Interview Question Bank
Multilingual Text Classification with BERT
Prepared for interviews at non-technical / business-focused companies
Focus: what you built, why it mattered, and how you explain it simply
How to Use This Document
Non-technical interviewers rarely ask deep model-architecture questions. They care about what problem you
solved, why it mattered, how you approached it, what went wrong, and what you learned — told in plain
language. Each question below has a short note on the angle they are probing for. Practice answering out loud in
30–60 seconds per question; avoid jargon like 'macro F1' or 'tokenization' unless you first explain it in one plain
sentence.
Section 1: Project Overview (Opening Questions)
These are almost guaranteed. They open the conversation and set the tone.
Q1. Can you walk me through this project in simple terms?
How to answer: Give the one-line version first: 'I taught a computer program to automatically read product reviews and
figure out whether they were positive or negative, in multiple languages.' Then briefly mention scale (200,000 reviews)
and the interesting finding (it worked great in one language but struggled across languages).
Q2. What problem were you trying to solve, and why does it matter for a business?
How to answer: Tie it to a real business use case: companies get thousands of customer reviews and can't read them all
manually. Automatic classification saves time, flags unhappy customers faster, and helps identify trends across
markets/countries.
Q3. What was your role in this project?
How to answer: Be honest and specific — e.g., 'I set up and trained the models, ran the experiments, and analyzed why
performance changed across languages.' If it was a bootcamp/team project, clarify what you personally owned.
Q4. Why did you choose this project / this technology?
How to answer: Connect it to curiosity or a skill gap you wanted to close — e.g., wanting hands-on experience with
modern AI language tools that companies increasingly rely on for customer feedback analysis.
Q5. In one sentence, what is BERT, in language a non-technical person would understand?
How to answer: Something like: 'BERT is a computer program that has already read a huge amount of text and learned
how language works, so I only had to teach it my specific task instead of teaching it English from zero.'
Section 2: Results and Business Impact
Q6. What results did you achieve?
How to answer: Lead with the headline number: 97% accuracy on same-language classification. Then be upfront about
the honest finding: performance dropped sharply when the model was tested on a different language than it was trained
on.
Q7. Were you happy with the results? Why or why not?
How to answer: Show balanced judgment: proud of the strong single-language result, but genuinely curious/concerned
about the cross-language drop, since that's the more realistic real-world scenario for a global company.
Q8. How would this be useful in a real company?
How to answer: Examples: auto-flagging negative reviews for customer support, tracking sentiment trends by region,
reducing manual review time, prioritizing urgent complaints.
Q9. If this were a live product, what risk would you flag to your manager?
How to answer: The cross-language weak spot: if the company operates in multiple countries and only tests in English,
the tool could quietly perform badly for non-English customers without anyone noticing — a real business/reputational
risk.
Q10. How did you measure success? Why that metric and not something simpler?
How to answer: Explain accuracy vs. a fairer metric in plain terms: 'Accuracy can look good even if the model ignores
smaller groups. I used a metric [macro F1] that checks performance on every language/category fairly, not just the
biggest one — that's actually what exposed the problem.'
Section 3: Process and Approach
Q11. Walk me through your process, step by step.
How to answer: Data (200k reviews) → chose two pre-trained models suited to the task → trained/fine-tuned them →
tested in two setups (same-language and cross-language) → compared results → investigated why performance
dropped.
Q12. How long did this take, and what did a typical day/week look like?
How to answer: Be concrete and honest about time spent — setup, training runs, debugging, analysis, writing up
conclusions.
Q13. Did you work alone or with a team? How did you divide the work?
How to answer: If part of a bootcamp cohort, describe any collaboration, code/data sharing, or peer review, and what
you specifically drove.
Q14. What tools did you use, and why?
How to answer: Name them simply: pre-trained AI language models (BERT variants), Python, and standard machine
learning libraries — no need to go deeper unless asked.
Section 4: Challenges and Problem-Solving (Very Common)
Non-technical interviewers love this category because it reveals how you think and handle setbacks, without
needing to judge your code.
Q15. What was the biggest challenge in this project?
How to answer: The cross-language performance collapse (F1 from 0.51 to 0.15) is your strongest story — a clear,
measurable problem you had to explain and investigate.
Q16. Something didn't work as expected — what did you do?
How to answer: Describe your investigation process: forming a hypothesis (why would it fail across languages?),
checking the data/vocabulary handling, and reasoning through the explanation rather than guessing.
