-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsamplepaper.tex
168 lines (134 loc) · 13.4 KB
/
samplepaper.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
% This is samplepaper.tex, a sample chapter demonstrating the
% LLNCS macro package for Springer Computer Science proceedings;
% Version 2.20 of 2017/10/04
%
\documentclass[runningheads]{llncs}
%
\usepackage{bbding} % 用于加通讯作者的小信封
\usepackage{graphicx}
\usepackage{amsmath,amssymb,amsfonts}
\usepackage{algorithm}
\usepackage{algorithmic}
% 支持多行注释的宏包
\usepackage{verbatim}
\usepackage{CJKutf8} % 兼容中文
% 支持插入并排图片的4个宏包
\usepackage{float}
\usepackage{subfigure}
\usepackage{caption}
% Used for displaying a sample figure. If possible, figure files should
% be included in EPS format.
%
% If you use the hyperref package, please uncomment the following line
% to display URLs in blue roman font according to Springer's eBook style:
% \renewcommand\UrlFont{\color{blue}\rmfamily}
\setlength{\abovecaptionskip}{-0.1cm} %调整图片标题与图距离，有用
\captionsetup{belowskip=-35pt} % 有用,调整图片距离下文的距离
\begin{document}
%
\title{S\textsuperscript{2}AP: Sequential Senti-weibo Analysis Platform}
\author{Shuo Wan \and
Bohan Li \inst{ (}\Envelope\inst{)} \and
Anman Zhang \and
Wenhuan Wang \and
Donghai Guan
}
%
\authorrunning{S. Wan et al.}
% First names are abbreviated in the running head.
% If there are more than two authors, 'et al.' is used.
%
\institute{
Nanjing University of Aeronautics and Astronautics, Nanjing, China \\
\email{\{bhli, shuowan\}@nuaa.edu.cn}
}
% \vspace{-0.5cm} %调整图片与上文的垂直距离
%
\maketitle % typeset the header of the contribution
%
\begin{abstract}
Microblogging sentiment analysis aims at exploring people's opinion on social networks such as Twitter and Weibo. Existing work mainly focus on the English corpus based on Distant Supervision, which ignores the noise data in corpus and internationalization. The field of Weibo sentiment analysis lacks a large-scale and complete corpus for application and evaluation. In this work, we formulate the problem of corpus construction into an Information Retrieval problem and construct a Weibo sentiment analysis corpus called Senti-weibo. We also release a weibo pre-processing toolkit in order to unify the pre-processing rules of Weibo text. Eventually, we apply these works to implement a real-time Weibo sentiment analysis platform: S\textsuperscript{2}AP, which serves to analyze and track the sequential sentiment of Weibo topics.
\keywords{Corpus construction \and Weibo sentiment analysis \and Text pre-processing.}
\end{abstract}
\section{Introduction}
% 介绍微博情感分析的意义
Weibo, the most popular microblogging social network in China, has 497 million monthly active users in September 2019. People can express their opinion about breaking news, current affairs, politics and other topics. These subjective data are rich in sentiment information~\cite{zhang2019cyberbullying}, which brings great convenience to the research of sentiment analysis.
% 介绍情感分析与语料库的中外现状与不足之处
Sentiment analysis on Twitter has made significant progress in sentiment corpus. Most of the methods for constructing sentiment corpus use Distant Supervision~\cite{go2009twitter}, which labels the sentiment of Twitter with emotion symbols such as ``:-) '' and ``:( ''. However, emotion symbols cannot fully represent the sentiment polarity of tweet, and corpus mentioned above are actually corpus with noise. Compared with Twitter sentiment analysis, the field of Weibo sentiment analysis lacks a large-scale and complete corpus.
% 介绍文本预处理
%Pre-processing is usually used for denosing and dimensionality reduction. \cite{haddi2013role} demonstrates that appropriate text pre-processing methods can significantly enhance the classifier's performance. Weibo sentiment analysis is short of a unified rule for text cleaning. A unified pre-processing tool for training and online environment is indispensable. Supposing that we have applied a couple of word segmentation tool and cleaning rule to preprocess dataset and train a corresponding classifier. Then if we want to make full use of this classifier in online environment, the pre-processing rules should be consistent with training environment. Otherwise, a large number of unknown words will be generated.
