AIF-C01（日语） 练习题 — AIF-C01:AI 从业者（日语）

The question requires content that provides transparency and explainability of ML forecast models to business stakeholders, who are typically non-technical and need to understand what drives model predictions to trust the forecasts and make informed operational decisions. Partial dependence plots (PDPs) are a standard explainable AI (XAI) tool designed to visualize the marginal impact of individual input features on model output, while holding all other features constant. This allows stakeholders to easily interpret how factors such as historical demand trends, seasonality, or operational inputs influence the quarterly demand forecasts, directly meeting the transparency and explainability requirements outlined. This aligns with the AIF-C01 domain focus on responsible AI, explainability, and non-technical stakeholder communication for ML use cases. Option Analysis: A. Code for model training is incorrect. Model training code is a technical artifact intended for ML engineering teams to debug or replicate model training, but it is not interpretable for non-technical business stakeholders, does not explain what drives the model's forecast outputs, and may contain sensitive intellectual property, so it does not meet the stated requirements. B. Partial dependence plots (PDPs) is correct. As a core XAI visualization, PDPs clearly demonstrate how relevant input features affect the model's predicted demand values, providing actionable, easy-to-understand insights for stakeholders to understand what factors drive forecast results. This directly addresses the requirements for transparency and explainability for stakeholders who need to trust and act on the forecast outputs to optimize operations. C. Sample data for training is incorrect. Raw sample training data does not explain how the model uses input data to generate forecasts. It also does not provide insights into model decision logic, may include sensitive customer or business data, and is not useful for non-technical stakeholders seeking to understand forecast drivers. D. Model convergence tables are incorrect. Convergence tables are technical artifacts that confirm the model training process completed successfully and stabilized during training, but they do not explain what features drive the model's final forecast outputs, so they do not meet the explainability requirement for business stakeholders. Key Concepts: 1. Explainable AI (XAI) for Stakeholder Communication: This AIF-C01 core concept covers translating complex ML model logic into interpretable insights accessible to non-technical users, focusing on tools that demonstrate prediction drivers rather than technical training artifacts. 2. ML Transparency and Governance: This core domain covers requirements for building stakeholder trust in ML outputs by providing clear, relevant explanations of model behavior, rather than artifacts only useful for technical ML teams, aligned with responsible AI practices. 3. Partial Dependence Plots (PDPs): This standard XAI tool is explicitly included in the AIF-C01 explainability topic as a method to visualize the marginal effect of features on model predictions, enabling stakeholders to understand model behavior. References: AWS What is Explainable AI (XAI)?, AWS Well-Architected Framework Machine Learning Lens: Explainability, https://docs.aws.amazon.com/wellarchitected/latest/machine-learning-lens/explainability.html

The question requires an LLM-powered solution that extracts key points from long legal documents, a use case that falls under the AIF-C01 domain of generative AI task identification and use case alignment. The suggested answer, a summarization chatbot, directly addresses this requirement because text summarization is a core LLM capability designed to condense long-form unstructured text into concise outputs that retain all critical, relevant information. For legal use cases, summarization models can be fine-tuned on legal domain corpora to accurately identify and extract key points such as case holdings, contractual obligations, filing deadlines, and argument highlights without extraneous detail, perfectly matching the law firm's needs. Option Analysis: A. Incorrect. Named Entity Recognition (NER) is a natural language processing task that only identifies and categorizes predefined specific entities such as names, dates, organization names, or case numbers from text. It cannot extract broader, context-dependent key points or summarize the overall meaning of a legal document, so it is too limited to meet the stated requirement. This task is covered in the AIF-C01 domain of traditional NLP capabilities, where candidates learn to distinguish NER from more complex generative AI tasks. B. Incorrect. Recommendation engines are AI systems that suggest relevant content, products, or actions to users based on historical behavior, preference data, or similarity matching. They have no functionality related to extracting key points from existing documents, so this solution is unrelated to the law firm's use case. Recommendation engines are covered in AIF-C01 as a separate AI use case category unrelated to text processing for key point extraction. C. Correct. Text summarization is a core generative AI task explicitly covered in the AIF-C01 exam objectives, where LLMs process long input text to generate a condensed output that preserves all critical information. A summarization chatbot built for legal use cases can process entire legal documents, identify contextually relevant key points, and output them in a structured, easy-to-consume format, which directly meets the requirement to extract key points from legal documents. D. Incorrect. Multi-language translation systems convert text from one source language to a target language, and do not include functionality to summarize or extract key points from input text. This solution is unrelated to the stated requirement, and translation is covered as a separate NLP task in the AIF-C01 exam domains. Key Concepts: 1. LLM Task Classification: A core AIF-C01 knowledge domain that requires candidates to identify appropriate LLM tasks for specific business requirements, including summarization, NER, translation, and content generation. Summarization is the correct task for extracting key points from long documents. 2. AI Use Case Alignment: A foundational AIF-C01 competency that involves matching business needs to the correct AI solution, avoiding unrelated or insufficiently capable AI systems that do not address the stated requirement. 3. Unstructured Text Processing: A key AIF-C01 concept covering AI solutions for processing unstructured data such as legal documents, which have no predefined structured format, to extract actionable insights or condensed information. References: AWS Certified AI Practitioner (AIF-C01) Official Exam Guide, https://aws.amazon.com/certification/certified-ai-practitioner/ AWS Generative AI Use Cases: Summarization

