NCU-NLP at ROCLING-2021 Shared Task: Using MacBERT Transformers for Dimensional Sentiment Analysis

985

Authors

nan

Abstract

AbstractWe use the MacBERT transformers and fine-tune them to ROCLING-2021 shared tasks using the CVAT and CVAS data. We compare the performance of MacBERT with the other two transformers BERT and RoBERTa in the valence and arousal dimensions, respectively. MAE and correlation coefficient (r) were used as evaluation metrics. On ROCLING-2021 test set, our used MacBERT model achieves 0.611 of MAE and 0.904 of r in the valence dimensions; and 0.938 of MAE and 0.549 of r in the arousal dimension.

Keywords

nan

Full Content

� � �� ���� MacBERT ��� CVAT � CVAS � � � � � � � � � � ROCLING 2021 ��������� MacBERT � BERT � RoBERTa ��� ������� Valence �Arousal �� ���������������� (MAE) ����� (r) ������� ��������� Valence �� MAE � r ��� 0.611 � 0.904��� Arousal �� MAE � r ��� 0.938 � 0.549 � ��� Abstract We use the MacBERT transformers and fine-tune them to ROCLING-2021 shared tasks using the CVAT and CVAS data. We compare the performance of MacBERT with the other two transformers BERT and RoBERTa in the valence and arousal dimensions, respectively. MAE and correlation coefficient (r) were used as evaluation metrics. On ROCLING-2021 test set, our used MacBERT model achieves 0.611 of MAE and 0.904 of r in the valence dimensions; and 0.938 of MAE and 0.549 of r in the arousal dimension. ������������������ Keywords: affective computing, learning emotions, deep learning 1� �� ���� (affective computing) ������� ������������������� �������(sentiment analysis) ���� ������������������� ������������������� �������������������� ���������������� (Calvo and Kim, 2013)� ��������� (categorical) ����� ������������������� �������������� �� Ekman (1992) ������������ (anger)�� � (happiness)��� (fear)��� (sadness)� �� (disgust) ��� (surprise)������� ��� (Schouten and Frasincar, 2015; Pontiki et al., 2015)� ������������������ � � � � � � � � � � � � � � � � � (dimensional) �������������� ���������� (Russell, 1980)��1 � ��� Valance-Arousal ���������� ������� Valance � (�� V) ���� ���������� 1 � 9 ���� 1 �� ������ 9 �������� 5 ���� �������� Arousal � (�� A) ��� ������ 1 �����9 �����5 � ������������������� ������ VA ����������� � � � � � � � � � � � � � � � � � (CVAW) (Yu et al., 2016a) � VA ���� 6.8 � 7.2��� VA �������������� ������������ (V: 4.0, A: 2.8) �� ������������������� ��������� VA ��������� NCU-NLP at ROCLING-2021 Shared Task: Using MacBERT Transformers for Dimensional Sentiment Analysis ��� Man-Chen Hung, ��� Chao-Yi Chen ��� Pin-Jung Chen, ��� Lung-Hao Lee ������������ Department of Electrical Engineering National Central University {109521068, 107501543, 107303034}@ncu.edu.tw, lhlee@ee.ncu.edu.tw The 33rd Conference on Computational Linguistics and Speech Processing (ROCLING 2021) Taoyuan, Taiwan, October 15-16, 2021. The Association for Computational Linguistics and Chinese Language Processing 380 � � ������������������� ������� IALP 2016 �������������� �������� (Chinese words) (Yu et al; 2016b)�IJCNLP 2017 ���������� (Chinese phrases)������������ ���������� (Yu et al., 2017) � ROCLING 2021 ������������� ���(domain adaption)��������� �������������������� ������������������� ������ Valence (V) � Arousal (A) �� ������������������� 1 � 9 ���������� 1 ������� ������� V � 6.8 � A � 5.2 � �� Input: 1, ������������� ������������� �� Output: 1, 6.8, 5.2 ����� NCU-NLP �������� MacBERT ��������� (Cui et al., 2020)���� CVAT ��� (Yu et al., 2016b) ������� CVAS ���������� �MacBERT ����������Valence� MAE � r ���� 0.611 � 0.904�� Arousal � MAE � r ���� 0.938 � 0.549 � ������������������ ������������������� ������������������� ������� 2� ���� ������� VA����������� � � � : � � � (lexicon-based) � � � � (regression-based)��������(neural- network-based)� ������������������ ��������������� VA��� � � � � � � � � � � VA � � � � � (Paltoglou et al., …

Content truncated to 500 words

Labeled Text

AbstractWe use the MacBERT transformers and fine-tune them to ROCLING-2021 shared tasks using the CVAT and CVAS data. We compare the performance of MacBERT with the other two transformers BERT and RoBERTa in the valence and arousal dimensions, respectively. MAE and correlation coefficient (r) were used as evaluation metrics. On ROCLING-2021 test set, our used MacBERT model achieves 0.611 of MAE and 0.904 of r in the valence dimensions; and 0.938 of MAE and 0.549 of r in the arousal dimension.

Publication Details

Paper ID:
985
Year:
2021
Publication:
TIPSTER
Publication Type:
conf
DOI:
nan