藏文句向量预训练模型在嵌入式系统中的应用研究  

Research on the Application of Tibetan Sentence Vector Pre-trained Model in Embedded Systems

作　　者：吕皓 吕慧 雍宾宾 多拉 李妍[1] 周庆国[1] 周睿[1] L Hao;L Hui;YONG Binbin;DUO La;LI Yan;ZHOU Qingguo;ZHOU Rui(School of Information Science and Engineering,Lanzhou University,Lanzhou 730000,China;The State Key Laboratory of Tibetan Intelligent Information Processing and Application,Qinghai Normal University,Xining 810008,China)

机构地区：[1]兰州大学信息科学与工程学院,兰州730000 [2]青海师范大学省部共建藏语智能信息处理及应用国家重点实验室,西宁810008

出　　处：《小型微型计算机系统》2025年第1期8-15,共8页Journal of Chinese Computer Systems

基　　金：国家重点研发计划项目(2020YFC0832500)资助;国家自然科学基金项目(61402210)资助;国家自然基金项目(62266037)资助。

摘　　要：本文研究了将藏文句向量预训练模型部署到嵌入式系统上进行推理和测试的问题.在机器学习中,对文本进行编码和表征存在困难,因此句向量技术成为自然语言处理领域的重要研究方向.然而,在藏文自然语言处理领域,句向量研究相对较少.为此,本文分析了藏文领域已有的预训练模型和句向量表示方法,并设计了一种改进的无监督SimCSE方法(Improved Simple Contrastive Learning of Unsupervised Sentence Embeddings,I-SimCSE).实验结果显示,使用I-SimCSE方法得到的藏文句向量模型性能优于其他方法.同时,本文探讨了边缘计算与预训练模型相结合的应用,并讨论了预训练语言模型在嵌入式系统上的潜在应用场景.最后,本文将I-SimCSE句向量模型部署在嵌入式设备Jetson TX1上,并测试了其平均单次推理时间,结果表明在嵌入式系统上部署预训练语言模型进行推理是可行的.综上所述,本文的研究对于藏文句向量预训练模型在嵌入式系统上的应用研究提供了有益的参考,并为未来藏文大模型在嵌入式系统的发展提供了指导和启示.This paper investigates the problem of deploying Tibetan sentence vector pre-trained models to embedded systems for inference and testing.In machine learning,encoding and representation of text pose challenges,leading to the prominence of sentence vector techniques in the field of natural language processing(NLP).However,research on sentence vectors in the Tibetan NLP domain is relatively limited.Therefore,this paper analyzes the existing pre-trained models and sentence vector representation methods in the Tibetan language field,and designs an improved simple contrastive learning of unsupervised sentence embeddings(I-SimCSE)method.The experimental results indicate that the performance of Tibetan sentence vectors model trained using the I-SimCSE method outperforms other methods.Furthermore,this paper explores the integration of edge computing with pre-trained models and discusses the potential application scenarios of pre-trained language models on embedded systems.Finally,the I-SimCSE sentence vector model is deployed on the Jetson TX1 embedded device,and its average single inference time is tested,demonstrating the feasibility of deploying pre-trained language models for inference on embedded systems.In conclusion,this research provides a useful reference for the application of Tibetan sentence vector pre-trained models on embedded systems,and offer guidance and insights for the future development of Tibetan large model on embedded systems.

关 键 词：藏文 句向量表示 嵌入式系统 预训练模型 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]