Question 20 A global commerce catalog contains French descriptions that English-language agents need to read while preserving product measurements and part numbers. Which task directly addresses the language barrier? Machine translation from French to English with terminology and part-number preservation checks. Clustering all French descriptions by similarity. Extracting only prices and dimensions. Generating entirely new product stories without using the source. ExplanationCorrect Answer A:Translation maps text across languages, while preserving critical identifiers must be validated separately. Other Answers B:Clusters are not English-language versions of product descriptions. C:Field extraction omits the requested readable translated descriptions. D:Unanchored generation can invent dimensions and product attributes.
Question 21 An accounting team needs to populate amount_due, invoice_date and vendor_id from a large collection of invoices. Which GenAI task best maps source content to these structured fields? Open-ended storytelling based on invoice headings. Information extraction into a typed invoice-field schema, with downstream validation. Sentiment scoring on each vendor name. Automatic language identification as the only processing step. ExplanationCorrect Answer B:The target is a set of named factual attributes rather than a summary or sentiment label. Other Answers A:A creative narrative does not produce reliable vendor IDs or amounts. C:The task is document field extraction, not opinions or affect. D:Language detection may aid extraction but cannot produce the required invoice fields.
Question 22 An online retailer wants customers to find products similar to a search phrase even when the wording differs from item descriptions. Which capability most directly supports semantic matching? Count only exact keyword matches and reject all paraphrases. Perform language translation on every product into the same original language. Encode queries and product text in a compatible embedding space and retrieve by vector similarity. Generate a random list of product IDs without indexing product text. ExplanationCorrect Answer C:Embeddings capture semantic proximity, enabling retrieval beyond exact token overlap when indexing and evaluation are suitable. Other Answers A:Exact keyword matching alone misses semantic matches expressed with different vocabulary. B:Translation is not by itself a semantic-nearest-neighbor retrieval system. D:Random selection cannot reflect a customer’s stated search meaning.
Question 23 A bank wants an internal assistant to answer current fee questions using authoritative policy documents, including the policy section cited in each reply. What application pattern most directly addresses this need? Unconditional free-form generation from a base model without retrieval. Sentiment analysis of the user’s fee question. Fine-tune only on last year’s fees and never update the knowledge source. Retrieval-augmented question answering over approved current policy material, with citations and access controls. ExplanationCorrect Answer D:RAG combines retrieval of authoritative current context with answer generation, reducing dependence on stale parametric memory. Other Answers A:A model-only answer can be stale or uncited and ignores the authoritative policy requirement. B:The user needs a factual fee answer and citation, not an attitude label. C:Static fine-tuning cannot reliably reflect new policy documents as they change.
Question 24 A quality-assurance team wants product reviews grouped into complaints about durability, fit and delivery, allowing more than one theme. Which approach best matches this requirement? Multi-label text classification with clearly defined theme labels and a review set for overlapping categories. Force exactly one label for every review regardless of mixed topics. Generate a one-number similarity score with no category definitions. Translate the review to English and treat the translation as the theme labels. ExplanationCorrect Answer A:A single review can mention fit and delivery, so independent category labels better match the stated multi-theme output. Other Answers B:Single-label routing loses valid additional complaints in reviews discussing several issues. C:A lone similarity value cannot indicate which named complaint themes apply. D:Translation may prepare text but does not identify durability, fit or delivery topics.
Question 25 An operations team wants a daily readable digest of 500 incident notes emphasizing the few significant customer-impacting events. Which task and control are most appropriate? Classify only the language of each note. Document summarization with severity-aware selection and source-linked factual review. Convert all notes into embeddings and present the vectors to managers. Generate a generic incident report unrelated to the day’s records. ExplanationCorrect Answer B:The output is a concise digest that preserves significant facts from many notes; importance selection must be checked. Other Answers A:Language labels do not convey what happened or the customer impact. C:Vectors are useful for search, not a readable management digest. D:A generic report cannot accurately summarize the actual daily incidents.
Question 26 A manufacturer wants to spot when two differently worded service bulletins describe the same underlying failure. What capability is best for candidate matching? Sort bulletins alphabetically by title and assume adjacent ones describe the same fault. Translate both English bulletins into English again. Semantic similarity using text embeddings, followed by a human confirmation step for high-impact matches. Label every bulletin as positive or negative sentiment. ExplanationCorrect Answer C:Embedding similarity identifies candidate conceptual matches even when exact phrases differ; it is evidence for review rather than proof of equivalence. Other Answers A:Lexicographic order does not represent causal similarity. B:No language conversion is needed to detect semantic overlap. D:Sentiment polarity does not establish that failures are technically the same.
Question 27 A content platform wants a warning when user reviews contain explicit abusive language before those reviews are published. Which model task fits the decision? Generate a promotional summary of the review. Use semantic nearest-neighbor search without any policy threshold or label. Translate every abusive review and then publish it unreviewed. Safety or toxicity classification, with clear policy categories and a review route for uncertain results. ExplanationCorrect Answer D:This is a policy-based classification decision about a text, not a request to paraphrase abusive content. Other Answers A:A summary does not directly determine whether publication violates the abuse policy. B:Similarity alone does not define which abusive categories should block publication. C:Translation does not constitute a moderation control.
Question 28 A health benefits portal asks callers to supply a policy identifier and then must extract the corresponding coverage limit from the current plan document. Which system design best meets the business objective? Retrieve the exact authorized plan section by policy ID, then extract the coverage-limit field with source validation. Generate a likely coverage limit from generic health-insurance training data. Analyze whether the caller sounds pleased with the policy. Translate the identifier into several languages and treat each translation as a limit. ExplanationCorrect Answer A:The limit is a factual field in a versioned document; retrieval and extraction are both necessary to keep it tied to current policy. Other Answers B:A plausible number is not the actual plan value and may be dangerously wrong. C:Sentiment does not identify the contractual coverage limit. D:Policy IDs are lookup keys, not coverage figures or natural-language translation tasks.
Question 29 A local agency receives free-text service requests and wants to find all named street addresses and dates in them for a scheduling system. Which GenAI task is most direct? Summarize every service request into one opinion sentence. Named-entity or schema-based information extraction for addresses and dates, with parser validation. Perform text-to-image generation of a map without extracting locations. Run only language detection and store the detected language. ExplanationCorrect Answer B:The requested output consists of entities with types and spans/values rather than a summary of the request. Other Answers A:A single opinion may omit addresses and dates necessary for scheduling. C:A generated image does not provide the typed values the scheduling system needs. D:Language identity is not the requested address/date data.
Question 30 A company has a collection of expense memos and wants a one-sentence description of each memo that preserves the main expense purpose. What NLP task should drive the output? Generate a new memo about an unrelated expense. Return only whether the memo is written in English. Abstractive or extractive summarization constrained to one faithful sentence. Produce the embedding vector and display its raw numbers. ExplanationCorrect Answer C:The requirement is to condense each original memo while preserving intent rather than assign a category. Other Answers A:Creating different content is not a faithful one-sentence gist of the source. B:Language detection does not describe the expense purpose. D:Embedding coordinates are not a readable summary sentence.
