Efficient Hyperdimensional Computing Accelerator Design for Chinese Text Classification Task
-
摘要: 随着物联网、智能穿戴及离线终端等边缘计算场景的蓬勃发展，在本地设备上进行低延迟、高隐私的中文文本分析已成为智能交互与数据处理的核心需求。尽管基于神经网络、Transformer的语言模型展现出卓越的准确率，但其庞大的参数规模与计算开销，使其难以在功耗与存储极度受限的边缘设备上落地。超维计算(HDC)作为一种新兴的类脑计算范式，通过2K-10K维度的超向量表征文本，以轻量级的“编码-查询”机制替代了复杂的多层网络结构，为边缘侧文本任务提供了硬件友好型的技术路径。然而，现有HDC研究主要面向由少量字母构成的拼音文字系统如英文。对于拥有庞大字符集的中文，现有方法会因基超向量数量的剧增而导致存储开销显著上升，这直接破坏了其在边缘侧部署的轻量化优势。为突破中文文本在边缘侧的高效表征瓶颈，该文基于汉字独特的书写与结构特征，提出了一种高效的汉字超维编码方法，并构建了专为边缘侧中文文本分类设计的超维计算框架HDChinese。此外，为进一步适配边缘硬件环境，该文还设计了与HDChinese适配的加速器架构。实验结果表明，在维持分类准确率相当的前提下，HDChinese及其硬件加速器能够将模型复杂度大幅缩减99%以上，同时实现4.91%-77.36%的训练时间缩短以及6.67%-47.25%的推理延迟降低。该研究有效兼顾了准确性与极低开销，为边缘设备上的中文文本分析任务提供了理想的硬件解决方案。Abstract:
Objective With the proliferation of edge computing in IoT, smart wearables, and offline terminals, low-latency and high-privacy Chinese text analysis on local devices has become a core requirement. While neural network-based and Transformer-based language models achieve excellent accuracy, their massive parameter scales and computational demands make them unsuitable for power- and storage-constrained edge devices. Hyperdimensional Computing (HDC), an emerging brain-inspired paradigm, represents text through 2K-10K dimensional hypervectors and employs lightweight encoding-query mechanisms instead of complex multi-layer networks, offering a hardware-friendly approach for edge-side text tasks. However, existing HDC research has focused exclusively on alphabetic writing systems such as English, where a limited set of letters suffices for N-gram encoding. For Chinese, with its thousands of commonly used characters, directly applying existing methods would cause prohibitive storage overhead from the explosion of base hypervectors, fundamentally undermining the lightweight advantage of HDC. This paper aims to overcome this bottleneck by proposing a novel character encoding method and a dedicated HDC framework with hardware acceleration for Chinese text classification. Methods Drawing inspiration from the structural characteristics of Chinese characters as ideographic symbols, this paper proposes a hyperdimensional encoding method based on character glyph structure. Specifically, the method reverse-engineers the encoding process of the Wubi input method, decomposing each Chinese character into its constituent strokes and radicals and mapping them to equivalent letter sequences for efficient hyperdimensional encoding. A retrieval library covering 3,500 commonly used Chinese characters, as defined in the General Standard Chinese Characters Table issued by the Ministry of Education of the People's Republic of China, is constructed. For each character, the Wubi library is queried to obtain a letter sequence of length 3 or 4, where each letter position corresponds to a base hypervector. Through cyclic shift and binding operations applied to the base hypervectors of consecutive letters, the character is encoded into a character-level hypervector that simultaneously captures both letter identity and sequential order information, effectively preventing confusion arising from different letter permutations. This approach reduces the storage overhead associated with base hypervectors by more than 95% compared to the method of directly assigning independent hypervectors to each Chinese character ( Table 1 ). Building upon this character encoding scheme, the HDChinese framework is developed to support both training and inference procedures. During inference, the sentence-level hypervector, obtained by bundling all character hypervectors, queries the class hypervectors via cosine similarity, which is efficiently