Pre-processing for Twitter data is commonly used for denosing and dimensionality reduction. The experiment results of~\cite{haddi2013role} show that appropriate text pre-processing methods can significantly enhance the classifier's performance. Weibo sentiment analysis is short of a unified rule for text cleaning. A unified pre-processing tool for training and online environment is indispensable. For example, supposing that we have applied a couple of word segmentation tool and cleaning rule to preprocess Weibo dataset and train a corresponding classifier. Then if we want to make full use of this classifier in online environment, the segmentation tool and cleaning rules should be consistent with training environment. Otherwise, a large number of unknown words will be generated.
Our works are summarized below:
\begin{enumerate}
% 构建语料库的问题被转化：利用爬虫检索微博数据库的情感微博，提高召回的情感微博的精确率和召回率的问题。我们以提高召回微博精确率和召回率为目标对语料库进行迭代，对语料库进行降噪并扩展，并构建了一个
\item The problem of corpus construction is formulated into an Information Retrieval (IR) problem and we build a Weibo sentiment analysis corpus: Senti-weibo\footnote{Available at: http://bit.ly/2IEzTw1}, a collection of 671,053 weibos (equals tweet of Twitter).
\item We unify weibo pre-processing rules and package it into a toolkit in Python: weibo-preprocess-toolkit\footnote{https://pypi.org/project/weibo-preprocess-toolkit}, which is applied for weibo cleaning.
\item A Weibo sentiment analysis platform called S\textsuperscript{2}AP\footnote{http://sentiweibo.top} is constructed to analyze the real-time sentiment of Weibo topics. A Weibo Topic Spider is designed to crawl real-time weibos for analysis and fresh sentiment weibos for corpus iteration. In particular, we take two topics as concrete examples and track the sentiment trend of them.
\end{enumerate}
\section{Corpus Iteration and Construction}
% 利用 Tweet 的 emotion symbols 或 emojis 来构建语料库，实际上等价于根据检索词检索相关情感 Tweet。
Using emotion symbols or emojis to build corpus is actually equivalent to retrieving related sentiment tweets based on specified queries. The problem of corpus construction can be transformed into the problem of using spider to retrieve sentiment weibos from Weibo database, and improving the $Precision$ and $Recall$ of retrieved sentiment weibos.
\newtheorem{myDef}{Definition}
\begin{myDef}
Given a database $D$(public dataset or Weibo server), the retrieved weibos as $R(D)$. $Precision$ is the sentiment relevant weibos of $R(D)$, $Recall$ is the sentiment weibos in $R(D)$ retrieved from $D$:
\scriptsize{ % 公式字体大小
% 空格：\
% 公式对齐：align* 就是用来包裹需要对齐的公式， 以 & 开头，以 \\ 结尾，最后一行不需要 \\
\begin{align}
&Precision = \frac{\#\left ( sentiment\ weibos\ retrieved \right )}{\#\left ( retrieved\ weibos \right )} = P\left ( sentiment | R(D) \right ) \\
&Recall = \frac{\#\left ( sentiment\ weibos\ retrieved \right )}{\#\left ( sentiment\ weibos \right )} = P\left (retrieved\ sentiment | D\right )
\end{align}
} % end of 字体大小
\end{myDef}
In order to quickly initialize corpus, we firstly query sentiment weibos from public dataset with 40 typical emojis. The sentiment classifier trained on the initialized corpus with fastText~\cite{joulin2016bag} has the initial precision of 87.04\%. The test dataset comes from COAE2014~\cite{Yang2015Task} and we manually labeled 1,790 weibos.
Like most corpus~\cite{go2009twitter,iosifidis2017large} built with Distant Supervision, the initialized corpus also contains noise and slowly decays over time. From the perspective of IR, low $Precision$ is responsible for noise in the recalled weibos, while the decay of corpus corresponds to low $Recall$. In order to construct corpus and train model with high performance, we need to optimize the $Precision$ for high precision of sentiment classification model and improve the $Recall$ for generalized performance.
