School of Computer & Information Engineering, Jiangxi Normal University, Nanchang 330022, China
Show less
文章历史+
Received
Accepted
Published
2024-11-29
2025-03-08
Issue Date
2025-10-30
PDF (1836K)
摘要
表情符号(emoji)作为一种有效的情感表达工具,已被广泛用于基于互联网的日常交流。目前,大多数emoji预测模型采用词嵌入方法构建emoji表示,但并未考虑emoji与情绪的直接关联和emoji在目标数据集中的共现模式,导致学习到的emoji表示的情绪区分度不足。针对上述问题,该文基于与情绪直接关联的emoji情感分布构建emoji表示,提出一种融合情感分布的多标签emoji预测模型(Emotion distribution Information Fusion for multi-label Emoji Prediction, EIFEP)。EIFEP模型的结构由三个模块组成:文本语义模块、emoji信息模块和信息融合预测模块。文本语义模块采用基于转换器的双向编码器表示(Bidirectional Encoder Representations from Transformers,BERT)模型提取文本语义表示;emoji信息模块使用多层图卷积网络构建融合emoji情感分布和共现模式信息的emoji嵌入表示;信息融合预测模块利用emoji嵌入注意力机制将文本语义表示与emoji嵌入表示矩阵融合,并进行emoji预测。在Mu-Emoji英文数据集上进行的对比实验表明,EIFEP模型在准确率(Accuracy,Acc)、微平均F1分数(Micro-F1 score,micF1)、杰卡德系数(Jaccard Score,JS)和汉明损失( Hamming Loss,HL)四个评价指标上均优于现有的多标签emoji预测模型,分别达到60.88%、58.79%、74.05%和6.92%。
Abstract
Emoji, as an effective tool for emotional expression, has been widely used in daily internet-based communication. Currently, most emoji prediction models employ word embedding methods to construct emoji representations, but consider neither the direct association between emoji and emotions, nor the co-occurrence patterns of emoji in the target dataset, resulting in insufficient emotional discriminability of the learned emoji representations. To address these issues, this paper proposes a multi-label emoji prediction model based on the emotional distribution of emojis directly related to emotion, which is named EIFEP (i.e., Emotion distribution Information Fusion for multi-label Emoji Prediction). The EIFEP model consists of three modules: a semantic information module, an emoji information module, and an information fusion prediction module. The semantic information module uses BERT to extract text semantic representation; the emoji information module applies a multi-layer graph convolutional network to construct emoji embedding that integrate emotion distribution and co-occurrence pattern information; the information fusion prediction module uses emoji embedding attention mechanism to integrate the text semantic and the emoji embedding matrix for emoji prediction. The comparative experiments on the Mu-Emoji English dataset demonstrate that the EIFEP model outperformers the existing multi-label emoji prediction model, achieving 60.88% in Acc, 58.79% in JS, 74.05% in micF1, and 6.92% in HL, respectively.
表情符号(emoji)作为一种视觉化的符号语言,在短文本信息的情感表达和视觉增强方面具有重要作用,目前已被广泛应用于社交媒体平台的日常交流中[1]。与纯文本相比,emoji能够有效地辅助理解语义和情感的细微差别。Emoji预测任务的目标是针对给定文本预测与其最相关的emoji,在情感分析、情绪识别、反讽检测等自然语言处理(Natural Language Processing,NLP)任务中具有广泛的应用价值[2]。
