用語集
各項目の用語表をまとめた一覧です。用語をクリックすると解説を表示します。
1.1 AI の基本概念と用語
| 日本語 | 読み | English |
|---|
| | Agentic AI |
| | Algorithm |
| （AI） | じんこうちのう | Artificial Intelligence |
| ひどうきすいろん | Asynchronous inference |
| じりつせい | Autonomy |
| | Batch inference |
| | Bias |
| | Classic / traditional ML |
| ぶんるい | Classification |
| | Clustering |
| | Cold start |
| | Computer Vision |
| しんそうがくしゅう | Deep Learning |
| | Discriminative |
| | Endpoint |
| こうへいせい | Fairness |
| | Fit |
| | Generative |
| せいせいエーアイ | Generative AI |
| | Image data |
| すいろん | Inference |
| | Labeled data |
| （LLM） | | LLM |
| （ML） | きかいがくしゅう | Machine Learning |
| | Model |
| | Neural Network |
| しぜんげんごしょり | NLP |
| | Real-time inference |
| かいき | Regression |
| きょうかがくしゅう | Reinforcement learning |
| ほうしゅう | Reward |
| | Self-supervised |
| | Semi- |
| | Serverless inference |
| | Structured |
| きょうしありがくしゅう | Supervised learning |
| | Tabular data |
| | Text data |
| じけいれつデータ | Time series data |
| | Training |
| | Unlabeled data |
| | Unstructured |
| きょうしなしがくしゅう | Unsupervised learning |
1.2 AI のユースケース
| 日本語 | 読み | English |
|---|
| | Amazon Comprehend |
| | Amazon Lex |
| | Amazon Personalize |
| | Amazon Polly |
| | Amazon Rekognition |
| | Amazon SageMaker AI |
| | Amazon Textract |
| | Amazon Transcribe |
| | Amazon Translate |
| | Automation |
| | Binary |
| ひようたいこうか | Cost-benefit analysis |
| いしけっていのしえん | Decision support |
| とくてい | Deterministic outcome |
| せつめいかのうせい | Explainability |
| きばんモデル | Foundation model (FM) |
| ふせいけんしゅつ | Fraud detection |
| | Human-in-the-loop |
| | Insufficient data |
| | Knowledge base |
| | Label |
| | Multiclass |
| うんようじょうのせいやく | Operational constraints |
| | Personalization |
| よそく | Prediction |
| | Recommendation system |
| きせい | Regulation |
| | Rule-based system |
| | Scalability |
| おんせいにんしき | Speech recognition |
| | Traditional ML model |
1.3 AI/ML の開発ライフサイクル
| 日本語 | 読み | English |
|---|
| せいかいりつ | Accuracy |
| | Amazon Bedrock |
| | Amazon Quick |
| | Cost per user |
| | Custom-trained model |
| | Customer feedback |
| しゅうしゅう | Data collection |
| まえしょり | Data preprocessing |
| | Deployment |
| ひょうか | Evaluation |
| じっけん | Experimentation |
| （EDA） | たんさくてきデータぶんせき | Exploratory Data Analysis |
| | F1 score |
| | Feature engineering |
| | Hyperparameter tuning |
| | Kiro |
| | License |
| | Managed API service |
| （MLOps） | えむえるうんよう | MLOps |
| | Model monitoring |
| | Model training |
| | Monitoring |
| | Open-source pretrained model |
| うんようふか | Operational overhead |
| | Pre-training |
| てきごうりつ | Precision |
| ほんばんかどう | Production readiness |
| | Proprietary model |
| さいげんりつ | Recall |
| | Repeatable process |
| | Retraining |
| （ROI） | とうししゅうえきりつ | Return on Investment |
| | SageMaker JumpStart |
| | Scalable system |
| | Self-hosted API |
| ぎじゅつてきふさい | Technical debt |
2.1 生成 AI の基本概念
| 日本語 | 読み | English |
|---|
| （A2A） | | Agent-to-Agent (A2A) |
| | AI assistant |
| | Chunking |
| | Code generation |
| ようやく・あっしゅく | Context compaction |
| | Context engineering |
| | Context window |
| | Customer service agent |
| せんたく | Data selection |
| かくさんモデル | Diffusion model |
| うめこみ | Embedding |
| | Feedback |
| | Fine-tuning |
| | Graph |
| | Image / video / audio generation |
| | Input tokens |
| | instructions |
| | Memory |
| | Memory management |
| （MCP） | | Model Context Protocol |
| | Model selection |
| | Multi-agent system |
| | Multimodal model |
| | On-demand |
| | Output tokens |
