基于圖像知識(shí)增強(qiáng)的中文多模態(tài)反諷檢測(cè)方法

打開文本圖片集

Chinese multimodal irony detection method based on image knowledgeenhancement

LI Yueying，CAO Hui， ZHANG Jisai， XIA Xiaotian （KeyLaboratoryofLinguisticandCulturalComputingMinistryofEducation，InstituteofChineseEthnicInformationTechnology， Northwest Minzu University，Lanzhou 73Oo3O，China）

Abstract：With therapid development of social mediaandonlinecontent，theuseof ironyhasbecomecommonin the onlinecommunicationandinformationdisemination.However，thetraditionaltextanalysismethodsoftenfailtocapturethe meaningofironyaccurately，andrelyingsolelyontextualinformationhaslimitationsandisofinstability.Inthispaper，a Chinesemultimodal ironydatasetisconstructed.Thedatasetincludes5964annotateddatasamples，includingtwomodesoftext andimage.Theimagesplayanimportantroleinmultimodalironydetectiontasks.Inordertofullexplorethehidden information inimages，theimagecaptioninggenerationmodelViT-GPT-image-captioningisusedtogeneratethedescription informationoftheimageforimageknowledgeenhancement，soastoenhancetheunderstandingandcognitionoftheimage. Moreover，amultimodalatentionnetwork modelCMANetthatintegratesmodal informationforironydetectionisproposedtoget ridof theinsuffcientinformationcorelationbetweenmodesandlackofdataintheprocessofmulti-modaldatafusion. Experimental verification was performed on the dataset. The results show that the F1 -score of the proposed CMANet model has been improved by 1.49% and its accuracyby 1.89% in comparison with those of the baseline model.

Keywords：multimodal; irony detection; attention mechanism; cross-modal; deep learning; network fusion

0 引言

反諷是一種特殊的情感表達(dá)技巧，讓人難以直接理解表達(dá)者的真實(shí)意圖，達(dá)到委婉而含蓄的表達(dá)目的。（剩余11168字）