realized through XNOR-based inner product computations. During training, class hypervectors are derived by bundling sentence hypervectors belonging to the same category and applying a sign-based binarization operation. A subsequent retraining phase iteratively fine-tunes the non-binary class hypervectors using misclassified samples with a learning rate of 0.25, thereby mitigating the influence of outlier hypervectors on the cluster center. Furthermore, a dedicated hardware accelerator architecture tailored to HDChinese is designed. The architecture comprises three main modules: an encoding module, a querying module, and a class hypervector update module, which can be selectively activated to support inference-only, training, and retraining modes respectively (Fig. 2 ). The accelerator is prototyped on an Ultra96v2 FPGA development board featuring a Xilinx ZYNQ SoC.Results and Discussions Four open-source datasets covering both binary and multi-class tasks are used ( Table 2 ). The Wubi-based encoding outperforms the Pinyin-based method in accuracy while saving 31.46% storage (Table 4 ), and ignoring rare characters (average frequency below 0.5%) has negligible accuracy impact (Table 4 ). A dimension of 2K is adopted for hypervectors as higher dimensions yield no significant gains (Fig. 3 ). FPGA measurements show HDChinese achieves a model size of approximately 16 KB—over 99% smaller than kNN, SVM, and Random Forest (Table 6 ). Inference time is reduced by 6.67%-47.25% and total training time by 4.91%-77.36% compared to SVM and Random Forest, while accuracy after retraining reaches comparable levels (Fig. 5 ,Table 6 ). Additionally, the KB-scale model enables operation without external memory chips, unlike comparison algorithms requiring over 400 MB of runtime memory. At 100 MHz, power consumption is 0.278 W, representing the optimal throughput-power trade-off (Table 5 ).Conclusions This paper proposes a hyperdimensional encoding method based on Chinese character glyph structure, resolving the incompatibility between existing HDC methods and Chinese text. The HDChinese framework and its dedicated hardware accelerator achieve over 99% model complexity reduction while maintaining competitive accuracy, with training time reductions of 4.91%-77.36% and inference time reductions of 6.67%-47.25%. The proposed approach effectively balances accuracy with extremely low overhead, providing an ideal hardware solution for Chinese text analysis on edge devices. -
表 1 汉字超维编码方法对比
编码方法 检索库空间占用(拼音+顺序；五笔) 基超向量空间占用 总空间占用 超向量直接表征的方法 0 3500 ×2000 bit=7 Mbit7Mbit 基于拼音的方法 3500 ×(6×5 bit)+3500 ×6 bit=126 Kbit26× 2000 bit=52 Kbit178 Kbit 基于五笔的方法 3500 ×(4×5 bit)=70 Kbit26× 2000 bit=52 Kbit122 Kbit 
下载: 导出CSV
表 2 评估所用数据集
数据集名称 类别数 内容简介 waimai_10k* 2 某外卖平台收集的用户评价，正向 4000 条，负向8000 条weibo_100k* 2 带情感标注新浪微博评论文本，正负向评论约各5万条 online_shopping* 2 共6万余条电商评论数据，正、负向评论各约3万条 TelecomFraud[22] 5 包含4种电信诈骗类型的文本，各约 1000 条，以及1类非诈骗的文本约8000 条*开源地址： https://github.com/SophonPlus/ChineseNlpCorpus
下载: 导出CSV
表 3 硬件参数
硬件描述 主频 LUT FlipFlop BRAM DSP 功耗 CPU 4核Cortex-A53 CPU 1.2GHz - - - - 1.71W FPGA Ultrascale+ FPGA 100MHz 25060 27877 56.5 0 0.278W
下载: 导出CSV
表 4 不同超维编码方法准确率对比
waimai_10k weibo_100k online_shopping TelecomFraud 基于拼音的方法 83.82% 90.27% 80.1% 93.5% 基于五笔的方法 84.86% 91.4% 80.2% 94.18% 基于五笔的方法* 84.73%
(0.38%)90.76%
(1.12%)79.92%
(0.21%)94.14%
(0.2%)*保留对生僻字( 3500 个常用汉字以外的字)的超维表征，括号中是生僻字在数据集中出现的频率
下载: 导出CSV
表 5 跨域泛化准确率对比
waimai_10k weibo_100k online_shopping 在waimai_10k上训练 84.86% 64.91% 71.67% 在weibo_100k上训练 65.01% 91.4% 58.63% 在online_shopping
上训练71.35% 59.58% 80.2%
下载: 导出CSV
表 6 加速器在不同主频下的实测性能
主频 waimai_10k weibo_100k online_shopping TelecomFraud 训练
