Huge amount of unstructured data is posted on the cloud from various sources for the purpose of feedback and reviews. These review needs require classification for many a reasons and sentiment classification is one of them. Sentiment classification of these reviews quite difficult as they are arriving from many sources. A robust classifier is needed to deal with different data distributions. Traditional supervised machine learning approaches not works well as they require retraining when domain is changed. Deep learning techniques perform well to handle these situations, but they are more data hungry and computationally expensive.

Transfer learning is a feature in the cross-domain sentiment classification where features are transferred from one domain to another without any training. Moreover, transfer learning allows the domains, tasks, and distributions used in training and testing to be different. Therefor transfer learning mechanism is required to transfer the sentiment features across the domains.

This paper presents a transfer learning approach using pretrained language model, ELMO which helps in transferring sentiment features across domains. This model has been tested on text reviews posted on twitter data set and compared with deep learning methods with and without pretraining process, also our model delivers promising results. This model permits flexibility to plug and play parameters with target models with easier domain adaptivity and transfer sentiment features. Also, model enables sentiment classifiers by using the transferred features from an already trained domain and reuse the sentiment features by saving the time and training cost.

来自各种来源的大量非结构化数据被发布到云中，以进行反馈和审查。这些审查需求出于多种原因需要分类，而情感分类就是其中之一。这些评论的情感分类非常困难，因为它们来自许多来源。需要一个强大的分类器来处理不同的数据分布。传统的监督式机器学习方法效果不佳，因为当领域发生变化时，它们需要重新培训。深度学习技术可以很好地处理这些情况，但是它们需要大量数据，而且计算量大。

转移学习是跨域情感分类中的一项功能，该功能无需任何培训即可将功能从一个域转移到另一个域。此外，转移学习允许在培训和测试中使用的域，任务和分布是不同的。因此，需要转移学习机制来跨域转移情感特征。

本文提出了一种使用预训练语言模型ELMO的迁移学习方法，该方法有助于跨领域转移情感特征。该模型已在Twitter数据集上发布的文本评论上进行了测试，并与有无预训练过程的深度学习方法进行了比较，我们的模型也提供了可喜的结果。该模型允许灵活地将参数插入和播放目标模型，并具有更容易的域适应性和传递情感特征。此外，模型通过使用已训练的域中传输的特征来启用情绪分类器，并通过节省时间和训练成本来重用情绪特征。