Extract companies from noisy text
Role: Data Scientist
Category: Machine Learning
Difficulty: hard
Interview Round: Onsite
You receive messy resume text and scraped web snippets; the task is to extract company names. (a) Design a hybrid system combining rule-based patterns (e.g., legal suffixes, context windows) with a machine-learned NER model; discuss handling of casing, Unicode noise, and misspellings. (b) Explain feature choices or embeddings (e.g., subword, contextual) and how to incorporate a company gazetteer with fuzzy matching while avoiding label leakage. (c) Define evaluation metrics (entity-level precision/recall/F1) and error analysis procedures for acronyms and ambiguous tokens (e.g., Apple vs apple).
Overview: This question evaluates named entity recognition, noisy-text preprocessing, entity disambiguation, and hybrid rule- and model-based pipeline design for extracting organization names from unstructured resumes and web snippets.