| 以下仅列出文章摘要、目录等部分内容，如需获取完整论文资料，或原创定制、长期合作，请随时联系。 微信QQ:312050216 |
 |
 |
| 基于深度多任务网络的方面级情感分析研究与应用
摘 要
随着互联网应用的发展和普及，人们更倾向于在线上社区、博客、微博等网络平台上发表自己的看法、态度或心情，由此产生大量的评论文本数据，使得情感分析在处理文本数据方面成为一个研究热点。这些评论文本数据资源，同时也是服务平台或产品汲取意见的依据。本文针对评论文本进行细粒度情感分析技术研究，构建了基于深度多任务网络的方面级情感分析模型，用于联合学习评论实体方面项提取及情感分类任务，并在三种数据集上实验验证模型有效性。本文主要工作内容如下：
（1）构建了一种联合学习ATE和ACD任务的深度多任务网络模型。现有大多数ABSA研究只针对某一项子任务，这就会导致丢失任务间的相关信息，本论文将用于ATE任务的BiLSTM层和用于ACD任务的CNN层结合在一个多任务框架中，在进行ACD任务时可以学习ATE任务重的信息，实现交叉学习，验证了多任务学习的有效性。
（2）利用BERT、BiLSTM、CRF、CNN构建了一种联合处理方面术语提取和方面情感极性分类的深度多任务网络模型。采用流水线策略解决ABSA中相关子任务时，模型的误差会从上游任务传递到下游任务，从而导致错误不断累积，所以本论文在已有特征提取模型BiLSTM-Attention和命名实体识别模型BERT-BiLSTM-CRF之上，引入CNN层，构建深度多任务网络模型，提高了方面术语提取及其情感极性分类性能。
（3）基于本文第四章构建模型，设计开发了基于做业反馈文本分析的学业情绪识别系统。该系统可提取大量作业反馈文本中学生对教学活动或作业的观点与意见，并按学业情绪极性分类。该系统提供了一种科学分析学生学习状况的有效手段，为教师优化教学活动、改进教学内容提供依据。
关键词： 方面级情感分析；多任务学习；BERT语言模型；双向LSTM
Abstract
With the development and popularization of Internet applications,people are more inclined to post their opinions,attitudes or moods on online communities, blogs,microblogs and other network platforms, resulting in a large amount of comment text data, making sentiment analysis processing text data This aspect has become a research hotspot. These comment text data resources are also the basisfor the service platform or product to draw opinions. This paper conducts fine-grained sentiment analysis technology research on review text, and builds an aspect-level sentiment analysis model based on deep multi-task network for joint learning of review entity aspect item extraction and sentiment classification tasks, and experimentally verifies the model on three data sets Effectiveness. The main contents of this paper are as follows:
A deep multitasking network model for joint learning of ATE and ACD tasks is constructed.Most of the existing ABSA studies only focus on a certain sub-task,which will lead to the loss of relevant information between tasks. In this paper, the BiLSTM layer for ATE tasks and the CNN layer for ACD tasks are combinedinto a multi-task framework, which can learn the heavy information of ATE tasks during ACDtasks, achieve cross-learning, and verify the effectiveness of multi-task learning.
Using BERT, BiLSTM, CRF, and CNN,a deep multi-task network model for joint processing of aspect terminology extraction and aspect sentiment polarity classification is constructed. When the pipeline strategy is used to solve related subtasks in ABSA, the error of the model will be passed from the upstream task tothe downstream task, which will lead to the accumulation of errors. Therefore, this paper uses the existing feature extraction model BiLSTM-Attention and the named entity recognition model BERT-BiLSTM- On top of CRF,the CNN layer is introduced to build a deep multi-task network model,which improves the performance of aspect term extraction and emotional polarity classification.
Based on the model constructed in the fourth chapter of this paper,an academic emotion recognition system based on homework feedback text analysis is designed and developed. The system can extract a large number of homework feedback texts of middle school students' views and opinions on teaching activities or assignments, and classify them according to the polarity of academic emotion. The system provides an effective means to analyze students' learning situation scientifically,and provides a basis for teachers to optimize teaching activities and improve teaching content.
|
