基于深度学习的手语翻译：过去、现状与未来

打开文本图片集

中图分类号：TP183 文献标志码：A 文章编号：1001-3695（2025）08-001-2241-14

doi：10.19734/j.issn.1001-3695.2025.01.0001

Deep learning-based sign language translation： past， present， and future

Zhang Lei 1a，1b，1c ，Wang Zhenyula，1b，1c，Lian Shuaishuaile2，Pu Bingqian 1a，1b，1c ， Liu Yutao 1a，1b，1c ， Qin Mingzhe 1c，2† （1.a.SchoolffoocoreKboofoostelice&for cesing，c.ibellotsiUsi longjiang ，China；2.HandanPolytechnic Collge，Handan Hebei O56046，China）

Abstract：SLTbasedondeplearning aims totranslatesign language gestures into natural language using depleaming techniques to improve translation’saccuracySLTreducescommunication barrers between normal hearing individualsandthose withhearing impaiments.However，SLTfaces numerouschallngesdue tothelackofstandardizationacross diferentsiglanguages andthe mismatch between sign language gestures and spoken language sentence structures.With thedevelopmentof deeplearning technologies，SLThas gained widespreadatentionfromresearchers.This paper summarized recent approaches on SLTbasedondeeplearning and classifiedtheminto four categories accrding tomodel structureanddevelopment history： linearstructure-basedSLT，encoder-decoderarchitecture-basedSLT，largemodelfine-tuning-basedSLT，andcontrastive learning-based SLT.Byanalyzingthecharacteristicsandperformanceofthese methods，thisstudyprovidedacomprehensive evaluationof theprogressinSLTmethods.Finaly，thepaperoutlined futureresearchdirections，focusingonthepotentialand developmenttrendsofkeytechnologies，cludingrea-timetranslation，onrastivelearing-basedST，andlargemodelfintuning-based SLT.

Key words：deep learning；sign language translation（SLT）；machine translation;contrastive learning;large language models；encoder-decoder

0 引言

根据世界卫生组织统计，全球约有4.3亿人患有听力障碍[1]。（剩余48036字）