RAI 시험 - GARP실제시험문제와 답 - 300문항

Question No : 1

A bank wants to analyze customer reviews, where certain words like "rate" may have different meanings depending on the context.
Which feature of the transformer model would be most beneficial in distinguishing between the different meanings?

A.Sequence-to-sequence structure
B.Attention mechanism
C.Static embeddings
D.Recurrent connections

정답:
Explanation:
B is correct. The attention mechanism allows transformers to assign contextualized embeddings, differentiating words based on their context within the sentence
A is incorrect as the sequence-to- sequence structure refers to the model’s encoder-decoder format
C is incorrect as static embeddings do not adjust for context
D is incorrect as transformers lack recurrent connections, which are instead present in RNNs.

Question No : 2

A bank wants to analyze long text sequences to understand customer sentiment over time. However, their RNN model faces difficulty in learning from words far apart in sentences.
What model modification could help with long-range dependencies?

A.Increase the training epochs
B.Switch to a Convolutional Neural Network
C.Use a Long Short-Term Memory (LSTM) network
D.Reduce the sequence length

정답:
Explanation:
C is correct. LSTM networks are specifically designed to handle long-range dependencies by using gates to control information flow
A is incorrect as simply increasing epochs does not resolve long-range dependency issues
B is incorrect as CNNs are unsuitable for sequential text data
D is incorrect as reducing sequence length can decrease context understanding.

Question No : 3

A multimedia company wants to develop an AI model that can generate both text and images based on user prompts.
Which type of model is most appropriate for this task?

A.Multimodal model
B.Text-only model
C.Audio-only model
D.Transformer model limited to NLP

정답:
Explanation:
A is correct. Multimodal models are designed to generate multiple types of content, such as both text and images, based on a single prompt.
B and C are incorrect because they are limited to single content types
D is incorrect as transformers designed only for NLP cannot generate images.

Question No : 4

A bank analyst wants to classify a new feedback document using Naïve Bayes. The document contains words previously marked as negative.
Which step should the analyst prioritize to classify this document?

A.Calculate the posterior probability for each class based on word occurrences.
B.Remove commonly used words that appear in every document.
C.Set equal probability for all classes before calculating.
D.Use only the prior probability of each class to classify.

정답:
Explanation:
A is correct. Naïve Bayes calculates the posterior probability for each class using word occurrences and assigns the document to the class with the highest probability.
B is irrelevant to classification
C is incorrect as probabilities differ based on words.
D ignores the role of individual word probabilities.

Question No : 5

A bank is analyzing customer feedback to classify it as "Good," "Bad," or "Indifferent" using a Naïve Bayes classifier.
What is a major assumption made by the Naïve Bayes approach in this context?

A.Each feedback word is dependent on others.
B.Words are equally important for classification.
C.Words occur independently of each other.
D.Words do not contribute to the class label.

정답:
Explanation:
C is correct. Naïve Bayes assumes independence among features (words), meaning each word is treated independently
A is incorrect as Naïve Bayes assumes independence
B is incorrect; importance varies based on probabilities
D is incorrect because words are essential for determining class labels.

Question No : 6

A financial analyst is building an NLP model to analyze customer feedback on loan services. She notices many comments use negation, such as “not helpful” and “not satisfied.”
To improve sentiment accuracy, what technique should she consider?

A.Stop word removal
B.Part-of-speech tagging
C.Stemming
D.Using n-grams

정답:
Explanation:
D is correct. Using n-grams, especially bi-grams, helps capture phrases with negation like “not helpful.”
A is incorrect, as removing stop words could discard important negation terms.
B is unrelated to capturing multi-word phrases.
C is about reducing words to their root, not handling negations.

Question No : 7

An investment firm’s sentiment analysis of earnings call transcripts is skewed by excessive repetition of words like “growth” and “profit.”
To ensure no single word overpowers the sentiment vector, what approach should the firm take?

A.Remove high-frequency words
B.Apply L2 normalization to the vectors
C.Use stemming to reduce repeated words
D.Increase the threshold for sentiment words

정답:
Explanation:
B is correct. L2 normalization controls the impact of frequent words, balancing each word’s influence
A is incorrect as it would remove valuable context.
C does not reduce frequency impact but consolidates similar words.
D is irrelevant to managing word repetition.

Question No : 8

A bank is analyzing customer feedback on its services and encounters a review that heavily repeats the word “bad,” causing this term to dominate the analysis.
Which pre-processing step can help mitigate the impact of this repeated word in the analysis?

A.Removing the repeated words manually
B.Using L2 normalization on the word vectors
C.Increasing the vocabulary size
D.Using binary bag-of-words representation

정답:
Explanation:
B is correct. L2 normalization scales the vector, reducing the impact of any single term, even if repeated
A is incorrect as manual removal is inefficient for large datasets.
C does not address the dominance of frequent words.
D would only indicate presence or absence, not reduce the weight of repeated terms.

