Semantic Similarity Calculation based on Adaptive Semi-supervised Method

Semantic similarity calculation is a key problem in natural language processing tasks. Supervised language models have been widely used in semantic similarity calculation. However, when a small part of the labeled data is used for training, the accuracy of the supervised model will be greatly reduced. An important issue is how to effectively use labeled and unlabeled data to build models. Thus, we propose an adaptive semi-supervised semantic similarity calculation method that combines the learning content of the supervised model Siamese LSTM with the topic information of the unsupervised LDA topic model through an attention layer. Inspired by active learning, we apply an adaptive mechanism to actively optimize model parameters. We apply this method to standard semantic datasets for ablation experiments. The experiment results demonstrate that the proposed method considerably outperforms other state-of-the-art methods.