 85% while keeping training accuracy below 92%.

You can only change the model architecture and training parameters.

You cannot add more data or use external datasets.

Hint 1

Hint 2

Hint 3

Solution

NLP

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Embedding, LSTM, Dense, Dropout
from tensorflow.keras.callbacks import EarlyStopping

# Assume X_train, y_train, X_val, y_val are preprocessed multilingual text data and labels

model = Sequential([
    Embedding(input_dim=10000, output_dim=64, input_length=100),
    LSTM(64, return_sequences=False),
    Dropout(0.5),
    Dense(32, activation='relu'),
    Dropout(0.5),
    Dense(1, activation='sigmoid')
])

model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
              loss='binary_crossentropy',
              metrics=['accuracy'])

early_stop = EarlyStopping(monitor='val_loss', patience=3, restore_best_weights=True)

history = model.fit(X_train, y_train,
                    epochs=20,
                    batch_size=64,
                    validation_data=(X_val, y_val),
                    callbacks=[early_stop])

Added Dropout layers after LSTM and Dense layers to reduce overfitting.

Lowered learning rate to 0.001 for smoother training.

Added EarlyStopping callback to stop training when validation loss stops improving.

Results Interpretation

Before: Training accuracy was 95%, validation accuracy was 70%, showing overfitting.

After: Training accuracy dropped to 90%, validation accuracy improved to 87%, and validation loss decreased, indicating better generalization.

Adding dropout and early stopping helps the model avoid memorizing training data and improves its ability to understand new multilingual sentences.

Bonus Experiment

Try using a pretrained multilingual transformer model like mBERT or XLM-R for sentiment classification.

💡 Hint

Use Hugging Face transformers library to load a pretrained model and fine-tune it on your dataset.

Practice

1. What is the main advantage of using a multilingual sentiment analysis model?

easy

A. It can analyze sentiment in multiple languages with one model.

B. It only works for English text.

C. It requires training a new model for each language.

D. It ignores the language and treats all text the same.

Solution

1. Step 1: Understand multilingual sentiment models

   These models are designed to handle text in many languages without needing separate models for each.

2. Step 2: Compare options

   It can analyze sentiment in multiple languages with one model. correctly states the advantage. Options B, C, and D are incorrect because they limit the model to one language or misunderstand its function.

3. Final Answer:

   It can analyze sentiment in multiple languages with one model. -> Option A

4. Quick Check:

   Multilingual model = multiple languages [OK]

Hint: Multilingual means many languages, not just one [OK]

Common Mistakes:

• Thinking it only works for English
• Believing you need separate models per language
• Assuming language is ignored

2. Which of the following is the correct way to load a pretrained multilingual sentiment model using Hugging Face Transformers in Python?

easy

A. model = AutoModelForSequenceClassification.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')

B. model = AutoTokenizer.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')

C. model = AutoConfig.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')

D. model = AutoModel.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')

Solution

1. Step 1: Identify the correct class for sentiment classification

   For sentiment tasks, use AutoModelForSequenceClassification to load the model with classification head.

2. Step 2: Review options

   model = AutoModelForSequenceClassification.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment') uses AutoModelForSequenceClassification correctly. model = AutoModel.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment') loads a base model without classification head. model = AutoTokenizer.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment') loads tokenizer, not model. model = AutoConfig.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment') loads config only.

3. Final Answer:

   model = AutoModelForSequenceClassification.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment') -> Option A

4. Quick Check:

   SequenceClassification = sentiment model [OK]

Hint: Use AutoModelForSequenceClassification for sentiment tasks [OK]

Common Mistakes:

• Using AutoModel without classification head
• Confusing tokenizer with model
• Loading only config without weights

3. Given the following Python code snippet using the 'nlptown/bert-base-multilingual-uncased-sentiment' model, what will be the output sentiment label for the input text "Je suis très content" (French for "I am very happy")?

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')
model = AutoModelForSequenceClassification.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')

inputs = tokenizer("Je suis très content", return_tensors="pt")
outputs = model(**inputs)
probs = torch.nn.functional.softmax(outputs.logits, dim=1)
label = torch.argmax(probs).item() + 1  # labels 1 to 5
print(label)

medium

A. 1 (Very Negative)

B. 5 (Very Positive)

C. 3 (Neutral)

D. 2 (Negative)

Solution

1. Step 1: Understand the input sentiment

   The French sentence "Je suis très content" means "I am very happy", which is a positive sentiment.

2. Step 2: Interpret model output labels

   The model outputs labels from 1 (very negative) to 5 (very positive). Since the sentence is very positive, the highest probability label should be 5.

3. Final Answer:

   5 (Very Positive) -> Option B

4. Quick Check:

   Positive sentence = label 5 [OK]

Hint: Happy words usually map to highest positive label [OK]

Common Mistakes:

• Confusing label numbers with sentiment polarity
• Ignoring language and assuming English only
• Not adding 1 to zero-based index

4. You run this code to analyze sentiment but get an error:

from transformers import AutoTokenizer, AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')
tokenizer = AutoTokenizer.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')

inputs = tokenizer('Das ist schlecht', return_tensors='pt')
outputs = model(inputs)

What is the cause of the error?

medium

A. Missing import for torch library.

B. Tokenizer is loaded after the model, causing mismatch.

C. The input text is in German, which the model cannot process.

D. Model expects keyword arguments, but inputs passed as positional argument.

Solution

1. Step 1: Check how model is called

   The model expects inputs as keyword arguments like model(**inputs), but here inputs are passed as a single positional argument.

2. Step 2: Analyze other options

   Tokenizer order does not cause error. The model supports German. Missing torch import would cause a different error.

3. Final Answer:

   Model expects keyword arguments, but inputs passed as positional argument. -> Option D

4. Quick Check:

   Use model(**inputs) not model(inputs) [OK]

Hint: Pass inputs with ** to model call [OK]

Common Mistakes:

• Passing inputs without unpacking as keyword args
• Blaming language support incorrectly
• Ignoring error message details

5. You want to build a multilingual sentiment analysis app that supports English, Spanish, and Chinese. Which approach best balances accuracy and simplicity?

hard

A. Train separate sentiment models for each language from scratch.

B. Translate all texts to English and use an English-only sentiment model.

C. Use a pretrained multilingual sentiment model like 'nlptown/bert-base-multilingual-uncased-sentiment'.

D. Use a simple keyword-based sentiment dictionary for each language.

Solution

1. Step 1: Evaluate training effort and coverage

   Training separate models is costly and complex. Keyword-based methods lack accuracy. Translating text adds errors and latency.

2. Step 2: Consider pretrained multilingual models

   Pretrained multilingual models support many languages with good accuracy and easy setup, balancing simplicity and performance.

3. Final Answer:

   Use a pretrained multilingual sentiment model like 'nlptown/bert-base-multilingual-uncased-sentiment'. -> Option C

4. Quick Check:

   Pretrained multilingual = best balance [OK]

Hint: Pretrained multilingual models save time and support many languages [OK]

Common Mistakes:

• Assuming training separate models is easier
• Ignoring translation errors
• Overestimating keyword-based method accuracy