Question No : 9

In preparing text from central bank communications for analysis, a researcher encounters phrases like “interest rates might increase.”
Which preprocessing step can help convert these words to their base forms, making them easier to analyze?

A.Stop word removal
B.Lemmatization
C.Tokenization
D.Feature extraction

정답:
Explanation:
B is correct. Lemmatization reduces words to their base form, making “increase” and “increasing” identical in analysis
A is incorrect as stop word removal removes non-informative words without transforming base forms
C is incorrect because tokenization splits text into components but does not simplify word forms
D is incorrect as feature extraction converts text into numerical formats post- processing.

Question No : 10

A data scientist at a hedge fund uses grid search to optimize hyperparameters for a trading model. Concerned about computational efficiency, they consider switching to a random search.
What is a key advantage of using random search over grid search in this context?

A.It guarantees finding the exact optimal hyperparameter value
B.It always outperforms grid search
C.It reduces computation time by sampling fewer hyperparameter values
D.It ensures all possible values are tested

정답:
Explanation:
C is correct. Random search is generally faster as it samples fewer points and avoids exhaustive grid testing
A is incorrect; random search does not guarantee the exact optimal value.
B is incorrect, as random search may not outperform grid search.
D is incorrect, as random search does not test all possible values.

Question No : 11

A financial analyst wants to apply a regularized regression model that reduces extreme coefficient values without eliminating any features, as they all provide valuable insights.
Which regularization method is most appropriate?

A.LASSO
B.Ridge
C.Elastic Net
D.Stepwise selection

정답:
Explanation:
B is correct. Ridge regression (L2 regularization) reduces coefficient magnitudes without setting them to zero, making it ideal when all features should be retained
A is incorrect as LASSO can set coefficients to zero, potentially eliminating features.
C is partially correct as Elastic Net may also zero out some coefficients if L1 is significant
D is incorrect as stepwise selection does not perform regularization.

Question No : 12

A bank decides to use a simple linear regression model to understand the impact of economic variables on credit risk. They aim for a clear understanding of causal relationships.
What is likely sacrificed in this choice?

A.Prediction accuracy
B.Interpretability
C.Model simplicity
D.Bias in predictions

정답:
Explanation:
A is correct because opting for a simpler, interpretable model may sacrifice prediction accuracy compared to complex models
B is incorrect as interpretability is not sacrificed in this case; it’s actually prioritized
C is incorrect since simplicity is maintained, not sacrificed
D is incorrect as bias may still exist, but it is not directly due to interpretability.

Question No : 13

A financial institution uses a neural network model with thousands of parameters to predict loan defaults. On the training dataset, the model has a nearly zero residual sum of squares (RSS). However, it performs poorly on new data.
What does this indicate?

A.The model has insufficient parameters
B.The model is underfitted
C.The model is overfitted
D.The model has perfect generalization

정답:
Explanation:
C is correct because poor performance on new data with excellent performance on the training set indicates overfitting, where the model has memorized the training data rather than generalizing
A is incorrect as the model likely has too many parameters, not too few
B is incorrect as underfitting typically results in poor performance on both training and validation sets
D is incorrect because perfect generalization would imply good performance on both training and new data.

Question No : 14

A financial institution is implementing a neural network to predict loan default risk. They use gradient descent with backpropagation.
What is the primary purpose of applying the chain rule in this process?

A.To calculate the initial weights for each neuron.
B.To calculate the partial derivatives needed to update weights in each layer.
C.To ensure the model does not overfit the data.
D.To validate the accuracy of the final model predictions.

정답:
Explanation:
B is correct because the chain rule is used to compute the partial derivatives of the loss function with respect to each weight, enabling accurate updates
A is incorrect as initial weights are typically randomized, not derived from the chain rule
C is incorrect as the chain rule itself does not prevent overfitting
D is incorrect because the chain rule does not validate predictions but helps calculate gradients for weight updates.

Question No : 15

During training, a credit scoring model using stochastic gradient descent shows fluctuating performance and fails to converge smoothly.
Which of the following is the most likely cause?

A.The learning rate is too high, causing oscillations.
B.The model is overfitting to the training data.
C.The model is not receiving enough data.
D.Batch gradient descent is more suitable due to consistent updates.

정답:
Explanation:
A is correct. A high learning rate can cause oscillations in SGD, leading to unstable convergence
B is incorrect as overfitting is unrelated to convergence issues
C is incorrect because SGD can function well with smaller data batches.
D is partially correct but does not explain the root cause of the fluctuating convergence.

GARP RAI 시험