情感分布学习（emotion distribution learning， EDL）采用情感分布记录给定样本在各个情绪上的表达程度， 在处理具有模糊性的多标签情绪分析任务时具有明显优势。情感分布标签增强技术将已标注的情绪单标签增强为情感分布， 可以解决EDL缺乏已标注情感分布的实验数据集的问题。然而， 已有的情感分布标签增强方法采用离散空间情绪模型表示情绪， 存在情绪间的相关信息丢失和情绪表达不连续等问题。针对上述问题， 该文引入基于连续维度的效价-唤醒-支配（valence-arousal-dominance， VAD）心理学情绪模型， 提出融合VAD情绪知识的文本情感分布标签增强方法（VAD emotion knowledge-based text emotion distribution label enhancement， VADLE）。VADLE方法基于先验的VAD情绪模型中的情绪距离， 先为英文句子的真实情绪标签和句中情感词的情绪标签分别生成先验情感分布， 再通过分布叠加将2种先验情感分布统一。通过英文单标签文本情感数据集的对比实验表明：VADLE方法在情绪预测任务方面的性能优于已有的情感分布标签增强方法。
Abstract
[Objective] Existing emotion distribution label enhancement (EDLE) methods construct the emotion distribution based on a discrete spatial emotion model; hence, expressing the correlation between emotions in a granular manner with continuous values is challenging. Therefore, herein, a valence-arousal-dominance(VAD) emotion knowledge-based text emotion distribution label enhancement (VADLE) method is proposed based on the VAD continuous-dimensional psychology emotion model. Unlike existing EDLE methods, VADLE uses VAD emotion knowledge in a three-dimensional continuous space to model emotion correlations and generate a more nuanced emotion distribution. The VADLE method comprises several steps: (1) Extraction of emotion word information via referencing lexicon and extracting emotion words from a given sentence. (2) Generation of priori emotion distributions for emotion labels using a local linear-weighting algorithm. The algorithm measures the effect of secondary emotion on the primary emotion based on the VAD emotional spatial distance and assigns weights to nearby emotions using a Gaussian kernel. (3) Construction of sentence-level emotion distribution by combining the prior emotion distributions of sentence and textual emotion words. Furthermore, this study uses joint loss to train a multitask emotion distribution learning model based on the robustly optimized bidirectional encoder representations from transformers pretraining approach (RoBERTa) pretrained language model. This approach simultaneously optimizes the prediction of emotion distribution and classification. The sentence text features extracted using the RoBERTa pretrained model are then passed through a fully connected layer to generate a probability distribution over all emotion labels. Based on this probability distribution, the model utilizes the Kullback-Leibler (KL) loss for measuring the distance between the predicted and actual distributions, optimizing the emotion distribution prediction task. Simultaneously, cross-entropy loss is employed for optimizing the emotion recognition task. To evaluate the performance of the proposed VADLE method, extensive comparative experiments is performed on several single-label English datasets using four baseline EDLE methods: emotion wheel and lexicon-based emotion distribution label enhancement (EWLLE), lexicon-based emotion distribution label enhancement (LLE), Mikels emotion wheel-based emotion distribution label enhancement (MWLE), and One-Hot. Moreover, this study explores the effect of the bandwidth parameter (τ) in the local linear-weighting algorithm on the balance between the primary and secondary emotions in the generated emotion distribution. The performance of the model's emotion prediction was assessed using four classification evaluation metrics (Precision, Recall,F1-score, and Accuracy) and four emotion distribution prediction metrics (Canberra, Chebyshev, Cosine, and Intersection). The experimental results demonstrated that the VADLE method was superior to the baseline methods. Specifically, the VADLE method achieved superior performance on the emotion classification task over the EWLLE, LLE, and MWLE methods across all four indicators. The VADLE method also exhibited excellent performance for the emotion distribution prediction task. For instance, on the Cosine metric, the VADLE method outperformed the suboptimal EWLLE method by 2.6 % and exhibited considerable improvements over the LLE, MWLE, and One-Hot methods. The results showed that the optimal balance could be achieved by setting τ to 0.6, enabling the highest level of performance in the emotion distribution generation. Unlike existing EDLE methods, the VADLE method employs a fine-grained approach to studying emotions. It combines the prior emotion knowledge in the VAD continuous space with the linguistic information inherent in the sentiment words for generating more reasonable emotion distributions. Experimental results reveal that the VADLE method outperforms existing methods in terms of enhancing the emotion distribution labels in emotion prediction tasks.
