Pre-trained language models have demonstrated excellent performance in downstream natural language understanding tasks by learning general grammar and semantic features from large-scale unlabeled text data. However, most existing pre-trained language models and cross-lingual models have yet to cover minority languages such as Tibetan, resulting in limited performance in processing these languages. In recent years, contrastive learning methods, such as SimCSE, have optimized data encoding by constructing positive and negative sample pairs. However, the simple division of positive and negative pairs may erroneously classify semantically similar positive pairs as negative pairs. To address this issue, this paper proposes a BERT-based label repairing alternating learning network. This network alternately trains two identical BERT models to encode Tibetan sentences and optimizes the division of positive and negative samples based on the encoding results, thereby obtaining more discriminative feature representations. Experimental evaluations on the Tibetan text similarity matching dataset demonstrate the cleaning ability of Bert-Label-Repair Alternating Learning for similar and negative pairs. Furthermore, experiments on two downstream understanding tasks,i.e., text classification and text similarity matching, validate the method's superiority. Finally, Ablation studies also confirm the effectiveness of each module, collectively highlighting the comprehensive advantages of the proposed approach.
自然语言理解是一项使计算机理解和解析人类语言的技术,主要包括命名实体识别、文本分类、文本相似度匹配等任务。传统卷积神经网络(Convolutional Neural Network,CNN)等方法只能设计具有有限泛化能力的特定任务模型,而预训练语言模型(Pre-trained Language Models,PLMs)则在一定程度上缓解了网络泛化能力弱的问题[1]。PLMs在大量无监督数据上进行训练,提取通用语义特征后,通过在下游任务数据上进行参数微调,使得在各种下游自然语言理解任务中取得更好的泛化能力和更高效的数据利用能力。
然而,现有PLMs的应用主要集中在英语、中文等使用广泛且数据资源丰富的语言上。对于藏语等资源相对有限的语言,尚缺乏性能优良的专用优化模型。为弥补这一不足,跨语言模型应运而生。如,Google推出的多语言预训练模型(Multilingual BERT,mBERT)[2],其能够有效增强PLMs的跨语言泛化能力;FAIR团队提出的跨语言模型XLM(Cross-lingual Language Model)[3],则通过联合训练不同语言,能够捕捉更多的跨语言信息,使低资源语言能够受益于从其他语言迁移而来的知识。然而,上述多语言模型并未包含藏语和国内其他少数民族语言。由哈尔滨工业大学讯飞联合实验室(HFL)开发的面向中国少数民族语言的多语言预训练模型(Chinese mINOrity pre-trained language model,CINO)[4],涵盖了藏语等少数民族语言和国内部分方言。已有研究表明,CINO在少数民族语言任务上表现优于通用多语言模型,但在某些特定下游任务中,其性能仍有待提升。
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: Human Language Technologies, 2019: 4171-4186.
[2]
PiresT, SchlingerE, GarretteD. How multilingual is multilingual BERT?[C]. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Florence, Italy: Association for Computational Linguistics, 2019:4996-5001.
[3]
AlexisC, KartikayK, NamanG, et al. Cross-lingual language model pretraining[C]. Proceedings of the 33rd International Conference on Neural Information Processing Systems,2019:7059-7069.
[4]
YangZ Q, XuZ H, CuiY M,et al. CINO: A Chinese minority pre-trained language model[C]. Proceedings of the 29th International Conference on Computational Linguistics,2022:3937-3949.
[5]
LiB H, ZhouH, HeJ X,et al. On the sentence embeddings from pre-trained language models[C]. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP),2020:9119-9130.
[6]
ZhangJ, KazhuoD, GadengL,et al. Research and application of Tibetan pre-training language model based on BERT[C]. Proceedings of the 2022 2nd International Conference on Control and Intelligent Robotics,2022: 519-524.
[7]
LiuS, DengJ, SunY, et al.TiBERT: Tibetan pre-trained language model[C].2022 IEEE International Conference on Systems, Man, and Cybernetics(SMC).IEEE,2022:2956-2961.
DengJ, ChenL, ZhangY,et al. TiKEM: Knowledge enhanced Tibetan pre-trained language mode[C]. Proceedings of the 22nd Chinese National Conference on Computational Linguistics,2023:135-144.
[10]
SechaJ, CizhenJ, CairangJ,et al. Automatic generation of Tibetan poems based on pre-training and control code method[C]. Proceedings of the 21st Chinese National Conference on Computational Linguistics,2022:366-373.
[11]
HuangS, YanX, OuYangX,et al. Abstractive summarization of Tibetan based on end-to-end pre-trained model [C]. Proceedings of the 22nd Chinese National Conference on Computational Linguistics,2023:113-123.
[12]
GaoT Y, YaoX C, ChenD Q.SimCSE: Simple contrastive learning of sentence embeddings[C]. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,2021:6894-6910.
[13]
ZhangD J, LiS W, XiaoW,et al. Pairwise supervised contrastive learning of sentence representations[C]. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,2021:5786-5798.
[14]
WangT, IsolaP. Understanding contrastive representation learning through alignment and uniformity on the hypersphere[C]. Proceedings of the International Conference on Machine Learning, 2020:9929-9939.
[15]
WuX, GaoC C, ZangL J,et al. ESimCSE: Enhanced sample building method for contrastive learning of unsupervised sentence embedding[C].Proceedings of the 29th International Conference on Computational Linguistics,2021:3898-3907.
SunY, LiuS, ChenC,et al.Construction of high-quality tibetan dataset for machine reading comprehension [C]. Proceedings of the 20th Chinese National Conference on Computational Linguistics,2021:208-218.
VaswaniA, ShazeerN, ParmarN,et al. Attention is all you need[J]. Advances in Neural Information Processing Systems, 2017:5998-6008.
[20]
JohnsonR, ZhangT. Deep pyramid convolutional neural networks for text categorization[C].Association for Computational Linguistics,2017:562-570.
[21]
LaiS, XuL, LiuK,et al. Recurrent convolutional neural networks for text classification[C]. Proceedings of the Twenty-ninth AAAI Conference on Artificial Intelligence,2015:2267-2273.
[22]
SuJ, CaoJ, LiuW,et al. Whitening sentence representations for better semantics and faster retrieval[J]. arXiv preprint arXiv:2021.