Task 1.3: Identify AI workloads

• AText analysis

  Incorrect

  Text analysis examines existing text, for example to find key phrases or sentiment. It doesn't write a new itinerary.

• BAgentic AI

  Incorrect

  Agentic AI automates tasks by using tools and acting in other systems. Writing a description from notes is content generation alone.

• CGenerative AI

  Correct

  Creating new, original text from a short prompt is the defining task of generative AI.

• DInformation extraction

  Incorrect

  Information extraction works on content that already exists. Here the goal is to produce new text.

Generative AI produces new content such as text, images or code in response to a prompt. Drafting documents as a starting point for people to refine is one of its most common uses.

• AKey phrase extraction

  Correct

  Key phrase extraction returns the main concepts in a text as a list, which suits tagging and indexing.

• BSentiment analysis

  Incorrect

  Sentiment analysis tells whether text is positive, negative or neutral. It doesn't name what the article is about.

• CLanguage detection

  Incorrect

  Language detection identifies the language a document is written in, often as a first routing step. It doesn't find topics.

• DText summarization

  Incorrect

  A summary condenses the article into a few sentences. It doesn't produce a set of short tags.

Key phrase or keyword extraction identifies the words and phrases that capture a document's main points, which helps with indexing, tagging and search. In Foundry, a general-purpose model can perform it from a natural language prompt.

• ASpeech recognition (speech to text)

  Incorrect

  That capability works in the other direction. Here the input is already text and the output is audio.

• BSpeech synthesis (text to speech)

  Correct

  Speech synthesis turns text into audible speech, which is exactly what reading a chapter aloud requires.

• CSpeech translation

  Incorrect

  Speech translation converts speech in one language into another language. The chapter is read in its own language.

• DSpoken language identification

  Incorrect

  Language identification works out which language is spoken in audio. Nobody is speaking into the app.

Speech synthesis, or text to speech, vocalizes text. Neural voices make the result sound natural, and reading articles or books aloud is a common scenario.

• ARecord the dictations with higher-quality microphones

  Incorrect

  Better audio helps when noise causes errors, but ordinary words are already recognized well, so the gap is vocabulary.

• BAdd Speech Synthesis Markup Language (SSML) to each request

  Incorrect

  SSML controls how text to speech pronounces and paces output. It doesn't affect how spoken audio is recognized.

• CTurn on profanity filtering for the transcripts

  Incorrect

  Profanity filtering masks inappropriate words in the output. It doesn't teach the model new terms.

• DTrain a custom speech model with radiology text and audio

  Correct

  Custom speech augments the base model with domain text and audio data, which improves recognition of specialized vocabulary.

The base speech recognition model works well for common language but can struggle with industry jargon. Custom speech adds domain-specific text and audio so the model recognizes those terms more accurately.

• AImage classification

  Incorrect

  Image classification predicts one label for the image as a whole. It doesn't locate several animals in the same photo.

• BFace detection with the Azure Face service

  Incorrect

  Face detection finds human faces and their locations. It doesn't identify animal species.

• CObject detection

  Correct

  Object detection finds each object in an image, labels it and returns the coordinates of a bounding box around it.

• DImage generation of a reference picture for each species

  Incorrect

  Generating new pictures doesn't locate animals in the real camera photos.

Image classification answers "what is this an image of?" Object detection goes further: it identifies multiple objects and returns where each one appears as a bounding box.

• AField extraction that maps recognized text to named schema fields

  Correct

  OCR says what text is on the page; field extraction decides what that text means and which business field it belongs to.

• BImage preprocessing that raises contrast and corrects page skew

  Incorrect

  Preprocessing improves how well characters are recognized. The output is still unlabeled text.

• CLanguage detection that identifies the language of each invoice

  Incorrect

  Knowing the invoice language can help routing, but it doesn't turn text into labeled values.

• DTranslation of the recognized text into English

  Incorrect

  Translating the text changes its language but doesn't separate it into labeled values.

Information extraction usually combines OCR, which finds and reads text, with an analytical model that maps values to specific fields, such as matching receipt text to the fields of an expense claim.

• ADiscarding characters whose OCR confidence score is below a threshold

  Incorrect

  A confidence filter catches uncertain readings. In this case the characters were read with high confidence, so nothing is filtered.

• BCross-field validation that checks line items add up to the total

  Correct

  Checking relationships between extracted fields catches values that are individually readable but inconsistent with the rest of the document.