关键词
情感分布标签增强 /
情感分布学习 /
VAD情绪空间 /
情感词典
Key words
emotion distribution label enhancement /
emotion distribution learning /
VAD emotion space /
affective lexicon
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] GARRITZMANN J L, NEIMANNS E, BUSEMEYER M R. Public opinion towards welfare state reform:The role of political trust and government satisfaction[J]. European Journal of Political Research, 2023, 62(1):197-220.
[2] LIU H, CAI Y C, LIN Z R, et al. Variational latent-state GPT for semi-supervised task-oriented dialog systems[J]. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023, 31:970-984.
[3] VATANI N, RAHMANI A M, JAVADI H H S. Personality-based and trust-aware products recommendation in social networks[J]. Applied Intelligence, 2023, 53(1):879-903.
[4] YU J F, MARUJO L, JIANG J, et al. Improving multi-label emotion classification via sentiment classification with dual attention transfer network[C]//Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Brussels, Belgium:Association for Computational Linguistics, 2018:1097-1102.
[5] GAO B B, XING C, XIE C W, et al. Deep label distribution learning with label ambiguity[J]. IEEE Transactions on Image Processing, 2017, 26(6):2825-2838.
[6] ZHOU Y, XUE H, GENG X. Emotion distribution recognition from facial expressions[C]//Proceedings of the 23rd ACM International Conference on Multimedia. Brisbane, Australia:ACM, 2015:1247-1250.
[7] GENG X. Label distribution learning[J]. IEEE Transactions on Knowledge and Data Engineering, 2016, 28(7):1734-1748.
[8] WU H Y, HUANG Y G, NAN G S. Doubled coupling for image emotion distribution learning[J]. Knowledge-Based Systems, 2023, 260:110107.
[9] QIN X R, CHEN Y F, RAO Y H, et al. A constrained optimization approach for cross-domain emotion distribution learning[J]. Knowledge-Based Systems, 2021, 227:107160.
[10] JIA X Y, LI Z C, ZHENG X, et al. Label distribution learning with label correlations on local samples[J]. IEEE Transactions on Knowledge and Data Engineering, 2021, 33(4):1619-1631.
[11] YANG J F, SHE D Y, SUN M. Joint image emotion classification and distribution learning via deep convolutional neural network[C]//Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence. Melbourne, Australia:AAAI Press, 2017:3266-3272.
[12] ZHANG Y X, FU J M, SHE D Y, et al. Text emotion distribution learning via multi-task convolutional neural network[C]//Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence. Stockholm, Sweden:AAAI Press, 2018:4595-4601.
[13] 曾雪强, 华鑫, 刘平生, 等. 基于情感轮和情感词典的文本情感分布标记增强方法[J]. 计算机学报, 2021, 44(6):1080-1094. ZENG X Q, HUA X, LIU P S, et al. Emotion wheel and lexicon based text emotion distribution label enhancement method[J]. Chinese Journal of Computers, 2021, 44(6):1080-1094. (in Chinese)
[14] STRAPPARAVA C, MIHALCEA R. SemEval-2007 task 14:Affective text[C]//Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007). Prague, Czech Republic:Association for Computational Linguistics, 2007:70-74.
[15] ZHENG Q H, ZHU J H, TANG H Y, et al. Generalized label enhancement with sample correlations[J]. IEEE Transactions on Knowledge and Data Engineering, 2023, 35(1):482-495.
[16] MIKELS J A, FREDRICKSON B L, LARKIN G R, et al. Emotional category data on images from the international affective picture system[J]. Behavior Research Methods, 2005, 37(4):626-630.