针对上述问题,本文采用与情绪直接关联的emoji情感分布构建emoji表示,提出了一种融合情感分布的多标签emoji预测(Emotion distribution Information Fusion for multi-label Emoji Prediction, EIFEP)模型。基于预训练的emoji情感分布表示,EIFEP模型采用在目标数据集中的emoji共现模式计算emoji的相关性,再利用多层图卷积网络(Graph Convolutional Network, GCN)将emoji情感分布增强为包含共现模式信息的emoji嵌入表示。EIFEP模型的结构由三个模块组成:文本语义模块、emoji信息模块和信息融合预测模块。文本语义模块采用基于转换器的双向编码器表示(Bidirectional Encoder Representations from Transformers,BERT)模型提取文本语义表示;emoji信息模块使用多层GCN构建融合emoji情感分布和共现模式信息的emoji嵌入表示;信息融合预测模块利用emoji嵌入注意力机制将文本语义表示与emoji嵌入表示矩阵融合,并进行emoji预测。在Mu-Emoji英文数据集上进行的对比实验结果表明,本文提出的EIFEP模型相较于现有的多标签emoji预测模型具有更优的性能。本文的主要贡献概括如下:
情感分布学习(Emotion Distribution Learning, EDL)[12]通过构建情感分布来定量描述情绪模糊性,对于处理文本情绪识别和情感分析任务,尤其是涉及情绪模糊性和复杂性的场景下具有明显的优势。在表达复杂情感方面,emoji所呈现的情绪与人脸表情具有相似性,均可被视为由多种基本情绪以不同强度组合而成的情感分布。观察图1发现, 不仅表达joy这一种情绪,还同时蕴含了optimism和love等多种其他情绪。
本文基于与情绪直接关联的emoji情感分布表示,提出了一种融合情感分布的多标签emoji预测(Emotion distribution Information Fusion for multi-label Emoji Prediction, EIFEP)模型。EIFEP模型将包含情感分布先验知识和目标数据集emoji相关性信息的emoji表示与文本语义信息融合,用于多标签emoji预测。模型由文本语义模块、emoji信息模块和信息融合预测模块组成。具体框架如图2所示。
在数据预处理部分,本文将句子中的网页链接、邮箱地址等替换成“〈url〉”、“〈email〉”的形式;去除换行符和停用词;对hashtag进行展开和标注,例如将“#ilikedogs”转换成“〈hashtag〉 I like dogs 〈/hashtag〉”;对延长单词进行恢复和标注,例如将“suuuucks”转换成“sucks 〈elongated〉”;对重复的标点符号丢弃并标注,例如将“!!!”标注成“!〈repeated〉”;对颜表情进行替换成对应的情绪单词,例如将“(^.^)”转换成“〈happy〉”;对所有英文单词进行词性还原、展开缩写并转换为小写的操作;最后丢弃tokens少于5的样本。经过数据预处理后,推文从583 048条减少到526 993条,减少了56 055条。
FRANCESCOB, LUISE A, HORACIOS. Revealing Patterns of Twitter Emoji Usage in Barcelona and Madrid[M]//Artificial Intelligence Research and Development. Amsterdam, Netherlands: IOS Press, 2016: 239-244. DOI:10.3233/978-1-61499-696-5-239 .
[2]
FELBOB, MISLOVEA, SØGAARDA, et al. Using Millions of Emoji Occurrences to Learn Any-domainrepresentations for Detecting Sentiment, Emotion and Sarcasm[C]//Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Stroudsburg, PA, USA: ACL, 2017: 1615-1625. DOI:10.18653/v1/d17-1169 .
[3]
YANGZ P, YEX, XUH T. TGCN-bert Emoji Prediction in Information Systems Using TCN and GCN Fusing Features Based on BERT[J]. Int J Semant Web Inf Syst, 2023, 19(1): 1-16. DOI:10.4018/ijswis.331082 .
[4]
MILLERH, KLUVERD, THEBAULT-SPIEKERJ, et al. Understanding Emoji Ambiguity in Context: The Role of Text in Emoji-related Miscommunication[C]//Proceedings of the International AAAI Conference on Web and Social Media. Menlo Park, California, USA: AAAI, 2017: 152-161. DOI:10.1609/icwsm.v11i1.14901 .
[5]
CHENY X, YUANJ B, YOUQ Z, et al. Twitter Sentiment Analysis via Bi-sense Emoji Embedding and Attention-based LSTM[C]//Proceedings of the 26th ACM International Conference on Multimedia. New York: ACM, 2018: 117-125. DOI:10.1145/3240508.3240533 .
[6]
EISNERB, ROCKTÄSCHELT, AUGENSTEINI, et al. Emoji2vec: Learning Emoji Representations from Their Description[C]//Proceedings of The Fourth International Workshop on Natural Language Processing for Social Media. Stroudsburg, PA, USA: ACL, 2016: 48-54. DOI:10.18653/v1/w16-6208 .
