Automated Construction of Lexicons to Improve Depression Screening With Text Messages

Area of Science:

• Computational linguistics
• Mental health informatics
• Machine learning applications

Background:

• Depression is a widespread mental illness requiring effective diagnostic tools.
• Current text message-based depression screening often uses formal lexicons, limiting performance due to the colloquial nature of messages.
• Need for unobtrusive and accurate methods for mental health assessment.

Purpose of the Study:

• To automatically construct alternative, informal lexicons for improved depression screening via text messages.
• To evaluate the effectiveness of these novel lexicons in machine learning classification models.
• To enhance the accuracy and relevance of features extracted from colloquial text data.

Main Methods:

• Generated 36 lexicons from diverse corpora (fiction, forums, news) to capture colloquial language.
• Extracted lexical category features from text messages using the constructed lexicons.
• Employed machine learning models to compare the depression screening performance of different lexicons against formal ones and bag-of-words.

Main Results:

• 14 out of 36 automatically constructed lexicons significantly outperformed the pre-existing formal lexicon and basic bag-of-words approach.
• The best-performing informal lexicon increased average F1 scores by 10% compared to the formal lexicon.
• Demonstrated the superiority of less formal lexicons for depression screening in text message data.

Conclusions:

• Informal, automatically constructed lexicons significantly enhance the performance of machine learning models for depression screening via text messages.
• This strategy overcomes limitations of formal lexicons in analyzing colloquial communication.
• The developed lexicons provide valuable resources for future research in computational mental health and text analysis.