Bigger But Not Better: Small Neural Language Models Outperform Large Language Models in Detection of Thought Disorder