Interactive Visual Text Analytics for Decision Making

Transcription

1 Interactive Visual Text Analytics for Decision Making Shixia Liu Microsoft Research Asia 1

2 Text is Everywhere We use documents as primary information artifact in our lives Our access to documents has grown tremendously in recent years due to networking infrastructure WWW Digital libraries 2

3 Big Question What can information visualization provide to help users in understanding and gathering information from text and document collections? 3

4 Outline Example tasks in text analytics Visually analyzing textual information Dynamic word cloud Topic-based visual text summarization TextFlow: towards better understanding of evolving topics in text Future work 4

5 Text Analytics: Our Understanding How can I find information buried inside the piles of text? Information finding 5

6 Text Analytics: Our Understanding What is in my text? What s inside the NHTSA Data: 450,000+ documents What are the major causes of injuries 70,000+ patient emergency room records What did my customers say about my hotels customerposted reviews Information Understanding: Text Summarization 6

7 Text Analytics: Our Understanding What is in my text? Which hotel features do my customers like/dislike customer reviews How customers sentiment have changed toward my hotels customerposted reviews How do customers feel about my new product launch thousands of e- opinion postings Insight Discovery: Sentiment Analysis 7

8 Text Analytics: Our Understanding What is in my text? What are the correlations of tire problems and highway death in the NHTSA Data: 450,000+ documents What are the correlations of patient gender and the cause of injury 70,000+ patient emergency room records Compare the customers attitude toward our product with theirs for our competitors thousands of e- opinion postings Decision Making and Problem Solving: Text Analysis++ 8

9 Major Challenges Huge amounts of complex information Understanding the meanings of free text is just hard Performing analysis on top of that is harder Different people want different things No one-size-fits-all solutions People may not know what they want Tell me something I don t know I will tell you when I see it Machines are *not* just smart enough. 9

10 Outline Example tasks in text analytics Visually analyzing textual information Dynamic word cloud TIARA: topic-based visual text summarization TextFlow: towards better understanding of evolving topics in text Future work 10

11 Dynamic word cloud Selected Projects Illustrate content evolution trend TIARA Topic-based visual text summarization and analysis TextFlow Towards better understanding of evolving topics in text 11

12 Dynamic Word Cloud Word clouds for content overview Aesthetic issues Inadequate for temporal patterns Standard time chart: trend Inadequate for correlations Cui et al. IEEE CG&A

13 Our Solution A evolution trend chart + word clouds Measure the evolution Ensure the semantic coherence between clouds 13

14 Evolution Measurement Conditional entropy: measure the amount of information contained by Xi but not by Xj 14

15 Word Cloud Layout Geometry meshes to ensure the semantic coherence Semantically related words stay together The same word in different clouds stay at the similar place 15

16 Example: CG&A Abstracts ,984 abstracts IEEE Computer Graphics and Applications (CG&A) from 1981 to

17 Comparison with Wordle 13,828 news articles 17

18 Dynamic word cloud Selected Projects Illustrate content evolution trend TIARA Topic-based visual text summarization and analysis TextFlow Towards better understanding of evolving topics in text 18

19 Demo Y axis encodes topic significance (3) Doc # (1) A layer represents a topic (2) Topic keywords X axis encodes time ~10,000 s in

20 Demo 20

21 Key Challenges Summarize text corpora Huge amounts of complex information Time-varying Visually explain summarization results Consistent visualization Provide feedback or articulate their needs Imperfect summarization results or varied user needs 21

22 TIARA Overview Text collection Text Preprocessing Text content + meta data Text Summarization Summarization results Data Transformation Transformed results Visualization User Interaction 22

23 TIARA Technical Focus Text collection Text Preprocessing Text content + meta data Text Summarization Summarization results Data Transformation Transformed results Visualization User Interaction 23

24 Text Summarization Latent Dirichlet Allocation (LDA) model [Blei et al. 03] High portability High compaction rate for scalability A finer grained model kth document in the collection {, P(T i D k ), } A set of topic probabilities {T 1, T i, T N } A set of topics {W 1,, W j,, W M } A set of keywords A set of word probabilities {, P(W j T i ), } 24

25 LDA Data Transformation kth document in the collection {, P(T i D k ), } A set of topic A set of topics probabilities {T1, T i, T N } Rank the topics to present most interesting ones first Select keyword sub-sets for time-based summary { } t-1, {, W j, } t, { } t+1, A set of keywords {W 1,, W j,, W M } A set of word probabilities {, P(W j T i ), } 25

26 Topic Ranking by User Interests Rank topics by strength Rank topics by distinctiveness rank( Tk ) f ( Tk ), ( Tk ), ( Tk ) M N ˆ m m 1 m, k coverage variance M ( Tk ) M N ˆ m m m k T 1, ( k ) N ( Tk ) M m m 1 v~ topic Lv~ T k k rank ( Tk ) l( Tk ) T v~ k Dv~ k graph degree matrix topic Domain-dependent activeness metric graph Laplacian m 1 for each T k,,,..., T v ~ k is normalized v k [Song et al. CIKM 09] ˆ ˆ N v k 1, k 2, k M, k ˆ m doc-topic distribution 26

27 Experiments Goal Measure which metric produces more important topics Data sets messages News 34,690 documents Method Users indicate the importance of top-k ranked topics Very important, somewhat important, Unimportant 27

