SubmissionNumber#=%=#11
FinalPaperTitle#=%=#Neural Nexus at PsyDefDetect: Fine-Tuning RoBERTa with Focal Loss and Role-Tagged Dialogue History for Defense Level Detection
ShortPaperTitle#=%=#
NumberOfPages#=%=#
CopyrightSigned#=%=#
JobTitle#==#
Organization#==#
Abstract#==#We present our system for the PsyDefDetect shared task at BioNLP 2026, which focuses on classifying help-seeker utterances into psychological defense mechanism levels using conversational context. Our approach fine-tunes RoBERTa with a composite objective combining focal loss, label smoothing, and dampened class weights to address severe class imbalance. We also use role-tagged dialogue history to capture context dependency. Our system achieves a macro-F1 of 0.2556 on the private test set ranking 12th out of 20.
Author{1}{Firstname}#=%=#Subhrajyoti
Author{1}{Lastname}#=%=#Basu
Author{1}{Username}#=%=#subh1234
Author{1}{Orcid}#=%=#
Author{1}{Email}#=%=#subhrajyoti479@gmail.com
Author{1}{Affiliation}#=%=#Heritage Institute of Technology

==========
èéáğö