Research On Sarcasm Recognition Based On Transfer Learning

Sarcasm is sophisticated linguistic expression and is commonly observed on social media and e-commerce platform.Sarcasm intent is mainly manifested in the incongruity between the context and the statement.By analyzing sarcastic comments in online texts,it has potential economic and social value for understanding users’personal emotional tendencies and public opinion issues.Traditional approaches rely heavily on discrete handcrafted features and will incur enormous human costs.Existing studies mostly formulate irony detection as a standard supervised learning text categorization task,However,supervised learning often requires a large amount of labeled data to train a robust classifier,and there are practical difficulties in collecting and labeling a large amount of data.Due to the complexity of sarcasm language expression,pattern recognition for sarcasm detection is also quite complex and difficult.Based on this,this paper introduces transfer learning and proposes two deep neural network models based on transfer learning for sarcasm detection to alleviate the problem of insufficient datasets for sarcasm detection tasks.The main work and innovations of this paper are as follows:(1)A transfer learning model based on self-matching network is proposed.Using the method of transfer learning,it solves the problem of insufficient target task data by learning the required resource knowledge in similar fields.Then a new self-matching network is combined to construct a self-matching attention vector by calculating the joint information in each word pair of the input sentence,which is used to capture the incongruent information of the sentence to improve the detection effect of the target sarcasm dataset.Finally,the effect of the migration of different network layers of the deep neural network on the model performance is investigated through ablation experiments.(2)A transfer learning model based on adversarial network is proposed.The BERT model is used to extract the semantic information representation of the text,and the LSTM network is used to extract the sequence feature representation of the text.Then,the feature confusion is carried out through the domain adversarial network,so that the model has more similarity when extracting the shared features of different domains,and realizes the knowledge transfer between the source domain and the target domain.Domain adversarial networks will also encourage BERT to fool the domain discriminator to generate domain-invariant features.Finally,the impact on model performance is explored through domain adaptation parameter analysis and different deep neural networks.The Experimental results on publicly available datasets show that the two transfer learning-based deep neural network models proposed in this paper are significantly better than the existing baseline models on standard evaluation metrics.It can alleviate the problem that the data of sarcasm detection task is difficult to obtain,and improve the accuracy of sarcasm detection.