| | Prompt caching |
| | Prompt engineering |
| | Provisioned throughput |
| | Recommendation |
| | results |
| | Search |
| ようやく | Summarization |
| | Supervisor pattern |
| | Swarm |
| | System prompt |
| | Token |
| | Token-based pricing |
| | Tool descriptions |
| | Tool use |
| （LLM） | | Transformer-based LLM |
| ほんやく | Translation |
| | Vector |
| | Workflow |
| | Workflow orchestration |
2.2 生成 AI の可能性と限界
| 日本語 | 読み | English |
|---|
| てきおうせい | Adaptability |
| | Average revenue per user (ARPU) |
| きのう | Capabilities |
| | Compliance |
| せいやく | Constraints |
| | Content generation |
| かいわのうりょく | Conversational ability |
| | Conversion rate |
| | Cost |
| | Cross-domain performance |
| （CLV） | こきゃくしょうがいかち | Customer lifetime value |
| こうりつせい | Efficiency |
| | Hallucination |
| | Human review |
| ふせいかく | Inaccuracy |
| かいしゃくかのうせい | Interpretability |
| | Latency |
| | Model complexity |
| | Model type |
| ひけっていせい | Non-determinism |
| | Performance requirements |
| おうとうせい | Responsiveness |
| | Simplicity |
2.3 生成 AI を支える AWS のサービス
| 日本語 | 読み | English |
|---|
| | Accessibility |
| | Amazon Bedrock AgentCore |
| | Amazon Nova |
| かようせい | Availability |
| ひようたいこうか | Cost-effectiveness |
| | Cross-Region inference |
| あんごうか | Encryption |
| さんにゅうしょうへき | Low barrier to entry |
| | Meeting business goals |
| | Performance |
| じょうちょうせい | Redundancy |
| てんかい | Regional coverage |
| せきにん | Responsibility |
| あんぜんせい | Safety |
| | Security |
| せきにんきょうゆうモデル | Shared responsibility model |
| | Speed to market |
| | Strands Agents |
3.1 基盤モデル活用の設計
| 日本語 | 読み | English |
|---|
| | AI agent |
| | Customization |
| じげんすう | Embedding dimensions |
| | Graph database |
| | Grounding |
| コンテキストないがくしゅう | In-context learning (ICL) |
| | Input/output length |
| さいだいしゅつりょくちょう | Maximum output length |
| | Modality |
| モデルじょうりゅう | Model distillation |
| | Model size |
| たげんご | Multilingual |
| | Orchestration |
| | pgvector |
| けんさくかくちょうせいせい | Retrieval-Augmented Generation (RAG) |
| るいじどけんさく | Similarity search |
| | Stop sequence |
| おんど | Temperature |
| | Tool |
| | Top K |
| | Top P |
| | Vector database |
| おもみ | Weights |
3.2 プロンプトエンジニアリング
| 日本語 | 読み | English |
|---|
| （Prompt Management） | | Amazon Bedrock Prompt Management |
| しこうのれんさ | Chain-of-thought (CoT) |
| | Compare variants |
| | Context |
| はっけん | Discovery |
| （DRAFT） | | Draft |
| ろしゅつ | Exposure |
| | Few-shot |
| | Guardrails |
| | Hijacking |
| | Improved response quality |
| | Input data |
| | Jailbreaking |
| | Negative prompt |
| しゅつりょくけいしき | Output format |
| | Poisoning |
| | Prompt |
| | Prompt builder |
| | Prompt injection |
| | Prompt optimization |
| | Prompt template |
| | Prompt version |
| | Single-shot (one-shot) |
| | Specificity and concision |
| | Using multiple comments |
| へんすう | Variable |
| | Zero-shot |
3.3 トレーニングとファインチューニング
| 日本語 | 読み | English |
|---|
| | Continued pre-training |
| | Data curation |
| じょうりゅう | Distillation |
| | Domain adaptation |
| はんよう | General |
| | Governance |
| | Instruction tuning |
| | Labeled |
| | Labeling |
| | Prompt-completion pair |
| だいひょうせい | Representativeness |
| （RLHF） | | RLHF |
| | Size |
| せんよう | specific |
| てんいがくしゅう | Transfer learning |
3.4 基盤モデルの評価
| 日本語 | 読み | English |
|---|
| | A |
| | Amazon Bedrock model evaluation |
| | Automatic evaluation |
| | B testing |
| | Baseline |
| | Benchmark dataset |
| | BERTScore |
| | Bilingual Evaluation Understudy |
