Orientation Mining Driven Approach to Analyze Web Public Sentiment

1
Orientation Mining Driven Approach to Analyze Web
Public Sentiment
2
Abstract

• Internet provides an unique opportunity to
  express and spread public sentiment.
• Since web public sentiment reflects people's
  attitude to society and politics, the public
  opinions orientation is significant to
  decision-makers.
• The research utilizes VSM (vector space model) to
  present the text orientation of web information
  and offer data-mining approaches to analyze
  public opinions orientation.

3
Abstract

• To achieve the goal of text orientation analysis,
  two ways are proposed.
• a novel text orientation analysis method is
  described to analyze the orientation of original
  web postings and their replies.
• an improved single-pass clustering algorithm is
  introduced to cluster the subject of web
  discussion and discover the hot topics.
• The research constructs a prototype system, named
  WPSAS (web public sentiment analysis system), as
  experimental platform to validate the presented
  methodology.

4
Introduction

• Public opinion may be partial or impartial, also
  may be of fairness or prejudice.
• Once public sentiment becomes public opinion, it
  will do powerful impact to society.
• The goal of web public sentiment analysis is to
  discover the orientation of public sentiment
  before it changes to negative public opinion or
  public emergency, which greatly influence
  realworld society.

5
Introduction

• Web public sentiment activity can be divided as
  Internet news and Netizens discussion.
• Bulletin Board System (BBS) becomes the most
  important approach for Netizen express their
  sentiment on Internet/network.
• The research takes BBS as the data source to
  analyze web public sentiment.

6
Conference for web public sentiment analysis

• Text Retrieval Conference (TREC), Special
• Interest Group on Information Retrieval (SIGIR),
  Topic Detection and Tracking (TDT). The most
  well-known study on web public sentiment analysis
  system is TDT project 2
• It has five basic study missions story
  segmentation, topic tracking, topic detection,
  first-story detection and link detection.

7
VSM

• Vector space model (or term vector model) is an
  algebraic model for representing text documents
  (and any objects, in general) as vectors of
  identifiers, such as, for example, index terms.
• Vector space model was firstly brought forward by
  Sahon G. in 1975.

8
VSM

• VSM is used in information filtering, information
  retrieval, indexing and relevancy rankings.
• VSM is first used in the SMART Information
  Retrieval System.

9
SVM for analyzing web public sentiment

• However, to analyze web public sentiment, TF is
  too simple to calculate the weight of Internet
  opinion.
• The weight of a term is determined by three
  factors
• how often the term di occurs in the topic j (the
  term frequency tfji),
• how often it occurs in the whole document
  collection
• how strong it presents the emotion.

10
SVM for analyzing web public sentiment

• the weight of a term di of topic Tj is
• M is the number of collective topics, mi is the
  number of topics which contain di.
• where eji ? (0, 1) is the parameter describing
  emotional strength, T is the number of original
  posting which contained in the topic Tj .

11
Flow of WPSAS

• Initialization stage
• Computation stage
• Encapsulation stage
• Analysis stage

12
Text Orientation Mining
13
Orientation vector matching

• Five types according to the matching result
• Fully support Orientation vector of replying
  postings is fully consistent to orientation
  vector of original posting
• Partially support Orientation vector of replying
  postings is partially consistent to orientation
  vector of original posting, and there is no
  contrary weight on the same item, such as -1 to
  1
• Partially object there is contrary weight on the
  same item between orientation vector of replying
  postings and orientation vector of original
  posting

14
Orientation vector matching

• Fully object Every item weight of orientation
  vector of replying postings is opposite to
  orientation vector of original posting
• Neutral All item weight of orientation vector of
  replying postings is 0.

15
Orientation Mining Algorithm
16
Orientation Mining Algorithm
17

• Hot Topic Clustering

18
single-pass clustering

• uses the first text as the basis to cluster
• the rest texts are sequentially compared with
  their similarities
• If the similarity reaches the specified
  threshold, the texts are clustered into a group,
  and their features are recalculated as the basis
  for matching with other texts
• Or else, textual information is directly
  extracted as basis of a new kind of topic.

19
single-pass clustering

• The computation complexity of single-pass
  clustering is O (nk), where k is the number of
  the class.
• However, for web public sentiment analysis, there
  are two shortcomings
• this method is greatly dependent on the order,
  that clustering on the same object by different
  order will get different result.
• it is easy to be tendentious to some big category
  where clustering, because of the uneven
  distribution of categories.

20

• New clustering algorithm based on multi-comparing

21
Data Flow of WPSAS
22
Chinese word segmentation

• WPSAS adopts famous ICTCLAS (Institute of
  Computing Technology, Chinese Lexical Analysis
  System) to help us achieve goal of Chinese word
  segmentation, which is developed by Institute of
  Computing Technology of Chinese Academy of
  Sciences.
• main features of ICTCLAS include Chinese word
  segmentation, part-of-speech tagging, named
  entity recognition, new words identification and
  user dictionary.
• The latest version of ICTCLAS is ICTCLAS3.0

23
Experimental Data

• Two Web forums are selected for the public
  sentiment analysis
• Tencent forum (??) (http//bbs.news.qq.com/b-10000
  83746)
• Tianya forum (??) (http//www.tianya.cn).

24
Emotional words statistic-based vs VSM-based
orientation

• Recall of supporting postings (Sup_recall),
• accuracy of supporting postings (Sup_accuracy),
• recall of negative postings (Neg_recall)
• accuracy of negative postings (Neg_accuracy)
• overall accuracy

25
WPSAS vs machinelearning-based
26
Time cost
27
Conclusion

• The research presents data mining approaches to
  discover network public opinion.

28
Conclusion

• The contributions
• a novel text orientation analysis method
• to analyze the orientation of original web
  postings and their replies
• an improved single-pass clustering algorithm
• to cluster the subject of web discussion and
  discover hot topics
• a web public sentiment analysis system, named
  WPSAS,
• to analyze public opinion of web forum and
  validate the presented methodology