A semi-supervised short text sentiment classification method based on improved Bert model from unlabelled data

Abstract Short text information has considerable commercial value and immeasurable social value. Natural language processing short sentiment analysis technology can organize analyze on the Internet. tasks such as classification have achieved satisfactory performance under a supervised learning framework. However, traditional relies large-scale high-quality manual labels obtaining label data costs lot. Therefore, strong dependence hinders application of deep model to large extent, which is bottleneck learning. At same time, datasets product reviews an imbalance in distribution samples. To solve above problems, this paper proposes method predict according semi-supervised mode implements MixMatchNL enhancement method. Meanwhile, Bert pre-training updated. The cross-entropy loss function improved Focal Loss alleviate datasets. Experimental results based public indicate proposed accuracy recognition compared with previous update other state-of-the-art models.