Research On Aspect-level Sentiment Analysis Algorithm Based On Multi-task Learning

Sentiment analysis,as one of the important research fields of natural language processing,has received more and more attention as a more fine-grained aspect-level sentiment analysis with the deepening of research and the increasing demand for practical applications of sentiment mining from web texts.become a research hotspot.The aspect-level sentiment analysis model is mainly based on deep learning methods,and there are many substantial breakthroughs.However,deep learning methods rely on annotated datasets for supervised learning,and there are few aspect-level sentiment analysis task datasets,which makes the model over-fitting at greater risk,and it is easy to fall into performance bottlenecks and application obstacles.In this thesis,a method based on multi-task learning is proposed,which uses the information of multiple related tasks to improve the generalization ability of the overall multi-task model and alleviate the problem of overfitting to a certain extent.Aspect-level sentiment analysis is divided into aspect word extraction task and aspect sentiment polarity classification task.Even though these two tasks are closely related,the task forms are quite different,which easily leads to negative transfer problems within the multi-task model,which significantly affects Overall performance of the multi-task model.In this thesis,the pre-trained model BERT is first used to train the word vector of the text.After that,the word vectors containing global semantic features are extracted by using the hybrid coding method,using Bi LSTM combined with CNN to extract global context features for the sentiment polarity classification task,and local features for the aspect word extraction task,so as to provide all Phrase boundary information is required.At the same time,the features extracted by Bi LSTM are classified as the private feature space of a specific task,and the features extracted by CNN are classified into the multi-task shared feature space.Then,for the negative transfer problem,a feature interaction learning module is designed,which mainly introduces the multi-head attention mechanism to realize the feature learning of multi-dimensional and different semantic subspaces;so as to effectively integrate private and shared features,as well as contextual representation and local features,and control the flow of information in different feature spaces to improve the quality of features acquired for specific tasks.In order to capture the potential correlation information between the aspect extraction task and the aspect sentiment polarity classification,a task interaction sharing module is designed in this thesis.Associated Information Features.Finally,the intermediate hidden state of a specific task is updated using the associated information semantics,so that the specific task can obtain semantically rich features.Finally,label learning is performed through a fully connected layer,and CRF is used to infer the output labels of the task.The model in this thesis is used for experiments on public datasets such as Sem Eval2014 and Twitter,which are commonly used in fine-grained sentiment analysis tasks.The results of the comparative experiments verify the effectiveness of the model in this thesis.Compared with multiple baseline models,the model proposed in this thesis has objective evaluation indicators such as precision,recall,and F1 value in the aspect word extraction task and aspect sentiment polarity classification task have some improvement.Furthermore,ablation experiments are carried out,and the experimental results further show that the feature interaction learning module and task interaction sharing module designed in this thesis can effectively improve the overall performance of the multi-task model.