The scenario presents two core requirements: first, a multi-class classification task to categorize genes into 20 discrete categories, and second, full visibility into the model's inner decision-making process to document how internal mechanisms impact output. The correct algorithm must satisfy both requirements. Decision trees are the only option that natively supports multi-class classification while offering complete, traceable interpretability, making them the appropriate choice for this use case. This aligns with AIF-C01 domain content covering algorithm selection based on use case requirements and explainable AI principles. Option Analysis: A. Decision trees: Correct. Decision trees natively support multi-class classification, so they can directly handle the 20 gene category task without modification. They are inherently interpretable, as every prediction can be traced through a clear path of feature-based split decisions from the root node to the final classification leaf node. This allows the company to explicitly document how each gene characteristic (feature) and corresponding threshold contributes to the output, fully meeting the scenario's requirements. B. Linear regression: Incorrect. Linear regression is a supervised learning algorithm designed for regression tasks that predict continuous numerical values, not discrete classification tasks. It cannot be used to assign genes to 20 distinct categories, so it fails the core use case requirement. C. Logistic regression: Incorrect. While logistic regression supports classification, it is natively a binary classification, and multi-class adaptations introduce additional complexity that reduces interpretability for 20 categories. Logistic regression provides feature importance via coefficients, but it does not offer the step-by-step, traceable decision process the company needs to document the model's inner mechanism impact on output. D. Neural networks: Incorrect. Neural networks can perform high-accuracy multi-class classification, but they are considered black box models. The interactions between input features and hidden layer weights are not easily interpretable, and it is extremely difficult to trace how specific gene characteristics drive the final classification. This violates the requirement to document the model's inner mechanism, so it is not a valid choice. Key Concepts: 1. Explainable AI (XAI): A core AIF-C01 domain concept referring to the ability to explain and validate how an ML model generates predictions. For high-stakes use cases like genomic classification, XAI is a critical requirement to ensure model decisions are transparent and auditable, which is the primary need in this scenario. 2. Multi-class Classification: A supervised learning task where the model assigns inputs to one of three or more discrete classes. This question's 20 gene categories define a multi-class task, so algorithm selection must include support for this task type. 3. Inherent Algorithm Interpretability: A key AIF-C01 knowledge point that different ML algorithms have varying levels of built-in interpretability. Decision trees are among the most interpretable classification algorithms, while complex models like deep neural networks have very low inherent interpretability. References: AWS Certified AI Practitioner (AIF-C01) Exam Guide, Amazon SageMaker Developer Guide: Machine Learning Interpretability, https://docs.aws.amazon.com/sagemaker/latest/dg/interpretability.html

