Research On Aspect Based Sentiment Analysis Algorithm Of English Comments Based On Deep Learning

The rapid development of information technology and digital economy has led to a significant increase in network data,which has brought convenience to people but has also created problems such as the burden of manual processing of large amounts of data.Aspect Based Sentiment Analysis(ABSA)is a fundamental task in the field of natural language processing(NLP)and plays a crucial role in many downstream applications,such as reading comprehension,public opinion analysis,and dialogue systems.Despite the good performance of language models in current deep learning algorithms on some public ABSA datasets,sentiment classification accuracy and recall rates in complex multi-objective texts are easily affected by sequential prediction interference,thus compromising the performance of sentiment analysis models.This paper aims to improve the performance of the ABSA algorithm for English review text.The paper addresses the problems by focusing on two subtasks,i.e.,evaluation object extraction and evaluation object emotion polarity analysis.The paper proposes and evaluates an ABSA algorithm for English text that incorporates weight sharing,pre-trained models,and attention mechanisms.This paper aims to optimize the F1 score of the model in complex multi-objective scenarios with inaccurate sequence prediction.The main contents are as follows:(1)To address the problem of ambiguous sentiment classification caused by inaccurate sequence prediction in English texts,this paper proposes a model named WSABSA in the third chapter.The model improves emotion classification performance in the ABSA task by sharing the network layer parameters of the two sub-tasks to calculate the feature fusion weights.The experimental results demonstrate that the method of sharing feature fusion weights works well for the ABSA task of English text.The WSABSA model has improved sentiment classification performance on three English review benchmark datasets to varying degrees when compared to traditional models.(2)This paper proposes a model named Tapt-WSFSA in the fourth chapter to address the poor finetuning effect of the target task.The model incorporates task adaptive pretraining into the TAPT network.By pre-training the model on task-related unlabeled corpora and then fine-tuning it on specific tasks,the model can better understand language features and overcome problems such as catastrophic forgetting and overfitting.This method improves target domain vocabulary coverage and model performance in downstream tasks.By continuing pre-training on task-related unlabeled corpora and then finetune specific tasks,the model can better understand language features and avoid problems of catastrophic forgetting and overfitting,thus improving the coverage of target domain vocabularies and model performance of downstream tasks.The experimental results show that embedding task-adaptive pre-training into TAPT network effectively enhances the model’s feature extraction abilities,time utilization,and accuracy of emotion classification.(3)This paper proposes a model named HAN-TAPT-WSFSA in the fifth chapter to tackle the issue of poor semantic information acquisition in texts,especially for long texts.By implementing the two-level attention mechanism to words and sentences and giving different priorities to different sentences in the document and different words in the sentence.This optimization and targeted representation of the model improves the performance of feature extraction.According to experimental results,the hierarchical attention mechanism HAN can effectively improve the level of semantic information acquisition in text.