吞吐率*推理
吞吐率*功率 训练
吞吐率*推理
吞吐率*功率 训练
吞吐率*推理
吞吐率*功率 训练
吞吐率*推理
吞吐率*功率 50 MHz 2350.24 2321.39 0.191 W 1370.3 1374.63 0.21 W 1221.13 1196.27 0.219 W 639.87 627.74 0.219 W 100 MHz 2523.42 2497.62 0.25 W 1991.08 1975.14 0.278 W 1835.72 1827.66 0.289 W 1212.1 1201.53 0.294 W 150 MHz 2497.14 2462.01 0.35 W 1993.73 1972.71 0.381 W 1897.14 1832.58 0.385 W 1389.09 1394.07 0.393 W *吞吐率指标SPS：Samples Per Second，每秒处理样本数量
下载: 导出CSV
表 7 性能对比
waimai_10k weibo_100k online_shopping TelecomFraud 训练
时间推理
时间模型
大小训练
时间推理
时间模型
大小训练
时间推理
时间模型
大小训练
时间推理
时间模型
大小kNN 1.52 s 24 s 43.97 MB 内存溢出 内存溢出 3.93 s 24.27 s 43.18 MB SVM 184.71 s 69.05 s 19.4 MB 运行时间超过1小时 运行时间超过1小时 75.55 s 34.56 s 9.5 MB Random Forest 68.29 s 1.82 s 28.96 MB 1284.28 s 19.96 s 102.3M B 583.52 s 10.94 s 118.2 MB 59.44 s 2.1s 11.96 MB HDChinese 3.8 s 0.96 s 15.38 KB 48.21 s 12.15 s 15.38 KB 27.35 s 6.87 s 15.38 KB 7.77 s 1.96 s 16.11 KB HDChinese* 22.8 s 0.96 s 15.38 KB 290.71 s 12.15 s 15.38 KB 164.1 s 6.87 s 15.38 KB 46.52 s 1.96 s 16.11 KB *在训练完成后加入了5轮的重训练，因此总训练时间(包括1轮训练和5轮重训练)更长，但不影响其他性能
下载: 导出CSV
-
[1] ZHAO Jinghua, SHU Xi, ZHAO Xiaohua, et al. Online opinion trend prediction for public health events based on time series transformer[J]. IEEE Transactions on Computational Social Systems, 2026, 13(1): 1159–1168. doi: 10.1109/TCSS.2025.3617491. [2] 朱振方, 李嘉欣, 徐富永, 等. 一种融合情感和策略信息的共情对话生成方法[J]. 电子与信息学报, 2024, 46(8): 3382–3389. doi: 10.11999/JEIT231417.ZHU Zhenfang, LI Jiaxin, XU Fuyong, et al. Empathetic dialogue generation via sentiment and support strategy[J]. Journal of Electronics & Information Technology, 2024, 46(8): 3382–3389. doi: 10.11999/JEIT231417. [3] LIU Fengyu. Research on user dynamic preference modeling and recommendation system based on text sentiment transfer[C]. 2025 2nd International Conference on Intelligent Computing and Robotics (ICICR), Dalian, China, 2025: 1088–1092. doi: 10.1109/ICICR65456.2025.00192. [4] LIU Pengfei, QIU Xipeng, and HUANG Xuanjing. Recurrent neural network for text classification with multi-task learning[C]. Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, New York, USA, 2016: 2873–2879. [5] ZHENG Yafeng, GAO Zhanghao, SHEN Jun, et al. Optimizing automatic text classification approach in adaptive online collaborative discussion-a perspective of attention mechanism-based bi-LSTM[J]. IEEE Transactions on Learning Technologies, 2023, 16(5): 591–602. doi: 10.1109/tlt.2022.3192116. [6] JIANG Ting, WANG Deqing, SUN Leilei, et al. LightXML: Transformer with dynamic negative sampling for high-performance extreme multi-label text classification[C]. Proceedings of the 35th AAAI Conference on Artificial Intelligence, 2021: 7987–7994. doi: 10.1609/aaai.v35i9.16974.(查阅网上资料,未找到本条文献出版地信息,请确认). [7] 王泽昊, 朱振华, 谢童欣, 等. 混合专家大语言模型的系统与架构优化技术综述[J]. 电子与信息学报, 2025, 47(11): 4055–4078. doi: 10.11999/JEIT250407.WANG Zehao, ZHU Zhenhua, XIE Tongxin, et al. A survey on system and architecture optimization techniques for mixture-of-experts large language models[J]. Journal of Electronics & Information Technology, 2025, 47(11): 4055–4078. doi: 10.11999/JEIT250407. [8] 刘文波, 姚翼荣, 张弓, 等. 超维计算概念、应用及研究进展[J]. 系统工程与电子技术, 2023, 45(7): 1938–1956. doi: 10.12305/j.issn.1001-506X.2023.07.04.LIU Wenbo, YAO Yirong, ZHANG Gong, et al. Concept, application, and research progress of hyperdimensional computing[J]. Systems Engineering and Electronics, 2023, 45(7): 1938–1956. doi: 10.12305/j.issn.1001-506X.2023.07.04. [9] 余天洋, 吴比, 陈珂, 等. 超维计算硬件设计: 进展、趋势和展望[J]. 