Accenture Skill Proficiency Test - Natural Language Processing (NLP) - Dec 2025

See All: Miscellaneous Interviews @ Accenture

📘 Accenture Skill Proficiency Test - NLP - Question Report

🧠 SECTION 1: NLP & WORD EMBEDDINGS

Q1. What is GloVe?

Extracted options (interpreted):

• Matrix factorization of (raw) PMI values with respect to squared loss

• Matrix factorization of log-counts with respect to weighted squared loss

• Neural network that predicts words in context and learns from that

• Neural network that predicts words based on similarity and embedding

✅ Correct Answer

✔ Matrix factorization of log-counts with respect to weighted squared loss

💡 Hint

• GloVe = Global Vectors

• Combines count-based methods (co-occurrence matrix) with predictive embedding ideas

• Objective minimizes weighted squared error between word vectors and log(co-occurrence counts)

---

🧠 SECTION 2: TRANSFORMERS & CONTEXTUAL MODELS

Q2. In which architecture are relationships between all words in a sentence modeled irrespective of their position?

Extracted options:

• OpenAI GPT

• BERT

• ULMFiT

• ELMo

✅ Correct Answer

✔ BERT

💡 Hint

• BERT uses bidirectional self-attention

• Every token attends to all other tokens simultaneously

• GPT = causal (left-to-right), not fully bidirectional

---

📊 SECTION 3: EVALUATION METRICS

Q3. Log loss evaluation metric can have negative values

Options:

• True

• False

✅ Correct Answer

✔ False

💡 Hint

• Log loss = negative log likelihood

• Probabilities ∈ (0,1) → log(probability) ≤ 0 → negative sign makes loss ≥ 0

• Log loss is always ≥ 0

---

Q4. Which metric is used to evaluate STT (Speech-to-Text) transcription?

Extracted options:

• ROUGE

• BLEU

• WER

• METEOR

✅ Correct Answer

✔ WER (Word Error Rate)

💡 Hint

• WER = (Insertions + Deletions + Substitutions) / Total words

• Standard metric for speech recognition

---

🧩 SECTION 4: NLP ALGORITHMS & TAGGING

Q5. Best Cut algorithm works for:

Extracted options:

• Text classification

• Coreference resolution

• POS tagging

✅ Correct Answer

✔ Text classification

💡 Hint

• Best Cut / Min Cut → graph-based partitioning

• Often used for document clustering / classification

---

Q6. BIO tagging is applicable for:

Extracted options:

• Text classification

• Coreference resolution

• NER

• N-grams

✅ Correct Answer

✔ NER (Named Entity Recognition)

💡 Hint

• BIO = Begin – Inside – Outside

• Used in sequence labeling tasks

• Especially common in NER and chunking

---

🎯 SECTION 5: ATTENTION MECHANISMS

Q7. Which among the following are attention mechanisms generally used in neural network models?

Extracted options:

• Bahdanau attention

• Karpathy attention

• Luong attention

• ReLU attention

• Sigmoid attention

✅ Correct Answers

✔ Bahdanau attention
✔ Luong attention

❌ Incorrect

• Karpathy (not an attention mechanism)

• ReLU / Sigmoid (activation functions, not attention)

💡 Hint

• Bahdanau = additive attention

• Luong = multiplicative (dot-product) attention

• Activations ≠ attention

---

📚 SECTION 6: TEXT PROCESSING & TOPIC MODELING

Q8. Porter Stemmer is used for:

✅ Correct Answer

✔ Stemming words to their root form

💡 Hint

• Example: running → run

• Reduces vocabulary size

• Rule-based, not dictionary-based

---

Q9. LDA stands for:

✅ Correct Answer

✔ Latent Dirichlet Allocation

💡 Hint

• Probabilistic topic modeling

• Documents = mixture of topics

• Topics = distribution over words

---

Q10. Which of the following are true to choose optimal number of topics in LDA for topic modeling?

Extracted options:

• Min coherence value

• Max model perplexity

• Max coherence value

• Min model perplexity

✅ Correct Answers

✔ Max coherence value
✔ Min model perplexity

💡 Hint

• Coherence → interpretability (higher is better)

• Perplexity → generalization (lower is better)

• Best practice: balance both

---

Tags: Interview Preparation,Natural Language Processing,