Abstract
Multi-domain sentiment classification aims to mitigate poor performance models due to the scarcity of labeled data in a single domain, by utilizing data labeled from various domains. A series of models that jointly train domain classifiers and sentiment classifiers have demonstrated their advantages, because domain classification helps generate necessary information for sentiment classification. Intuitively, the importance of sentiment classification tasks is the same in all domains for multi-domain sentiment classification; but domain classification tasks are different because the impact of domain information on sentiment classification varies across different fields; this can be controlled through adjustable weights or hyper parameters. However, as the number of domains increases, existing hyperparameter optimization algorithms may face the following challenges: (1) tremendous demand for computing resources, (2) convergence problems, and (3) high algorithm complexity. To efficiently generate the domain information required for sentiment classification in each domain, we propose a dynamic information modulation algorithm. Specifically, the model training process is divided into two stages. In the first stage, a shared hyperparameter, which would control the proportion of domain classification tasks across all fields, is determined. In the second stage, we introduce a novel domain-aware modulation algorithm to adjust the domain information contained in the input text, which is then calculated based on a gradient-based and loss-based method. In summary, experimental results on a public sentiment analysis dataset containing 16 domains prove the superiority of the proposed method.
Abstract (translated)
跨领域情感分类旨在通过利用来自不同领域的标记数据，来缓解单一领域中由于标注数据稀缺而导致的模型性能不佳问题。一系列同时训练领域分类器和情感分类器的模型已经展示了它们的优势，因为领域分类有助于生成情感分类所需的信息。直观来看，在多领域情感分类任务中，所有领域的感情分类任务的重要性是相同的；然而，不同领域的领域分类任务则是不同的，因为领域信息对情感分类的影响在各个领域间各不相同；这可以通过可调权重或超参数来控制。但是，随着领域的数量增加，现有的超参数优化算法可能会面临以下挑战：（1）计算资源的巨大需求、（2）收敛问题以及（3）算法复杂度高。 为了高效生成每个领域中情感分类所需的信息，我们提出了一种动态信息调节算法。具体来说，模型训练过程分为两个阶段。在第一阶段，确定一个共享的超参数，该超参数将控制跨所有领域的领域分类任务的比例。第二阶段引入一种新的基于领域的调节算法来调整输入文本中的领域信息，然后根据梯度和损失方法进行计算。 总之，在包含16个领域的公共情感分析数据集上的实验结果证明了所提出方法的优越性。
URL
https://arxiv.org/abs/2505.06630
