面向新闻文本的情感原因抽取算法研究

Abstract

Abstract

With the development of the news media, more and more netizens have learned about hot news via official Weibo, official WeChat public account, news client, etc. However, due to the large number of news and the continuous dissemination, there are repeated news, etc. People's browsing and reading alone can hardly systematically sort out all the news and extract valuable information from it; on the other hand, a large number of texts make it difficult for policy makers to objectively evaluate the accuracy of the news and respond in a timely manner. For example, rumors, etc.

The best sentiment analysis model can learn texts in depth, and can achieve accuracy of 80% to 90% in predicting the emotional polarity of texts. Unfortunately, although there are many applications that already analyze the emotions of text, it is not enough. In order to make emotional predictions better applied, decision makers need to know what caused the emotion. In the public opinion analysis system, decision makers need to know which root causes the distribution of public opinion, and this is the reason why people's emotional distribution is not involved in hot events.

Emotional reasoning the task of extracting tasks is to extract the clauses containing the causes from the long sentences with emotional colors. This article will study from the following three aspects: extraction based on emotional reasons of conditional random fields, extraction of emotional reasons based on neural network combined conditions, and extraction of emotional reasons based on memory network.

The experimental results show that the extraction based on the conditional random field is better than the knowledge-based and rule-based method. The neural network combined with the conditional random field model to identify the non-cause clauses is of great help. The reason extraction based on the memory network, the same word the memory matrix of the vector matrix, the memory network combined with the attention mechanism, and the memory network to increase the context window are helpful to the improvement of the experimental results. Changing the question sentence, changing the length of the sentences, and changing the number of layers have the same effect on the experimental results. Changing the question sentence, changing the length of the sentences, and changing the number of layers have the same effect on the experimental results directed impact.

Keywords:Sentiment Analysis, Cause Extraction, Memory Network, Conditional Random Fields, Bidirectional Neural Networks

目录

目录

摘要 .......................................................................................................................... I ABSTRACT............................................................................................................... II 第1章绪论 (1)

1.1课题背景及研究目的和意义 (1)

课题背景 (1)

课题研究的目的和意义 (1)

1.2国内外研究现状 (2)

基于规则的方法 (2)

基于统计机器学习的方法 (3)

基于深度神经网络学习的方法 (4)

国内外文献综述简析 (4)

1.3本文研究内容及章节安排 (5)

1.4本文研究内容 (5)

本文章节安排 (6)

第2章基于条件随机场的情感原因抽取 (7)

2.1引言 (7)

2.2条件随机场 (7)

CRF介绍 (7)

CRF++介绍 (8)

2.3使用条件随机场抽取原因 (9)

2.4实验设置 (10)

语料集 (10)

对比实验模型 (12)

工具集与评价指标 (12)

2.5实验结果及分析 (13)

2.6本章小结 (13)

第3章基于神经网络结合条件随机场的情感原因抽取 (15)

3.1引言 (15)

3.2神经网络结合条件随机场 (15)

3.3使用神经网络结合条件随机场抽取原因 (18)

3.4实验设置 (19)

目录

工具集与评价指标 (19)

3.5实验结果及分析 (20)

3.6本章小结 (20)

第4章基于记忆网络的情感原因抽取 (22)

4.1引言 (22)

4.2记忆网络模型及其改进 (22)

基于端到端的记忆网络 (22)

使用相同词向量矩阵的记忆网络改进 (25)

结合注意力机制的记忆网络改进 (27)

关键词-值的记忆网络 (28)

结合关键词-值网络的记忆网络改进 (31)

4.3实验设置 (32)

数据集构建 (32)

模型设置 (35)

工具集与评价指标 (36)

4.4实验结果及分析 (36)

不同实验模型 (36)

不同问句形式 (37)

不同句子长度 (37)

不同迭代层数 (38)

错误分析 (38)

对比实验模型 (39)

4.5本章小结 (40)

结论 (42)

参考文献 (43)

哈尔滨工业大学学位论文原创性声明和使用权限 (46)

致谢 (47)