[目的/意义] 大规模预训练模型的发展开启了人工智能领域的新阶段，越来越多的研究者将BERT、GPT等模型引入到各应用领域，但其复杂性和可解释性的问题给模型通用化带来一定程度的风险，与此同时，分类任务中训练数据的人工标注成本高，其质量直接影响模型的可用性，自动优化数据标注质量往往可以节约资源，提高模型的训练效率和模型的可用性。[方法/过程] 提出GSBERT（Gradient Saliency BERT Model）模型，该模型基于梯度显著度的方法探寻BERT模型决断依据，并通过可视化方式呈现BERT模型决断过程，使模型决策具有可解释性；在通过梯度显著度计算样本特征词的基础上，提取样本特征分布作为GSBERT检测标注的依据。为了检验GSBERT在分析数据一致性上的效果，利用GSBERT模型对新闻分类数据集的“误分类”样本进行重新标注，由于标注质量的提高，使模型在验证集上准确率提升12.55%。[结果/结论] 从大模型可解释性的角度出发，探究一种全新有效的挖掘数据价值的方法，为各领域更广泛地应用深度学习技术提供借鉴思路。
[Purpose/Significance] The development of large-scale pre-trained models has opened a new stage for artificial intelligence. More and more researchers have adopted models such as BERT and GPT in various application fields. However, the problems of complexity and interpretability have posed a degree of risk to the model’s generalization. Meanwhile, the manual annotation cost of training data in classification tasks is high, and its quality directly affects the usability of the model. Automated optimization of data annotation quality often leads to resource savings and enhances the training efficiency and usability of models. [Method/Process] This article proposed the GSBERT model. The model used gradient saliency to explore the decision-making basis of the BERT and presented the BERT model decision-making process through visualization, thereby making model decisions interpretable. Based on the calculation of feature words through gradient saliency, the distribution of sample features was extracted as the basis for GSBERT to detect annotations. To test the effectiveness of GSBERT in analyzing data consistency, the model was applied to re-annotate the “misclassified” samples from the news classification datasets. The improved annotation quality led to a 12.55% increase in accuracy on the validation set. [Result/Conclusion] This study explores a novel and effective method for mining the value of data from the perspective of large model interpretability, aiming to provide insights for broader applications of deep learning techniques across various fields.
[1] CAMBRIA E, MALANDRI L, MERCORIO F, et al. A survey on XAI and natural language explanations[J]. Information processing & management, 2023, 60(1): 103111.
[2] CHEN X, HE B, HUI K, et al. Dealing with textual noise for robust and effective BERT re-ranking[J]. Information processing & management, 2023, 60(1): 103135.
[3] 梁峥, 王宏志, 戴加佳, 等. 预训练语言模型实体匹配的可解释性[J]. 软件学报, 2023, 34(03): 1087-1108. (LIANG Z, WANG H Z, DAI J J, et al. Interpretability of entity matching based on pre-trained language model[J]. Journal of software, 2022, 34(3): 1087-1108.)
[4] ROGERS A, KOVALEVA O, RUMSHISKY A. A primer in BERTology: what we know about how BERT works[J]. Transactions of the association for computational linguistics, 2021, 8(1): 842-866.
[5] VIG J. Visualizing attention in transformer-based language representation models[J]. arxiv preprint, 2019, arxiv:1904.02679.
[6] POERNER N, ROTH B, SCHüTZE H. Interpretable textual neuron representations for NLP[J]. arxiv preprint, 2018, arxiv:1809.07291.
[7] OLAH C, SATYANARAYAN A, JOHNSON I, et al. The building blocks of interpretability[J]. Distill, 2018, 3(3): e10.
[8] VOITA E, SENNRICH R, TITOV I. The bottom-up evolution of representations in the transformer: a study with machine translation and language modeling objectives[J]. arxiv preprint, 2019, arxiv:1909.01380.
