AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

AI-103 · LESSON 4

Implement text analysis solutions10–15% · 2 topics

Learn text analysis with language models and speech solutions (speech recognition, speech synthesis, translation).

4.1Apply language model text analysis

Concept

Text analysis can be done by instructing an LLM or by using prebuilt Azure AI Language capabilities (sentiment analysis, key phrase extraction, named entity recognition, PII detection, summarization, language detection). Dedicated capabilities suit high-volume processing with fixed output; LLMs suit flexible analysis.

How it works / steps

  1. Decide the purpose of analysis (sentiment, classification, extraction, summarization) and the output format
  2. Choose Azure AI Language for routine processing and an LLM or custom model for proprietary classification
  3. With an LLM, fix the output with a JSON schema
  4. Detect and mask PII before handling documents that contain it
  5. Check accuracy with evaluation data

Workplace example

Aggregate free-text customer survey answers with sentiment analysis and key phrase extraction, and summarize only the notable answers in detail with an LLM.

Common confusions and exam tips

Mini quiz

Q1. You want to extract “customer name,” “product name” and “request” from inquiry emails and pass them to a downstream system. Which output format is most appropriate?

  1. Structured JSON output with a specified schema
  2. Free-form text with a line per item
  3. A Markdown table with item names as headers
  4. JSON mode output without a specified format
Show answer and explanation

Answer: A Structured JSON output with a specified schema

For machine processing downstream, structured output with a schema is easiest to handle. [Other options] Markdown and free text are hard for downstream systems to parse. JSON mode without a schema does not guarantee consistent field names or types.

Q2. You want to prioritize the most dissatisfied inquiries among customer feedback. Which capability is appropriate?

  1. Language detection
  2. Sentiment and tone detection
  3. Named entity recognition
  4. Key phrase extraction
Show answer and explanation

Answer: B Sentiment and tone detection

Sentiment and tone detection enables prioritization and routing. [Other options] Key phrases and entities extract content elements but do not show the degree of dissatisfaction. Language detection cannot judge priority either.

Reference links

4.2Implement speech solutions

Concept

Azure AI Speech provides speech recognition (Speech to Text: real-time, batch, speaker separation), speech synthesis (Text to Speech: neural voices, with pronunciation and pauses tuned by SSML) and speech translation. For real-time voice conversation, real-time speech APIs (such as Voice Live) can also be used.

How it works / steps

  1. Decide the use (captions, minutes, call center analytics, voice responses)
  2. Choose real-time or batch (batch transcription for large volumes of recordings)
  3. Improve accuracy on technical terms with phrase lists or custom speech
  4. Tune synthesized pronunciation and speed with SSML
  5. Distinguish speakers with speaker separation (diarization)

Workplace example

Batch-transcribe call center recordings every night with speaker separation, and use an LLM to create summaries and service evaluations.

Common confusions and exam tips

Mini quiz

Q1. Meeting recordings full of manufacturing jargon produce many transcription errors. What is the most appropriate improvement?

  1. Improve recognition of technical terms with a custom speech model
  2. Switch to speech translation and check the terms in the English translation
  3. Lower the recording sampling rate to make files lighter
  4. After transcription, have an LLM guess and correct errors
Show answer and explanation

Answer: A Improve recognition of technical terms with a custom speech model

Custom speech models are effective for industry and internal terms. [Other options] Guessing corrections afterward does not fix the cause of errors. Switching to translation leaves recognition errors. Lowering the sampling rate reduces accuracy.

Q2. You want to convey what is said in Japanese to overseas sites as English speech. Which capability is appropriate?

  1. Speech synthesis only
  2. Speech translation
  3. Speech recognition and sentiment analysis
  4. Text translation
Show answer and explanation

Answer: B Speech translation

Speech translation converts spoken language into another language. [Other options] Speech recognition and sentiment analysis do not translate. Text translation does not handle speech input and output. Speech synthesis alone cannot understand Japanese speech.

Reference links

Lessons are written by this site based on public information and do not guarantee exam content. Product specifications change, so check the latest information at the reference links.

Notes (unofficial, original questions, no pass guarantee)