• CNormalizing extracted dates and amounts to standard formats

  Incorrect

  Normalization makes dates and amounts consistent in format. It doesn't compare the total with anything.

• DDetermining the reading order of the text blocks on the page

  Incorrect

  Reading order puts text in the right sequence during OCR. It doesn't compare values to each other.

Confidence scores tell you how sure the recognizer was, not whether the business data is consistent. Cross-field validation, such as confirming that line items sum to the invoice total, catches errors that confidence scores miss.

• AEntity recognition to list the people mentioned in each call

  Incorrect

  Listing names that come up in a call doesn't tell whether the customer was satisfied.

• BSpeech translation to convert the calls into another language

  Incorrect

  Translation helps when callers and analysts speak different languages. Nothing in the scenario needs a second language.

• CSpeech recognition to transcribe the calls

  Correct

  Speech recognition, or speech to text, turns the recorded audio into a transcript that can be analyzed.

• DImage classification to label the call records

  Incorrect

  Image classification assigns labels to images. Call recordings are audio, not images.

• EText analysis to determine the sentiment of each transcript

  Correct

  Sentiment analysis classifies text as positive, negative or neutral, which answers whether the customer was satisfied.

Many solutions chain workloads. Transcribing calls is speech recognition; judging whether the customer was satisfied is sentiment analysis, a text analysis technique applied to the transcript.

• AStop word removal

  Incorrect

  Stop word removal drops common words such as "the"; it doesn't join words into phrases.

• BSentiment analysis

  Incorrect

  Sentiment analysis labels the tone of text; it doesn't join words into phrases.

• CLanguage detection

  Incorrect

  Language detection identifies the language of the text, not phrases.

• DN-gram extraction

  Correct

  N-gram extraction finds multi-term phrases, such as two-word bigrams, so they can be counted as a group.

Many text analysis tasks work better when frequent word sequences are treated as units. N-gram extraction finds unigrams, bigrams, trigrams and longer phrases such as "natural language processing".

• ARetrieval-augmented generation with a vector index

  Incorrect

  RAG grounds a generative model in retrieved content; it isn't a word count classifier.

• BFrequency analysis of each email's most common terms

  Incorrect

  Frequency analysis counts terms to suggest what a document is about. On its own it doesn't learn from labeled emails or predict a class.

• CA bag-of-words model with a Naive Bayes classifier

  Correct

  Bag-of-words turns text into word frequencies, and Naive Bayes applies Bayes' theorem to predict the most probable class.

• DTextRank over a graph of each email's sentences

  Incorrect

  TextRank is an unsupervised graph algorithm that ranks sentences or words by similarity. It doesn't predict a label such as spam.

Bag-of-words represents text as a vector of word frequencies or occurrences, ignoring grammar and order. It feeds classic machine learning classifiers such as Naive Bayes, for example in spam filtering or basic sentiment classification.

• AComputer vision, used for visual search

  Correct

  Analyzing photos to find visually similar items is a common computer vision scenario known as visual search.

• BImage generation, used to draw a matching chair

  Incorrect

  Image generation creates new pictures. The shopper wants existing catalog chairs that look like the photo, not a newly drawn one.

• COptical character recognition, used to read labels

  Incorrect

  OCR reads printed or handwritten text in an image. A chair's shape and style, not text, decide which catalog items are similar.

• DText analysis, used to match catalog descriptions

  Incorrect

  Text analysis works on written text. The shopper provides a photo, so the match has to come from what the image shows.

Computer vision deals with visual input such as photographs, videos and live camera feeds. Finding visually similar items from a photo, known as visual search, is one of its common scenarios.

• ALanguage detection on each voicemail transcript

  Incorrect

  Language detection identifies the language spoken, not the caller's details.

• BInformation extraction from the audio into defined fields

  Correct

  Extracting specific values from recordings into a schema is information extraction, here applied to audio.

• CSentiment analysis of the caller's tone in each voicemail

  Incorrect

  Sentiment tells whether the caller sounded positive or negative, not their contact details.

• DSpeech synthesis of a spoken reply to each voicemail

  Incorrect

  Speech synthesis creates audio; the team needs data out of the recordings.

Information extraction finds and structures data in unstructured sources, increasingly including audio and video. Turning recordings into named fields, such as caller and callback number, is a typical example.