1. The 2018 Portland Echo incident — where a private conversation was sent to a contact — resulted from what technical issue?
Correct. Amazon attributed the incident to background speech accidentally triggering the wake word, "send message," and a contact name in sequence.
Incorrect. Amazon's explanation was that normal speech accidentally triggered wake-word detection followed by misinterpreted send commands.
2. Amazon's Alexa UX research found that which error message style outperformed others on satisfaction?
Correct.
Amazon found that honest, simple phrases like "I'm not sure I understand" with a follow-up clarifying question consistently outperformed longer apologetic error messages.
3. KLM BlueBot's highest satisfaction scores came from which interaction type?
Correct. KLM documented that graceful escalation — where the bot clearly acknowledged its limits and connected to a human agent — produced higher satisfaction than successful self-service.
KLM's data showed escalation conversations with clear limit acknowledgment outperformed even successful automated transactions — redefining what "success" means in error handling.
4. Which approach to intent classification can work without domain-specific labeled training examples?
Correct. Zero-shot classification with an LLM requires only natural language descriptions of each intent — no labeled training examples. The LLM's pretrained knowledge handles classification, though accuracy may be lower than a task-specific trained model on common intents.
Fine-tuned BERT, SVM classifiers, and CRF models all require labeled training data per intent. Zero-shot LLM classification bypasses this requirement by describing intents in plain language and leveraging the LLM's pretrained knowledge to classify without domain-specific examples.
5. The Stanford CASA research framework (Nass & Reeves, 1994) established which principle relevant to chatbot design?
Correct.
CASA showed that humans are hardwired to treat computers with social cues as social agents — regardless of knowing it's a machine.
6. What is "slot filling" in chatbot flow design?
Correct.
Slot filling is the structured data collection phase — gathering departure city, dates, etc. — before the bot can act on the user's request.
7. What does "slot carry-forward" prevent in multi-turn dialogues?
Correct. Slot carry-forward passes entity values from previous turns into new intents, eliminating the frustrating user experience of re-entering information the bot already collected.
Slot carry-forward specifically prevents repetition — passing already-collected values (name, date, order number) through intent transitions so users don't have to repeat themselves.
Correct. Cross-session memory — storing user data between conversations — triggers GDPR Article 13 disclosure obligations and requires opt-in consent in most interpretations.
GDPR data protection obligations engage specifically when data is stored beyond the immediate interaction — cross-session user profiles require Article 13 transparency and typically opt-in consent.
9. Amazon's Alexa Skills Store contained approximately how many skills by 2019?
Correct. The Alexa Skills Store reached over 100,000 skills by 2019, though most saw minimal usage.
Incorrect. The Skills Store had exceeded 100,000 skills by 2019.
10. KLM's BlueBot processed over 16,000 conversations per week. What was its primary design success factor?
Correct.
KLM's success came from careful flow design, not AI sophistication. Mapped paths and graceful handoffs to agents were the foundation.
11. "John told Paul he was wrong" is an example of which type of linguistic ambiguity?
Correct. Referential ambiguity arises when a pronoun or definite noun phrase could refer to more than one antecedent. "He" could be John or Paul — resolving this requires coreference resolution, which remains one of the harder problems in NLP.
Referential ambiguity involves unclear pronoun or noun phrase reference. "He" could refer to either John or Paul. This is distinct from structural ambiguity (multiple parse trees), lexical ambiguity (words with multiple meanings), and scope ambiguity (quantifier scope).
12. Concatenative TTS produces its characteristic robotic quality primarily because of what?
Correct. Concatenative TTS stitches recorded audio segments; the seam points create prosodic discontinuities that sound robotic.
Incorrect. The robotic quality comes from prosodic discontinuities at the join points between spliced audio segments.
13. Word Error Rate is calculated as which formula?
Correct. WER counts substitutions, deletions, and insertions, divided by the total reference word count.
14. Which paper first systematically described turn-taking organization in human conversation?
Correct. The 1974 Sacks, Schegloff & Jefferson paper established the foundational framework for turn-taking in conversation, including transition-relevance places.
Sacks, Schegloff & Jefferson's 1974 paper "A Simplest Systematics for the Organization of Turn-Taking for Conversation" is the foundational reference for turn-taking theory.
15. In the standard voice processing pipeline, which stage runs entirely on-device before any audio is sent to the cloud?
Correct. Wake-word detection uses a tiny on-device model to confirm the trigger phrase before streaming begins.
Incorrect. Wake-word detection is the only stage guaranteed to run on-device before cloud streaming.
16. Bank of America's Erica assistant achieved one billion interactions by 2023. What tone change drove early adoption?
Correct. Shifting from "Your inquiry has been received" to "Got it. Let me pull up your balance." reduced anxiety signals and escalations — tone change with measurable impact.
Erica's tone evolution moved from formal banking language to warmer, direct phrasing — a design change that measurably reduced call center escalations.
17. Word2Vec was developed at which organization and published in which year?
Correct. Word2Vec was published by Tomas Mikolov and colleagues at Google in 2013. Stanford's GloVe followed in 2014; Facebook's FastText appeared in 2016–2017.
Word2Vec was published by Tomas Mikolov et al. at Google in 2013. Stanford published GloVe in 2014; Facebook AI Research published FastText around 2016–2017.
18. According to Forrester Research (2020), the single biggest predictor of high chatbot satisfaction was:
Correct.
Forrester found graceful escalation — not automation rates — was the top satisfaction predictor. Smooth handoffs with context transfer delight users more than bots that stubbornly avoid escalation.
19. "Persona fragmentation" most directly leads to which user behavior?
Correct. When a bot's persona shifts inconsistently, users lose their mental model and tend to abandon the task rather than continue with an unpredictable system.
Persona fragmentation — inconsistent tone across turns — correlates specifically with task abandonment, as users lose confidence in a bot whose "character" they cannot predict.
20. Which tokenization algorithm, adapted for NLP by Sennrich et al. in 2016, is the basis for most modern LLM tokenizers?
Correct. Byte-Pair Encoding, originally a data compression algorithm by Philip Gage (1994), was adapted for NLP subword tokenization by Rico Sennrich et al. in 2016 and is now used directly or as the basis for variants like WordPiece and tiktoken in most major LLMs.
Byte-Pair Encoding (BPE), adapted by Sennrich et al. in 2016, is the foundational algorithm. WordPiece and SentencePiece are related but distinct algorithms. TF-IDF is a document weighting scheme, not a tokenizer.