Context-Sensitive Spelling Correction of Clinical Text via Conditional Independence.

Area of Science:

• Clinical Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Misspellings are frequent in clinical medical records, posing challenges for natural language processing.
• Limited availability of labeled clinical datasets hinders the development of effective spelling correction models.

Purpose of the Study:

• To develop a probabilistic model for clinical spelling correction that requires minimal or no real-world data.
• To address the data scarcity issue in training machine learning models for clinical text.

Main Methods:

• A probabilistic model based on a conditional independence assumption was developed.
• The model was modularly decomposed into a language model and a corruption model.
• A deep character-level language model was trained on a large clinical corpus, complemented by an edit-based corruption model.

Main Results:

• The proposed model significantly outperforms existing baseline methods on healthcare spelling correction tasks.
• The model demonstrates effectiveness even with small or no labeled real-world clinical data.

Conclusions:

• The developed probabilistic model offers a robust solution for clinical spelling correction, overcoming data limitations.
• This approach enhances the reliability of processing medical records and advancing clinical natural language processing.