Refining Semantic Similarity of Paraphasias Using a Contextual Language Model.

Area of Science:

• Computational linguistics
• Neuroscience
• Speech-language pathology

Background:

• Aphasia, a language disorder, often involves anomia, characterized by word-finding difficulties.
• Paraphasias, or naming errors, provide crucial insights into anomia.
• Automated classification of paraphasias aids in understanding and assessing these deficits.

Purpose of the Study:

• To evaluate the performance of a modern language model, BERT, as a semantic classifier for paraphasias.
• To compare BERT's efficacy against the existing word2vec model within the ParAlg software.
• To identify challenges in classifying semantic similarity for aphasic errors.

Main Methods:

• Utilized a dataset of 11,999 paraphasias from the Philadelphia Naming Test.
• Trained the ParAlg software using both word2vec and BERT models.
• Compared model performance against human classifications and analyzed error types.

Main Results:

• BERT demonstrated a significant reduction in semantic classification errors, nearly halving them compared to word2vec.
• BERT showed improved handling of polysemy and reduced word-sense issues.
• Semantic ambiguity and errors in classifying associated or category-coordinate responses were common.

Conclusions:

• BERT significantly outperforms word2vec as a semantic classifier for paraphasias.
• BERT's enhanced performance, particularly with polysemy, advances automated assessment of anomia.
• This validates ParAlg with BERT as a more accurate tool for characterizing word-finding deficits.