集成电路与嵌入式系统, 2025, 25(8): 1–9. doi: 10.20193/j.ices2097-4191.2025.0047.YU Tianyang, WU Bi, CHEN Ke, et al. Hyperdimensional computing hardware: Progress, trends and prospects[J]. Integrated Circuits and Embedded Systems, 2025, 25(8): 1–9. doi: 10.20193/j.ices2097-4191.2025.0047. [10] RAHIMI A, KANERVA P, and RABAEY J M. A robust and energy-efficient classifier using brain-inspired hyperdimensional computing[C]. Proceedings of the 2016 International Symposium on Low Power Electronics and Design, San Francisco Airport, USA, 2016: 64–69. doi: 10.1145/2934583.2934624. [11] THAPA R, LAMICHHANE B, MA Dongning, et al. SpamHD: Memory-efficient text spam detection using brain-inspired hyperdimensional computing[C]. 2021 IEEE Computer Society Annual Symposium on VLSI (ISVLSI), Tampa, USA, 2021: 84–89. doi: 10.1109/isvlsi51109.2021.00026. [12] ALONSO P, SHRIDHAR K, KLEYKO D, et al. HyperEmbed: Tradeoffs between resources and performance in NLP tasks with hyperdimensional computing enabled embedding of n-gram statistics[C]. 2021 International Joint Conference on Neural Networks (IJCNN), Shenzhen, China, 2021: 1–9. doi: 10.1109/IJCNN52387.2021.9534359. [13] PARK S, KIM K, SEO J, et al. Unsupervised extractive dialogue summarization in hyperdimensional space[C]. ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Seoul, Korea, 2024: 12411–12415. doi: 10.1109/ICASSP48485.2024.10446698. [14] MIKOLOV T, CHEN Kai, CORRADO G, et al. Efficient estimation of word representations in vector space[C]. 1st International Conference on Learning Representations, Scottsdale, USA, 2013. [15] PENNINGTON J, SOCHER R, and MANNING C D. GloVe: Global vectors for word representation[C]. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), Doha, Qatar, 2014: 1532–1543. doi: 10.3115/v1/D14-1162. [16] DEVLIN J, CHANG Mingwei, LEE K, 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: Human Language Technologies, Volume 1 (long and short papers), Minneapolis, USA, 2019: 4171–4186. doi: 10.18653/v1/N19-1423. [17] LIU Caizhi, SEHNG Yanxiu, WEI Zhiqiang, et al. Research of text classification based on improved TF-IDF algorithm[C]. IEEE International Conference of Intelligent Robotic and Control Engineering (IRCE), Lanzhou, China, 2018: 218–222. doi: 10.1109/IRCE.2018.8492945. [18] YU Tianyang, WU Bi, CHEN Ke, et al. Fully learnable hyperdimensional computing framework with ultratiny accelerator for edge-side applications[J]. IEEE Transactions on Computers, 2024, 73(2): 574–585. doi: 10.1109/tc.2023.3337316. [19] GE Lulu and PARHI K K. Classification using hyperdimensional computing: A review[J]. IEEE Circuits and Systems Magazine, 2020, 20(2): 30–47. doi: 10.1109/mcas.2020.2988388. [20] GORBAN A N and TYUKIN I Y. Blessing of dimensionality: Mathematical foundations of the statistical physics of data[J]. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 2018, 376(2118): 20170237. doi: 10.1098/rsta.2017.0237. [21] SHIN S H, YOO Y, MOON Y, et al. STAR-PIM: Self-test and repair structure for processing-in-memory with adder tree-based MAC[J]. IEEE Transactions on Circuits and Systems I: Regular Papers, 2026, 73(3): 1990–2003. doi: 10.1109/tcsi.2025.3615631. [22] LI Jun, ZHANG Cheng, and JIANG Lanlan. Innovative telecom fraud detection: A new dataset and an advanced model with RoBERTa and dual loss functions[J]. Applied Sciences, 2024, 14(24): 11628. doi: 10.3390/app142411628. [23] 王杰, 王鹃. 面向RISC-V平台的安全高效固件可信平台模块设计与实现[J]. 电子与信息学报, 2025, 47(7): 2385–2395. doi: 10.11999/JEIT241112.WANG Jie and WANG Juan. The design and implementation of a secure and efficient firmware trusted platform module for RISC-V platforms[J]. Journal of Electronics & Information Technology, 2025, 47(7): 2385–2395. doi: 10.11999/JEIT241112. -
下载:
图(4) / 表(7)
计量
- 文章访问数: 72
- HTML全文浏览量: 20
- PDF下载量: 4
- 被引次数: 0
