Text Emotion Classification in Mental Health Counseling Conversations Using Soft Voting Ensemble of BERT and BiLSTM

DOI

10.2991/978-94-6239-828-3_10How to use a DOI?

Keywords

text emotion classification; mental health counseling; BERT; BiL- STM; soft voting ensemble; cross-dataset evaluation; pseudo-labeling

Abstract

Text-based mental health counseling is increasingly adopted due to its accessibility and anonymity; however, the absence of nonverbal cues makes emotion recognition challenging, particularly under high conversation volume and class imbalance. This study evaluates BERT, BiLSTM with GloVe embeddings, and their soft-voting ensemble for text emotion classification in counseling conversations, using only the client/help-seeker side of the Amod and MentalChat datasets. A semi-supervised pseudo-labeling approach (pre-trained emotion model, confidence threshold 0.7) was used, with low-confidence samples manually verified by a single non-expert annotator; this single-annotator setup, without inter-annotator agreement, is treated as a limitation. After cleaning and deduplication, 6,941 labeled samples across seven emotion classes were obtained. BERT was fully fine-tuned; BiLSTM was trained with a class-weighted loss to address imbalance. Simple and macro-F1-weighted soft voting were both evaluated. BERT consistently outperforms BiLSTM, achieving its best result on MentalChat (accuracy 0.804, macro F1 0.643). The ensembles do not consistently sur- pass BERT, reflecting the performance gap between base learners, though they substantially outperform standalone BiLSTM. Cross-dataset evaluation shows BERT has the most moderate degradation (≈12.3%). Per-class analysis shows the minority disgust class was correctly recognized in only 1 of 16 model–scenario combinations (BERT, Amod-trained and Amod-tested; recall 14.3%), and 0% elsewhere, highlighting the limits of class weighting under severe imbalance. These results suggest BERT is the more promising candidate among the models compared here for downstream emotion-support tools, though clinical deployment would require expert-validated labels and prospective clinical evaluation beyond this study’s scope.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Sabrina Ella Zafira
AU  - Fitra Abdurrachman Bachtiar
AU  - Fajar Pradana
PY  - 2026
DA  - 2026/11/04
TI  - Text Emotion Classification in Mental Health Counseling Conversations Using Soft Voting Ensemble of BERT and BiLSTM
BT  - Proceedings of the 2026 11th International Conference on Sustainable Information Engineering and Technology (SIET 2026)
PB  - Atlantis Press
SP  - 107
EP  - 117
SN  - 2589-4900
UR  - https://doi.org/10.2991/978-94-6239-828-3_10
DO  - 10.2991/978-94-6239-828-3_10
ID  - Zafira2026
ER  -