Business Context

LinguaPrep, an English-learning platform, wants to automatically score sentence completion questions where learners choose the grammatically correct word or phrase to fill a blank. The goal is to build an NLP system that predicts the correct completion option and supports fast feedback in practice sessions.

Data

The training set contains approximately 180,000 multiple-choice sentence completion items collected from grammar exercises and exam prep content. Each example includes a sentence with one blank, 4 candidate options, and the correct answer. Sentences are short to medium length (8-35 tokens), written in English, and cover common grammar topics such as verb tense, subject-verb agreement, articles, prepositions, and pronouns. Label distribution is balanced across answer positions, but grammar categories are unevenly distributed, with tense and preposition errors appearing most often.

Success Criteria

A production-ready solution should achieve at least 88% accuracy on a held-out test set and maintain low latency (<50ms per question) for real-time quiz feedback. The approach should also be interpretable enough to explain common failure modes.

Constraints

• Must run on a single CPU instance for inference
• Training should be feasible on one modern GPU or a strong laptop
• Only English text is in scope
• The system should handle noisy punctuation and minor formatting inconsistencies

Requirements

1. Formulate the task as an NLP prediction problem for grammar-based sentence completion.
2. Build a preprocessing pipeline for blank-based multiple-choice inputs.
3. Implement a baseline using TF-IDF or n-gram features and a stronger transformer-based model in Python.
4. Explain how you would score candidate options and select the final answer.
5. Evaluate performance by grammar category, not just overall accuracy.
6. Describe likely error cases such as semantically plausible but grammatically incorrect options.