Limitations of classical methods in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why classical ML methods struggle with high-dimensional text data

Classical machine learning methods like Naive Bayes or SVM often face challenges when working with text data represented by thousands of features (words). What is the main reason for this difficulty?

A They do not support categorical features like words

B They cannot handle numeric data and text is always numeric

C They always produce biased predictions regardless of data size

D They require very large amounts of labeled data to avoid overfitting in high dimensions

💡 Hint

Think about what happens when the number of features is much larger than the number of training examples.

Classical methods often overfit when the feature space is very large compared to the number of samples. This is common in text data where each word is a feature, leading to thousands of features but limited labeled examples.

❓ Model Choice

intermediate

Choosing a model to handle complex language patterns

Which classical machine learning model is least suitable for capturing complex word order and context in sentences?

A Logistic Regression with n-gram features

B Naive Bayes with bag-of-words features

C Recurrent Neural Network (RNN)

D Support Vector Machine with TF-IDF features

💡 Hint

Consider which model assumes independence between words and ignores order.

Naive Bayes assumes word independence and uses bag-of-words, so it cannot capture word order or context well. Other models like RNNs or n-gram features consider some order or context.

❓ Metrics

advanced

Evaluating classical methods on imbalanced text data

You train a classical classifier on a text dataset where 95% of examples belong to one class. The model achieves 95% accuracy but poor recall on the minority class. What metric better reflects the model's weakness?

A Recall on the minority class

B Overall accuracy

C Precision on the minority class

D Training loss

💡 Hint

Which metric measures how many true positive minority examples are correctly found?

Recall on the minority class measures the fraction of actual minority examples correctly identified. High accuracy can be misleading if the model ignores the minority class.

🔧 Debug

advanced

Why does this classical text classifier fail to generalize?

Consider this Python snippet using a classical method for text classification:

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.naive_bayes import MultinomialNB

texts = ["good movie", "bad movie", "great film", "terrible film"]
labels = [1, 0, 1, 0]

vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)

model = MultinomialNB()
model.fit(X, labels)

new_text = ["good film"]
X_new = vectorizer.transform(new_text)
pred = model.predict(X_new)
print(pred)

Why might this model give unreliable predictions on new text?

NLP

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.naive_bayes import MultinomialNB

texts = ["good movie", "bad movie", "great film", "terrible film"]
labels = [1, 0, 1, 0]

vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)

model = MultinomialNB()
model.fit(X, labels)

new_text = ["good film"]
X_new = vectorizer.transform(new_text)
pred = model.predict(X_new)
print(pred)

A The training data is too small and limited to learn reliable word associations

B CountVectorizer does not convert text to numbers

C MultinomialNB cannot handle sparse matrices

D The new text contains words not seen during training

💡 Hint

Think about the size and variety of the training examples.

The training set is very small with only four examples, so the model cannot learn strong patterns. This causes unreliable predictions on new inputs even if the words are known.

🧠 Conceptual

expert

Fundamental limitation of classical methods in capturing semantics

Classical machine learning methods for NLP often rely on word frequency features like bag-of-words or TF-IDF. What is a fundamental limitation of these features in understanding language?

A They require deep neural networks to compute

B They always produce dense vectors that are hard to interpret

C They ignore word order and context, so cannot capture meaning beyond individual words

D They cannot be used with linear classifiers

💡 Hint

Think about what meaning depends on besides word counts.

Bag-of-words and TF-IDF count words but ignore their order and surrounding words, so they miss context and semantic relationships essential for true language understanding.