Research On Multi-Task Sentiment Classification Algorithm Based On Hard Sharing Mechanism

Emotion classification is a common task,which is widely used in real life and has a great impact on people’s life and work.The data of this task often contains obvious emotion words or mood features.Therefore,the data of this task is often similar.Multi-task learning methods are good at extracting features from multiple similar task data to improve the effectiveness of each task.However,multi-task learning models often have the problem of insufficient feature extraction ability and mutual interference between multiple tasks.Therefore,this dissertation proposes two algorithms for transforming interference in the shared feature space to address this problem,which can effectively enhance the representation of the shared feature space.The main research contents of this dissertation are as follows:(1)To address the problem of insufficient feature extraction ability of the shared layer in the hard sharing model,this dissertation proposes a multi-task learning sentiment classification algorithm based on task recognition training.The algorithm first utilizes the multilayer Transformer Encoder model to constitute the shared layers.And the knowledge in the Bidirectional Encoder Representations from Transformers(BERT)model is used to enrich the features in the shared layer.Then a task recognition training method is proposed to transform the interference between multiple tasks into more discriminative features.It improves the recognition ability of the model for different tasks.Finally,we conduct experiments on two multi-task sentiment classification datasets.The experimental results show that the method proposed in this dissertation effectively improves the performance of the model in multi-task sentiment classification.(2)In this dissertation,we propose a multi-task learning sentiment classification algorithm based on multilayer residual connectivity Bi-directional Long Short-Term Memory(Bi-LSTM).The algorithm solves the problems brought by using BERT model,such as the increase in the number of model parameters,the increase in computation,the longer running time,and the lack of feature extraction capability of the multilayer Transformer Encoder model.Firstly,the attention mechanism is used to focus on the important words in the input sentences.Then a multi-level residual connected Bi-LSTM neural network is built to extract multi-level sequence features in the input sentences.We merge them with the features extracted by the attention mechanism as the output of the shared layer.Finally,a simple classifier is used to achieve better results.In this dissertation,experiments are conducted on a multi-task multi-sentiment category dataset.The experimental results show that the method works better on multi-sentiment category data than other multi-task learning methods.The number of layers of the Bi-LSTM neural network in the shared layer can be adjusted for different datasets to facilitate the extraction of sequence features at different layers.The stronger feature extraction capability allows the model to achieve better recognition results even with a fully connected layers classifier.