The question describes a discrete image classification task where the company needs to measure the total volume of correctly classified plant disease images. For the AWS Certified AI Practitioner AIF-C01 domain, candidates are required to match appropriate evaluation metrics to machine learning task types. This scenario requires a classification-specific performance metric that directly quantifies the share of correct predictions. Accuracy is the standard metric for this use case, as it calculates the ratio of correctly classified samples to the total number of samples evaluated, which directly addresses the company's stated requirement. Option Analysis: A. R-squared score: Incorrect. R-squared is a regression evaluation metric that quantifies the proportion of variance in a continuous target variable explained by a regression model. It is not applicable to discrete classification tasks with categorical output labels, so it cannot be used to count correct classification predictions. B. Accuracy: Correct. Accuracy is a core classification performance metric defined as the number of correct predictions divided by the total number of predictions made. It directly measures the share of images the model classified correctly, which aligns exactly with the company's requirement for this image classification use case. C. Root mean squared error (RMSE): Incorrect. RMSE is a regression metric that calculates the average magnitude of error between predicted continuous values and actual continuous target values. It is designed for tasks with numerical outputs, not discrete class labels, so it is not relevant for counting correct classification outcomes. D. Learning rate: Incorrect. Learning rate is a training hyperparameter, not an evaluation metric. It controls the size of weight updates applied to a model during training to minimize loss, and does not measure the performance of a trained model on inference tasks. Key Concepts: 1. Classification vs Regression Evaluation Metrics: A core AIF-C01 domain topic that distinguishes metrics for discrete label prediction tasks (classification) from metrics for continuous value prediction tasks (regression). Classification uses metrics like accuracy, precision, and recall, while regression uses metrics like R-squared, RMSE, and MAE. 2. Accuracy Metric: Defined as the ratio of correct predictions to total predictions, this is a fundamental classification metric for measuring overall correct prediction rate, which is the primary requirement in the given scenario. 3. Performance Metrics vs Training Hyperparameters: A key AIF-C01 concept that differentiates metrics that measure trained model performance on inference data from hyperparameters that are set prior to training to control model training behavior, such as learning rate, batch size, and epoch count. References: Amazon SageMaker Developer Guide: Model Quality Metrics for Classification Problems, https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor-model-quality-metrics.html#model-monitor-model-quality-classification-metrics AWS Certified AI Practitioner (AIF-C01) Official Exam Guide, https://aws.amazon.com/certification/certified-ai-practitioner/

This question aligns with the Generative AI domain of the AWS Certified AI Practitioner AIF-C01 exam, which tests knowledge of methods to align pre-trained large language model outputs to business use case requirements. The company has two explicit, rule-based requirements for LLM outputs: short length, and use of a specific target language. Prompt adjustment is the lowest effort, most cost-effective solution for these requirements, as it communicates the required output constraints directly to the pre-trained LLM at inference time without requiring model modification, retraining, or changes to inference parameters that impact output randomness rather than structural and linguistic constraints. Option Analysis: A. Correct. Prompt engineering, which involves adjusting input prompts to include explicit instructions for the LLM, is a standard method to enforce output requirements like length (e.g., including the instruction "respond in 2 sentences or less") and target language (e.g., including the instruction "write all outputs in Brazilian Portuguese"). This solution directly addresses the company's stated requirements without changes to the underlying model or inference sampling settings. B. Incorrect. Adjusting LLM size impacts overall model capability, complexity of tasks it can perform, and inference cost, but does not inherently enforce short responses or specific output language. Both smaller and larger LLMs rely on input prompts to understand required output formatting and language, so changing model size does not resolve the stated requirements. C. Incorrect. Temperature is an inference parameter that controls the randomness of LLM outputs. Increasing temperature makes outputs more creative and variable, but has no impact on enforcing response brevity or specific output language. This adjustment would make outputs less predictable, not more aligned to the company's requirements. D. Incorrect. Top K is a nucleus sampling inference parameter that limits the number of top candidate tokens the LLM considers when generating each output token, to control output diversity. Adjusting Top K does not enforce requirements for response length or target language, so it does not address the company's use case. Key Concepts: 1. Prompt Engineering: A core AIF-C01 knowledge area referring to the practice of designing input prompts to elicit desired, aligned outputs from generative AI models. Explicit instructions in prompts can control output formatting, length, tone, and language without modifying the underlying model. 2. LLM Inference Parameters: Configurable settings for LLM output generation including temperature, Top P, and Top K, which adjust randomness and diversity of outputs but do not enforce explicit rule-based requirements for output structure or language. 3. Pre-trained LLM Alignment: The process of adjusting LLM behavior to match business requirements, where low-complexity alignment needs (like output length and language) are addressed via prompt engineering, while more complex alignment needs require fine-tuning or reinforcement learning from human feedback. References: AWS Machine Learning Blog: Prompt engineering guidelines for generative AI on AWS, AWS Certified AI Practitioner (AIF-C01) Exam Guide