[7]
WIJERATNES, BALASURIYAL, SHETHA, et al. A Semantics-based Measure of Emoji Similarity[C]//Proceedings of the International Conference on Web Intelligence. New York: ACM, 2017: 646-653. DOI:10.1145/3106426.3106490 .
[8]
BARBIERIF, BALLESTEROSM, SAGGIONH. Are Emojis Predictable? [C]//Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics. Stroudsburg, PA, USA: ACL, 2017: 105-111. DOI:10.18653/v1/e17-2017 .
[9]
PENGD, ZHAOH. Seq2Emoji: A Hybrid Sequence Generation Model for Short Text Emoji Prediction[J]. Knowl Based Syst, 2021, 214: 106727. DOI:10.1016/j.knosys.2020.106727 .
[10]
LEES, JEONGD, PARKE. MultiEmo: Multi-task Framework for Emoji Prediction[J]. Knowl Based Syst, 2022, 242: 108437. DOI:10.1016/j.knosys.2022.108437 .
[11]
SINGHA, BLANCOE, JINW. Incorporating Emoji Descriptions Improves Tweet Classification[C]//2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg, PA, USA: ACL, 2019: 2096-2101. DOI:10.18653/v1/n19-1214 .
[12]
ZHOUY, XUEH, GENGX. Emotion Distribution Recognition from Facial Expressions[C]//Proceedings of the 23rd ACM International Conference on Multimedia. New York: ACM, 2015: 1247-1250. DOI:10.1145/2733373.2806328 .
[13]
XIONGH T, LIUH F, ZHONGB N, et al. Structured and Sparse Annotations for Image Emotion Distribution Learning[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Menlo Park, California, USA:AAAI, 2019: 363-370. DOI:10.1609/aaai.v33i01.3301363 .
LAIJ S, WANZ Y, ZENGX Q. The Image Emotion Distribution Learning Based on Emotion Wheel and Multi-task Convolutional Neural Network[J]. J Jiangxi Norm Univ Nat Sci Ed, 2022, 46(4): 363-371. DOI: 10.16357/j.cnki.issn1000-5862.2022.04.06 .
[16]
ZHANGB Q, ESSLG, MOWER PROVOSTE. Predicting the Distribution of Emotion Perception: Capturing Inter-rater Variability[C]//Proceedings of the 19th ACM International Conference on Multimodal Interaction. New York: ACM, 2017: 51-59. DOI:10.1145/3136755.3136792 .
GOYALP, FERRARAE. Graph Embedding Techniques, Applications, and Performance: a Survey[J]. Knowl Based Syst, 2018, 151: 78-94. DOI:10.1016/j.knosys.2018.03.022 .
[21]
CHENZ M, WEIX S, WANGP, et al. Multi-label Image Recognition with Graph Convolutional Networks[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE, 2019: 5177-5186. DOI:10.1109/cvpr.2019.00532 .
[22]
SCHOENEA M, TURNERA P, DETHLEFSN. Bidirectional Dilated LSTM with Attention for Fine-grained Emotion Classification in Tweets[J]. Affcon@AAAI, 2020, 2614: 100-117. DOI: 10.1007/s00521-022-08186-1 .
[23]
VUH T, NGUYENM T, NGUYENV C, et al. Label-representative Graph Convolutional Network for Multi-label Text Classification[J]. Appl Intell, 2023, 53(12): 14759-14774. DOI:10.1007/s10489-022-04106-x .
[24]
SINGHG V, FIRDAUSM, EKBALA, et al. Unity in Diversity: Multilabel Emoji Identification in Tweets[J]. IEEE Trans Comput Soc Syst, 2023, 10(3): 1029-1038. DOI:10.1109/TCSS.2022.3162865 .
[25]
DEVLINJ, CHANGM W, LEEK, et al. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding[C]//Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics. Stroudsburg PA, USA: ACL, 2019: 4171-4186. DOI: 10.18653/v1/N19-1423 .
[26]
CHIP H, CHUNGP H, WUT H, et al. Audio Albert: a Lite Bert for Self-supervised Learning of Audio Representation[C]//2021 IEEE Spoken Language Technology Workshop (SLT). New York: IEEE, 2021: 344-350. DOI:10.1109/SLT48900.2021.9383575 .