Sentiment analysis pipeline in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What is the main purpose of tokenization in a sentiment analysis pipeline?

In a sentiment analysis pipeline, why do we perform tokenization on the input text?

A To train the model on labeled data

B To convert sentiment scores into numerical values

C To split the text into smaller units like words or subwords for easier processing

D To remove stop words from the text

💡 Hint

Think about how a computer understands text before analyzing sentiment.

Tokenization breaks down text into smaller pieces such as words or subwords, which makes it easier for the model to process and understand the text.

❓ Predict Output

intermediate

What is the output of this preprocessing code snippet?

Given the following Python code for preprocessing text in a sentiment analysis pipeline, what is the output?

NLP

import re
text = "I love this product! It's amazing."
cleaned = re.sub(r'[^a-zA-Z ]', '', text).lower().split()
print(cleaned)

A ['I', 'love', 'this', 'product', 'its', 'amazing']

B ['I', 'love', 'this', 'product', 'It', 's', 'amazing']

C ['i', 'love', 'this', 'product', 'it', 's', 'amazing']

D ['i', 'love', 'this', 'product', 'its', 'amazing']

💡 Hint

Look at how punctuation is removed and text is converted to lowercase.

The regex removes all characters except letters and spaces. Then the text is converted to lowercase and split into words, so "It's" becomes "its" without the apostrophe.

❓ Model Choice

advanced

Which model architecture is best suited for capturing context in sentiment analysis?

You want to build a sentiment analysis model that understands the context of words in a sentence. Which model architecture is most suitable?

A A recurrent neural network (RNN) like LSTM or GRU

B A linear regression model on raw text

C A simple bag-of-words model with logistic regression

D A k-nearest neighbors model using word counts

💡 Hint

Think about models that can remember previous words to understand meaning.

RNNs like LSTM or GRU can capture the order and context of words, which is important for understanding sentiment in sentences.

❓ Metrics

advanced

Which metric is most appropriate to evaluate a sentiment analysis model on imbalanced data?

Your sentiment dataset has many more positive reviews than negative ones. Which evaluation metric should you prioritize?

A F1-score

B Recall

C Accuracy

D Precision

💡 Hint

Consider a metric that balances precision and recall.

F1-score balances precision and recall, making it better for imbalanced datasets where accuracy can be misleading.

🔧 Debug

expert

Why does this sentiment analysis model always predict the same class?

Here is a snippet of a sentiment analysis model training code. The model always predicts the same sentiment class regardless of input. What is the most likely cause?

NLP

import torch
import torch.nn as nn

class SimpleSentimentModel(nn.Module):
    def __init__(self, vocab_size, embed_dim, num_classes):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, embed_dim)
        self.fc = nn.Linear(embed_dim, num_classes)

    def forward(self, x):
        embedded = self.embedding(x)  # shape: (batch_size, seq_len, embed_dim)
        pooled = embedded.mean(dim=1)  # average over seq_len
        out = self.fc(pooled)
        return out

model = SimpleSentimentModel(vocab_size=1000, embed_dim=50, num_classes=2)

# Training loop omitted for brevity

# After training, model always predicts class 0.

A The model's output layer lacks a softmax or sigmoid activation, so predictions are not probabilities

B The training data is imbalanced and the model learned to predict the majority class

C The embedding layer is not updating because requires_grad is False

D The pooling operation averages embeddings, losing important sequence information

💡 Hint

Consider the effect of unbalanced classes on model predictions.

If the training data is heavily imbalanced, the model may learn to always predict the majority class to minimize loss, resulting in poor generalization.