毕业生知识库

随着人工智能技术的快速发展，情感识别作为人工智能领域的一个重要分 支，对于改善人机交互体验、提升社交媒体分析能力以及促进医疗诊断等方面具 有重要意义。然而，情感识别研究面临着数据标注成本高，情感识别数据少的挑 战，这些问题限制了情感识别技术的广泛应用。为了降低情感识别的标注成本和 充分利用具有弱标注的情感数据，在弱监督条件下展开情感识别研究的重要性 也与日俱增。而部分标签学习是弱监督学习的一个重要分支，因此本文针对部分 标签情感识别展开了研究。现有的方法存在两个主要缺陷，首先，往往忽略了情 感识别任务中严重的标签不平衡性，其次，往往忽略了情感数据在标注过程中普 遍存在的噪声。为了实现更加鲁棒的部分标签情感识别，本文针对其中的类别不 平衡问题和噪声标签问题展开了研究。本文的主要贡献总结如下： • 提出了一种缓解类别不平衡问题的部分标签情感识别方法。情感数据固 有的分布不均问题导致情感识别数据集中类别不平衡现象的普遍存在。这一挑 战往往未受到传统部分标签学习方法的充分重视，从而为构建鲁棒的部分标签 情感识别系统带来了额外的困难。本文提出了一种基于伪标签正则化的策略以 应对部分标签情感识别中的类别不平衡问题。该方法包含两个相互迭代优化的 过程：估计情感类别比例和动态更新伪标签。同时，通过挑选每个情感类别中的 高置信度样本，以及对同一个情感样本进行强弱增广的预测一致性正则，进一步 提升模型的识别准确性。本研究在多个情感识别数据集上进行了实验验证，实验 结果表明，与当前最先进的部分标签学习方法相比，本研究提出的方法展现出了 更好的情感识别效果。这一研究表明了采用伪标签正则化策略能够降低对常见 类别的过拟合，提升模型的鲁棒性，增强了模型对惊讶等长尾类别的识别能力。 • 提出了一种提升噪声部分标签条件下情感识别性能的方法。情感数据的 本质模糊性和标注过程的复杂性不可避免地导致了情感识别数据中标签噪声的 产生。然而，传统的部分标签学习方法常常忽视这一问题而假设真实标签总包含 在提供的候选标签集合之内，这种做法在追求鲁棒性的部分标签情感识别任务 中显得尤为不足。针对这一问题，本文提出了一种在标签集合信息和模型预测结 果之间寻求平衡的伪标签方法。本研究的创新之处在于，首次将标签信息的动态 平衡应用于部分标签学习，通过调整候选标签集合和模型预测结果之间的权重， 有效提升了部分标签情感识别在噪声条件下的性能。在多个情感识别数据集上 的实验证明，本文所提方法不仅在噪声条件下超过基线方法，还能进一步提升现 有部分标签学习方法的性能。这一研究表明了采用动态调整候选标签重要性的 策略能有效防止对噪声标签的过拟合，从而提高了噪声部分标签情感识别的准确性。

With the rapid development of artificial intelligence technology, emotion recogni tion, as an important branch in the field of artificial intelligence, is of great significance for improving human-computer interaction experience, enhancing social media analy sis ability and promoting medical diagnosis. However, emotion recognition research is faced with the challenges of high cost of data annotation and little emotion recognition data, which limit the wider application of emotion recognition technology. In order to reduce the labeling cost of emotion recognition and make full use of emotion data with weak labeling, the importance of emotion recognition research under weak supervision is increasing day by day. Partial label learning is an important branch of weakly su pervised learning, so this paper studies partial label emotion recognition. The existing methods have two major shortcomings, one is that they tend to ignore the serious la bel imbalance in the emotion recognition task, and the other is that they tend to ignore the widespread noise in the emotion data labeling process. In order to achieve a more robust partial label emotion recognition, the class imbalance problem and noise label problem are studied in this dissertation. The main contributions of this dissertation are summarized as follows: • Proposes a partial label emotion recognition method to alleviate the cate gory imbalance problem. The unevenness of distribution of emotion data leads to the existence of unevenness of categories in emotion recognition database. This challenge is often underappreciated by traditional partial label learning methods, which brings additional difficulties to the construction of robust partial label emotion recognition systems. In this paper, inspired by the modeling of transportation problems, a strategy based on pseudo-label regularization is proposed to deal with the category imbalance in the emotion recognition of partial labels. The method includes two iterative optimiza tion processes of estimating the proportion of emotion categories and dynamically up dating pseudo-labels, and is supplemented by the selection of high confidence samples in each emotion category and the prediction consistency regularity of the same emotion sample, so as to further improve the model recognition accuracy and the ability to ex tract emotion-related features. Through experiments on multiple emotion recognition data sets, the experimental results show that the proposed method has a better emotion recognition effect than the most advanced partial label learning methods. This study shows that using the pseudo-label regularization strategy can effectively improve the robustness of the model, reduce the over fitting of common categories, and enhance the recognition ability of the model for long-tail categories such as surprise classes. • Proposes a method to improve the performance of emotion recognition under noisy partial label conditions. The essential fuzziness of emotion data and the complexity of labeling process inevitably lead to label noise in emotion recognition data. However, traditional partial label learning methods often ignore this problem and assume that the true label is always included in the provided set of candidate labels, which is particularly inadequate in the pursuit of robust partial label emotion recogni tion task. To solve this problem, a novel pseudo-labeling method is proposed to find a balance between the information of the label set and the prediction of the model itself. The innovation of this study is that the dynamic balance of label information is applied to partial label learning for the first time, and the performance of partial label emo tion recognition under noisy conditions is effectively improved by adjusting the weight between the candidate label set and the model prediction. Experiments on several senti ment recognition databases show that the proposed text method not only performs well under noisy conditions, but also improves the performance of some existing label learn ing methods. Thisstudyshowsthatthestrategyofdynamicallyadjustingtheimportance of candidate labels can effectively improve the robustness of the model, and effectively prevent over-fitting of noisy labels, thus significantly improving the accuracy of emo tion recognition of noisy labels. This result provides a new perspective for dealing with the noise partial label problem.