Spell checker for consumer language (CSpell).

Area of Science:

• Computational linguistics
• Health informatics
• Natural language processing

Background:

• Automated understanding of consumer health inquiries is crucial for accessible health information.
• Misspellings in health questions pose a significant barrier to accurate automated processing.
• Existing spell-checking tools may not adequately address the nuances of health-related language.

Purpose of the Study:

• To develop a robust spell-checking tool, CSpell, for detecting and correcting diverse spelling errors in consumer health questions.
• To enhance the accuracy of automated analysis of health-related user queries.
• To improve the reliability of information retrieval from consumer health inquiries.

Main Methods:

• Development of a distributable spell-checking tool, CSpell.
• Implementation of a novel dual-embedding approach within Word2vec for context-dependent corrections.
• Integration of dictionary-based corrections within a two-stage ranking system.
• Development of specialized splitters and handlers for word boundary errors.

Main Results:

• Achieved an F1 score of 80.93% for spelling error detection.
• Achieved an F1 score of 69.17% for spelling error correction.
• The dual-embedding model improved F1 score by 9.13% compared to standard Word2vec cosine similarity.
• The two-stage ranking system showed a 4.94% improvement over the best one-stage system.

Conclusions:

• CSpell demonstrates superior performance over the state-of-the-art in misspelling detection and correction for consumer health questions.
• The tool provides near real-time automatic correction capabilities.
• The software and test set are publicly available for further research and application.