Business Context

LexiLearn, an English-learning platform, wants to automatically grade vocabulary questions where learners must identify the word most opposite in meaning to a target word. Build an NLP system for the prompt: identify the word most opposite in meaning to "mitigate".

Data

The training set contains 180,000 multiple-choice vocabulary items collected from prior assessments. Each item includes a target word, 4 candidate options, a correct antonym label, and optional example sentences. Text is English only. Target words are short tokens or short phrases, while supporting context ranges from 5-80 words. Label distribution is moderately imbalanced because some distractor types appear more often than true antonyms.

A representative item may include lexical pairs such as mitigate → worsen, abate → intensify, or expand → contract. Some examples contain morphology variants, polysemy, and domain-specific usage that make simple dictionary lookup unreliable.

Success Criteria

A good solution should achieve at least 90% accuracy on held-out easy/intermediate vocabulary items and maintain strong performance on unseen words through semantic generalization. The system should return the best antonym option with inference latency under 50 ms per item in batch scoring.

Constraints

• Must run in a standard Python inference service on CPU
• Should support explainable scoring for educational review
• External API calls are not allowed during inference

Requirements

1. Build a model that selects the candidate word most opposite in meaning to a target word.
2. Design preprocessing for short lexical inputs and optional sentence context.
3. Implement a modern Python solution using transformers or static embeddings as a baseline.
4. Evaluate against strong lexical baselines and report failure modes for polysemous words.
5. For the example item, determine the best antonym of "mitigate" from candidate options.