Q17. What was the hardest thing to explain to someone else about this project?
How to answer: Likely the difference between overall accuracy looking fine while performance for specific languages
was actually poor — a nuance that's easy to misread.
Q18. If you had more time or resources, what would you have done differently?
How to answer: Mention (in plain language) trying a model built for many languages more evenly, or giving the model
some exposure to the target language before the main task, rather than relying purely on 'zero prior exposure' transfer.
Q19. Tell me about a mistake or wrong assumption you made during this project.
How to answer: A safe, honest one: initially assuming the multilingual model would transfer well across languages 'for
free' without extra adaptation, then learning that wasn't the case.
Section 5: 'Why Not Just...' Questions (Testing Understanding, Not
Jargon)
Q20. Why couldn't you just build/train this model completely from scratch yourself?
How to answer: Analogy: 'It's like hiring someone who already knows the language and just needs training on our
specific job, versus teaching someone a language from zero — the first is faster, cheaper, and more reliable with the
amount of data I had.' Mention pre-training needs billions of words and huge computing cost, which wasn't practical for
a 200k-review project.
Q21. Why does it perform so differently across languages if it's the 'same' model?
How to answer: Plain explanation: the model was fine-tuned only on one language's examples, so it got very good at
that language's patterns specifically, and that specialization didn't transfer evenly to other languages — especially ones
less represented in its original training.
Q22. Why not just translate everything into English first and use the English model?
How to answer: Good question to anticipate even if not asked: translation can lose nuance, sarcasm, and cultural
context, and adds another point of failure/cost. It's a valid alternative worth mentioning as something you'd consider, not
dismiss.
Q23. How do you make sure the accuracy number is trustworthy and not misleading?
How to answer: Explain you looked beyond the single headline number (97%) and checked performance per
language/category, which is exactly how you caught the cross-language issue — this shows rigor, not just chasing a
good-looking number.
Section 6: Learning, Growth, and Reflection
Q24. What did you learn from this project that surprised you?
How to answer: The gap between 'looks impressive on paper' (97% accuracy) and 'works reliably in the real world
across all users' (F1 dropping to 0.15) — a genuinely useful lesson about evaluating things fairly.
Q25. What skills did this project help you build?
How to answer: Structuring an experiment, comparing results fairly, diagnosing why something failed, and explaining a
technical result to a non-technical audience — all transferable skills.
Q26. How does this project relate to the role you're applying for here?
How to answer: Tailor this to the company: if the role touches customer feedback, data-driven decisions, or
communicating analysis to non-technical stakeholders, draw a direct line to that.
Q27. What would you tell someone just starting a similar project?
How to answer: Something like: don't just look at the overall score — check whether it holds up fairly across every
group/language before trusting it.
Section 7: Curveball / Simplicity Checks
Interviewers sometimes deliberately ask a very basic question to see if you actually understand your own project
or just memorized a script.
Q28. If I knew nothing about AI, how would you explain what 'training' a model means?
How to answer: 'Training is showing the program lots of examples with the right answers already labeled, so it can
learn the pattern — similar to how a new employee learns by seeing worked examples before doing the job alone.'
Q29. What does '97% accuracy' actually mean in plain terms?
How to answer: Out of 100 reviews it hadn't seen before, it correctly guessed positive/negative for about 97 of them,
when tested on the same type/language of review it was trained on.
Q30. Explain, without any jargon, why performance dropped for other languages.
How to answer: 'The program only practiced on English examples, so it got really good at English specifically — like
someone who trained hard for one exam and then got tested on a different, related exam they hadn't studied for.'
Q31. What would you do next if this were your job, not a bootcamp project?
How to answer: Show product thinking: talk to stakeholders about which languages/markets matter most, prioritize
fixing those, add ongoing monitoring per language, and set expectations that the tool isn't equally reliable everywhere
yet.
Quick Reference: Your 60-Second Summary
Keep this ready — it answers 'tell me about this project' and can be trimmed or expanded depending on how deep
they want to go.
• What: Built AI models to automatically classify 200,000 product reviews as positive or negative.
• How: Used two ready-made language models — one for English, one built for many languages — and
trained each on the review data.
• Result: 97% accuracy when tested on the same language it was trained on.
• The twist: When tested on a different language than it was trained on, performance dropped sharply — a
realistic and important finding for any global business.
• Why it matters: Shows both the power and the limits of off-the-shelf AI tools, and the importance of testing
fairly across every group of users, not just looking at one overall score.