% 和大多数采用 DSL 方法构建的语料库一样，初始化的微博语料库包含有大量的噪声样本，而且语料库会慢慢腐烂，失去泛化性能。从信息检索的角度思考，对于我们召回的情感微博，大量的噪声样本对应了低精确率，而语料库的腐烂对应了低召回率。
Weibo Topic Spider is applied to retrieve sentiment weibos. We use spider with one emoji and its two synonyms words generated by word2vec as query to retrieve sentiment weibos from Weibo server. The combination of multiple features can improve the sentiment $Precision$ of retrieved weibos, and the similarity between the features can ensure $Recall$. Spider continuously collects fresh sentiment weibos with higher sentiment consistency and accumulates them for corpus iteration. In the process of iteration, we divide the corpus into training set and verification set, and sample subset of training dataset to train the classifier. Multiple supervised learning algorithms are applied to train sentiment classifiers to verify the sentiment label of verification dataset, and then sift out the weibos whose classification results do not match the original label. After several rounds of iteration, the precision of classifier trained by fastText has increased by 3.07\%. Finally, we build a Weibo sentiment analysis Corpus: Senti-weibo.
\begin{figure}[ht]
\vspace{-0.5cm} %调整图片与上文的垂直距离
\centering
\subfigure[unigram]{
\includegraphics[width=0.40\textwidth]{images/unigram.png}}
\subfigure[bigrams]{
\includegraphics[width=0.40\textwidth]{images/bigram.png}}
\caption{Multi-model evaluation of corpus}
\label{fig:corpus-evaluation}
\end{figure}
\section{Platform Implementation and Demonstration}
We design a demonstration platform called S\textsuperscript{2}AP to present our works mentioned in each section. Fig.~\ref{fig:Senti-weibo} depicts the architecture of S\textsuperscript{2}AP, which has two main function models: \textit{offline scripts} and \textit{front-end visualization}.
\begin{figure}[ht]
\vspace{-0.5cm} %调整图片与上文的垂直距离
\centering
\begin{minipage}[t]{0.43\textwidth}
\centering
\includegraphics[width=5.2cm]{images/Architecture-of-Senti-weibo-3.png}
\caption{Architecture of S\textsuperscript{2}AP}
\label{fig:Senti-weibo}
\end{minipage}
\begin{minipage}[t]{0.43\textwidth}
\centering
\includegraphics[width=5.2cm]{images/model-precision-on-test-dataset-cutted-by-each-tool.png}
\caption{Chinese segmentation tools}
\label{fig:segmentation-tools-precision}
\end{minipage}
\end{figure}
\textit{Offline scripts} model encapsulates scripts such as spider, corpus iteration, and text pre-processing. Fig.~\ref{fig:segmentation-tools-precision} shows the performance of the model caused by different pre-processing rules. We take six Chinese segmentation tools to cut test dataset, and apply the pre-trained model segmented by jieba\footnote{\url{https://github.com/fxsjy/jieba}} to classify each test dataset. Jieba is not the best Chinese segmentation tool, but the consistent rules in the training and test environment maximize the performance of model.
\textit{Front-end visualization} provides interface to interact with back-end. Fig.~\ref{fig:huawei-1day} represents the daily sentiment trend of Huawei from March to July. The negative sentiment of Huawei since May 16 accurately reflects the event that Huawei is added to the Entity List. Fig.~\ref{fig:trade-5min} reflects the sentiment trend of China-US trade from May 29 to May 30 in a five-minutes frequency, which corresponds to the intense emotional changes of netizen to the anchors' debate between China and US. Another feature of platform is the real-time analysis to reflect the latest sentiment trend, which can be available by visiting S\textsuperscript{2}AP.
\begin{figure}[ht]
\vspace{-0.5cm} % 调整图片与上文的垂直距离
\centering %图片全局居中
\subfigure[Huawei]{
\label{fig:huawei-1day}
\includegraphics[width=0.40\textwidth]{images/huawei-alert.png}}
\subfigure[China-US trade]{
\label{fig:trade-5min}
\includegraphics[width=0.40\textwidth]{images/china-us-detate-5min.png}}
\caption{Sequential sentiment analysis}
\label{fig:sentiment-trend}
\end{figure}
\section{Conclusion}
% 文章篇幅、预处理工具的实现细节，平台的实现细节等各种细节。We also open-source weibo data for two topics collected from April to July。以构建语料库为中心，顺带 more details of corpus construction and 更多的细节，例如爬虫的设计、文本预处理工具的设计、情感分析平台的构建，
In this work, we aim at the construction and iteration of Chinese Weibo sentiment analysis corpus. More datasets can be constructed using Information Retrieval and we will continue to abstract the problem of corpus construction. The details of S\textsuperscript{2}AP can be available on GitHub\footnote{https://github.com/wansho/senti-weibo}, including introduction video, design of pre-processing tool, implement of web demonstration and others.
\bibliographystyle{splncs04.bst}
\bibliography{mybibliography.bib}
\end{document}