[9] NOSTALGEBRAIST. Interpreting GPT: the logit lens[EB/OL]. [2024-10-12]. https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens.
[10] DENIL M, DEMIRAJ A, DE FREITAS N. Extraction of salient sentences from labelled documents[J]. arxiv preprint, 2014, arxiv:1412.6815.
[11] SHRIKUMAR A, GREENSIDE P, SHCHERBINA A, et al. Not just a black box: learning important features through propagating activation differences[J]. arxiv preprint, 2016, arxiv:1605.01713.
[12] ATANASOVA P, SIMONSEN J G, LIOMA C, et al. A diagnostic study of explainability techniques for text classification[M]. Accountable and explainable methods for complex reasoning over text. Denmark: Springer Nature Switzerland, 2024.
[13] TORRES J M M, MEDINA-DEVILLIERS S, CLARKSON T, et al. Evaluation of interpretability for deep learning algorithms in EEG emotion recognition: a case study in autism[J]. Artificial intelligence in medicine, 2023, 143(9): 102545.
[14] ZHENG P, ZHOU H, LIU J, et al. Interpretable building energy consumption forecasting using spectral clustering algorithm and temporal fusion transformers architecture[J]. Applied energy, 2023, 349(21): 121607.
[15] HU Y, ALASKAR A A, ALTHOEY F, et al. Strength evaluation sustainable concrete with waste ingredients at elevated temperature by employing interpretable algorithms: optimization and hyper tuning[J]. Materials today communications, 2023, 36(3): 106467.
[16] WILDE H, KNIGHT V, GILLARD J. Evolutionary dataset optimization: learning algorithm quality through evolution[J]. Applied intelligence, 2020, 50(4): 1172-1191.
[17] GONG Y, LIU G, XUE Y, et al. A survey on dataset quality in machine learning[J]. Information and software technology, 2023, 162(10): 107268.
[18] KROMMYDA M, KANTERE V. Semantic analysis for conversational datasets: improving their quality using semantic relationships[J]. International journal of semantic computing, 2020, 14(3): 395-422.
[19] ANDREW N. MLOps: from model-centric to data-centric AI[EB/OL]. [2024-10-15]. https://www.deeplearning.ai/wpcontent/uploads/2021/06/MLOps-From-Model-centric-to-Data-centric-AI.PDF.
[20] 王诚文, 董青秀, 穗志方, 等. 自然语言处理评测数据集质量评估研究[J]. 中文信息学报, 2023, 37(2): 26-40. (WANG C W, DONG Q X, SUI Z F, et al. A Study on the quality evaluation of natural language processing evaluation datasets[J]. Journal of Chinese information processing, 2023, 37(2): 26-40.)
[21] 张涛, 马海群. 智能情报分析中数据与算法风险识别模型构建研究[J]. 情报学报, 2022, 41(8): 832-844. (ZHANG T, MA H Q. Research on the construction of data and algorithm risk identification model in intelligent intelligence analysis[J]. Journal of the China Society for Scientific and Technical Information, 2022, 41(8): 832-844.)
[22] DEVLIN J, CHANG M W, LEE K, et al. Bert: Pre-training of deep bidirectional transformers for language understanding[J]. arxiv preprint, 2018, arxiv:1810.04805.
[23] BASTINGS J, FILIPPOVA K. The elephant in the interpretability room: why use attention as explanation when we have saliency methods?[J]. arxiv preprint, 2020, arxiv:2010.05607.
[24] Clue benchmark[EB/OL]. [2024-10-15]. https://storage.googleapis.com/cluebenchmark/tasks/tnews_public.zip.
[25] MISRA R. News category dataset[J]. arxiv preprint, 2022, arxiv:2209.11429.
[26] ALAMMAR J. The illustrated BERT, ELMo, and co. (How NLP cracked transfer learning)[EB/OL]. [2024-10-11]. https://jalammar.github.io/illustrated-bert/.