[17] LE N, NGUYEN K, TRAN Q, et al. Uncertainty-aware label distribution learning for facial expression recognition[C]//2003 IEEE/CVF Winter Conference on Applications of Computer Vision. Waikoloa, USA:IEEE, 2023:6088-6097.
[18] XU N, LIU Y P, GENG X. Label enhancement for label distribution learning[J]. IEEE Transactions on Knowledge and Data Engineering, 2019, 33(4):1632-1643.
[19] TENG Z Y, VO D T, ZHANG Y. Context-sensitive lexicon features for neural sentiment analysis[C]//Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. Austin, USA:Association for Computational Linguistics, 2016:1629-1638.
[20] LI Z X, CHEN X H, XIE H R, et al. EmoChannel-SA:Exploring emotional dependency towards classification task with self-attention mechanism[J]. World Wide Web, 2021, 24(6):2049-2070.
[21] 赖金水, 万中英, 曾雪强. 基于情感轮和多任务卷积神经网络的图像情感分布学习[J]. 江西师范大学学报(自然科学版), 2022, 46(4):363-371. LAI J S, WAN Z Y, ZENG X Q. The image emotion distribution learning based on emotion wheel and multi-task convolutional neural network[J]. Journal of Jiangxi Normal University (Natural Science Edition), 2022, 46(4):363-371. (in Chinese)
[22] PLUTCHIK R. A psychoevolutionary theory of emotion[J]. Social Science Information, 1982, 21(4-5):529-553.
[23] TROIANO E, OBERLÄNDER L, KLINGER R. Dimensional modeling of emotions in text with appraisal theories:Corpus creation, annotation reliability, and prediction[J]. Computational Linguistics, 2023, 49(1):1-72.
[24] RUSSELL J A, MEHRABIAN A. Evidence for a three-factor theory of emotions[J]. Journal of Research in Personality, 1977, 11(3):273-294.
[25] EKMAN P. An argument for basic emotions[J]. Cognition and Emotion, 1992, 6(3-4):169-200.
[26] SURESH V, ONG D C. Using knowledge-embedded attention to augment pre-trained language models for fine-grained emotion recognition[C]//2021 9th International Conference on Affective Computing and Intelligent Interaction. Nara, Japan:IEEE, 2021:1-8.
[27] MOHAMMAD S M. Emotional tweets[C]//Proceedings of the 1st Joint Conference on Lexical and Computational Semantics-Volume 1:Proceedings of the main conference and the shared task, and Volume 2:Proceedings of the Sixth International Workshop on Semantic Evaluation. Montréal, Canada:Association for Computational Linguistics, 2012:246-255.
[28] ALM C O, SPROAT R. Emotional sequencing and development in fairy tales[M]//TAO J H, TAN T N, PICARD R W. Proceedings of the 1st International Conference on Affective Computing and Intelligent Interaction. Beijing, China:Springer, 2005:668-674.
[29] GHOLIPOUR SHAHRAKI A. Emotion mining from text[M]. Edmonton:University of Alberta, 2015.
[30] SCHERER K R, WALLBOTT H G. Evidence for universality and cultural variation of differential emotion response patterning[J]. Journal of Personality and Social Psychology, 1994, 66(2):310-328.
[31] MOHAMMAD S M, TURNEY P D. NRC emotion lexicon[R]. Ottawa:National Research Council of Canada, 2013.
[32] PORIA S, GELBUKH A, CAMBRIA E, et al. EmoSentic- Space:A novel framework for affective common-sense reasoning[J]. Knowledge-Based Systems, 2014, 69:108-123.
[33] LIU Y H, OTT M, GOYAL N, et al. RoBERTa:A robustly optimized BERT pretraining approach. arXiv. (2019-07-26). https://arxiv.org/abs/1907.11692.
基金
国家自然科学基金地区科学基金项目（62266021）；江西省教育厅科学技术研究项目（GJJ2200330）
PDF(5349 KB)
