Automatic Correction of Real-Word Errors in Spanish Clinical Texts.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Medical Informatics

Background:

• Real-word errors, actual dictionary terms used incorrectly, pose challenges in text correction.
• Traditional methods rely on word sequence frequency, while deep learning offers advanced contextual understanding.
• Accurate clinical text is crucial for patient safety and medical research.

Purpose of the Study:

• To implement and evaluate a deep learning model for correcting real-word errors specifically in clinical text.
• To compare the performance of Seq2seq Neural Machine Translation models trained on clinical versus general text corpora.
• To investigate the impact of pre-trained word embeddings (GloVe, Word2Vec) on error correction accuracy.

Main Methods:

• Developed a Seq2seq Neural Machine Translation model to map erroneous sentences to corrected ones.
• Generated rule-based real-word errors in correct sentences for training and evaluation.
• Trained and evaluated models on the Wikicorpus and a smaller, privacy-conscious clinical dataset.
• Utilized GloVe and Word2Vec pre-trained word embeddings.

Main Results:

• Seq2seq models trained on the smaller clinical corpus demonstrated superior performance compared to those trained on the larger Wikicorpus.
• Despite data limitations, clinical text-specific models achieved better real-word error correction.
• Pre-trained word embeddings showed varying performance, with potential for improvement.

Conclusions:

• Deep learning, particularly Seq2seq models, is effective for real-word error correction in clinical text.
• Smaller, domain-specific corpora can yield better results than larger, general corpora for specialized tasks.
• Acquiring more clinical text data is essential for further enhancing model performance and generalizability.