28 Results data (by F1 measure) News data (by F1 measure) 28

29 TIARA Technical Focus Text collection Text Preprocessing Text content + meta data Text Summarization Summarization results Data Transformation Transformed results Visualization User Interaction 29

30 Visualizing Text: Existing Work Visualize text at a high level Visualize text at a low level Few on explaining advanced text analysis results 30

31 Visual Text Summary Metaphor Data to be visualized: 1. Topics: {T 1, T i, T N } and their probabilities 2. For each T i,topic keywords by time: {,w i k,, } t, and their probabilities over time 3. For each T i, Topic strength: {..., S i (t), } over time Visual encoding: Augmented stacked graph 31

32 Enhanced Stacked Graph: Key Steps Computing geometry of layers Layer coloring Layer ordering Layer labeling 32

33 Enhanced Stacked Graph: Key Steps Computing geometry of layers Layer coloring Layer ordering Layer labeling Byron_Infovis08 33

34 Goals Layer Ordering Minimize distortion Maximize usable space Ensure semantic coherence unordered ordered 34

35 Layer Ordering - Comparison Observed topic 35

36 Layer Ordering - Comparison Same topic 36

37 Layer Ordering - Comparison Same topic 37

38 Enhanced Stacked Graph: Layer Labeling Goals Temporal proximity Informativeness 38

39 Enhanced Stacked Graph: Layer Labeling (cont d) Our approach [Liu et al. CIKM09] Constraint-based space allocation Particle-based layout [Luboschik et al. 08] + wordle s1 s2 t- t t+ 2t T 39

40 Enhanced Stacked Graph: Layer Labeling (cont d) 40

41 Interacting with Visual Summary people involved and their relationships 41

42 Application Example: Healthcare Visualize text to facilitate analysis Cause of injury Reason for visit Diagnosis Multiple fields of text data and their correlation Leverage structured data to help better illustrate text information Gender + Cause of injury 42

43 Correlation between Structured and Text Fields 43

44 Correlation between Text Fields Correlation between two fields, diagnosis and reason for visit 44

45 Dynamic word cloud Selected Projects Illustrate content evolution trend TIARA Topic-based visual text summarization and analysis TextFlow Towards better understanding of evolving topics in text 45

46 Problems Understanding topic evolution in large text collections is important Keep abreast of hot, new, and intertwining topics Gain insight into the latent topics 46

47 Scholars Applications Find related works in a publication set Business professional Examine a large collection of s and instant messages Politicians Examine online posts to identify the key public opinion and concern 47

48 Example 48

49 Example 49

50 Application Example: 933 VisWeek Pulications geometry/rendering flow/vector volume/rendering structure/layout exploration/analytics document/ temporal 50

51 Application Example: 933 VisWeek Pulications 51

52 Challenges Model Topic merging/splitting patterns Visually convey Topic merging/splitting patterns in an intuitive way Facilitate Analytical reasoning 52

53 Related Work Most of the existing work Studying the evolution of individual topics Little work Studying topic merging and splitting patterns Barely been touched Using visual analytics techniques to interactively analyze complex topic evolution 53

54 Our Solution Leverage hierarchical Dirichlet processes To model topic merging/splitting Augment familiar visual metaphors (rivers) To convey the complex analytic results Support interactions at different levels Smooth communication between visualization and the topic mining model 54

55 TextFlow Overview 55

56 Topic Data and Relationship Extraction Incremental Hierarchical Dirichlet Processes Online learning of the topics in text Automatically detect the topic numbers Extract the merging/splitting relationships Based on document topic change Online compute the merging/splitting probabilities 56

57 Splitting Relationship Time t Time t+1 Predict Refine 57

58 Merging Relationship Time t Time t+1 Predict Refine 58

59 Critical Event Extraction Types of critical events Birth, death, merge, and split Scoring the merging/splitting event Number of the branches Entropy of the branching probabilities 59

60 Keyword Correlation Discovery Extract noun phrases, verb phrases, and named entities in each document Count Co-occurrences among them Be used to illustrate why 60

61 Topic Evolution as Flow Alternative Our design 61

62 Critical Event as Glyph Emerge, dissolve, split, and merge 62

63 Keyword Correlation as Thread Alternatives 63

64 Keyword Correlation as Thread Intertwine to indicate co-occurrences 64

65 Visualization Design - Consistency Topic Critical event Keyword 65

66 Layout Algorithm Three-level directed acyclic graph (DAG) 66

67 Interactive Exploration ilter place glyphs ilter ilter ilter expand place threads 67

68 Application Example: Bing News events happening in Egypt protest in Egypt Obama government s actions budget/spending republican/campaign 68

69 Application Example: Bing News (a) (b) 69

70 Application Example: Bing News 70

71 Video 71

72 Outline Example tasks in text analytics Visually analyzing textual information Dynamic word clouds Topic-based visual text summarization TextFlow: towards better understanding of evolving topics in text Future work 72

73 Future Text Visualization Topics Interactive, incremental text analytics Multi-level visual text summarization (keywords + sentences) Multi-faceted text analytics (e.g., summarization + sentimental analysis) Multimedia document summarization (text + image + video) Interactive, visual social media analysis 73

74 Acknowledgements Weiwei Cui, Yangqiu Song, Furu Wei, Xin Tong (MSRA) Nan Cao, Yingcai Wu, Prof. Huamin Qu (HKUST) Dr. Michelle X Zhou (IBM Almaden Research Center) 74