| もくひょう | Business goal |
| | component-level evaluation |
| | Context relevance |
| | Cost per interaction |
| | End-to-end |
| | Escalation rate |
| ちゅうじつせい | Faithfulness / Groundedness |
| | Human-in-the-loop evaluation |
| （LLM-as-a-judge） | | LLM-as-a-judge |
| せいさんせい | Productivity |
| | Recall-Oriented Understudy for Gisting Evaluation |
| | Reference text |
| けんさくひんしつ | Retrieval quality |
| | Task completion |
| かんりょうりつ | Task completion rate |
| | Task engineering |
| | Tool-call accuracy |
| | User engagement |
| まんぞくど | User satisfaction (CSAT) |
4.1 責任ある AI の開発
| 日本語 | 読み | English |
|---|
| | Amazon Bedrock Guardrails |
| | ApplyGuardrail API |
| | AWS's eight core dimensions |
| | Balanced dataset |
| | Bias drift |
| | Biased model outputs |
| | Carbon footprint |
| ふきんこう | Class imbalance |
| | Content filters |
| | Contextual grounding check |
| | Curated data sources |
| | Demographic groups |
| | Denied topics |
| たようせい | Diversity |
| | End-user risk |
| | Energy use |
| かんきょうへのはいりょ | Environmental considerations |
| | Human audit |
| ほうかつせい | Inclusivity |
| ちてきざいさんけんしんがい | IP infringement claims |
| | Label quality analysis |
| | Loss of customer trust |
| | Overfitting |
| | Prompt attack protection |
| さいりよう | Reuse |
| てきせいサイズ | Right-sizing |
| けんろうせい | Robustness |
| | Sensitive information filters |
| | Subgroup analysis |
| じぞくかのうせい | Sustainability |
| | Trustworthiness |
| | Truthfulness |
| | Underfitting |
| ぶんさん | Variance |
| しんぴょうせい | Veracity |
| | Word filters |
4.2 透明性と説明可能性
| 日本語 | 読み | English |
|---|
| | Amazon SageMaker Clarify |
| | Amazon SageMaker Model Cards |
| かんさかのうせい | Auditability |
| | AWS AI Service Cards |
| | Black box |
| にんげんちゅうしんせっけい | Human-centered design |
| | Interpretable model |
| | Open-source models, data, licenses |
| | Post-hoc explanation |
| | Source citation |
| | Trade-off |
| とうめいせい | Transparency |
| | Transparency of AI decisions |
| | User feedback mechanism |
5.1 AI システムの保護
| 日本語 | 読み | English |
|---|
| | AgentCore Identity |
| | AgentCore Policy |
| | Amazon Bedrock Guardrails |
| | Amazon Macie |
| とくめいか | Anonymization |
| | Application security |
| | Audit trails and logging |
| | Automated reasoning checks |
| | AWS PrivateLink |
| | Confidence scoring |
| | Data access control |
| | Data catalog |
| かんぜんせい | Data integrity |
| | Data leak prevention |
| | Data lineage |
| | Data provenance |
| | Data quality assessment |
| | Encryption at rest / in transit |
| | Hallucination detection |
| | IAM roles, policies, permissions |
| | Infrastructure protection |
| さいしょうけんげん | Least privilege |
| | Masking |
| | Output filtering and validation |
| しゅつりょくけんしょう | Output validation |
| （PETs） | | Privacy-enhancing technologies |
| | RAG grounding |
| （SageMaker Model Cards） | | SageMaker Model Cards |
| きょういけんしゅつ | Threat detection |
| どくせい | Toxicity |
| ぜいじゃくせいかんり | Vulnerability management |
5.2 ガバナンスとコンプライアンス
| 日本語 | 読み | English |
|---|
| | Amazon CloudWatch |
| | Amazon Inspector |
| | AWS Artifact |
| | AWS CloudTrail |
| | AWS Config |
| | AWS Trusted Advisor |
| | AWS Well-Architected Tool |
| | Data classification |
| | Data lifecycle |
| | Data residency |
| | Generative AI Security Scoping Matrix |
| | Governance framework |
| ろぐきろく | Logging |
| かんさつ | Observability |
| | Policy |
| ほじ | Retention |
| | Review cycle |
| | Review strategy |
| | Team training requirements |
| とうめいせいきじゅん | Transparency standards |