SENTIMENT ANALYSIS. CS 498 Mar 6

Transcription

1 SENTIMENT ANALYSIS CS 498 Mar 6

2 Macbeth, Scene 1, Act 2 from Wordle

3 my Citeulike page

4 Brad Paley s TextArc.

5 Fernanda Viégas s Th .

6 Martin Wattenberg s recent Word Tree visualization, showing Alberto Gonzales s testimony.

7

8

9 PNNL s ThemeRiver.

10 PNNL s IN-SPIRE.

11 tools you know...practical stuff

12 Stanford s list LIWC SentiWordNet Pang & Lee s data at Cornell

13 analysis & design how we might use it and why

14 If I could figure out a way to determine whether people are more fearful or changing to more euphoric I can forecast the economy better than any way I know. e trouble is, we can't figure that out. Alan Greenspan, Jan 2008

15 1841

16 Nasdaq vs. LiveJournal anxious moods, Jan 3 Oct 26, 2007.

17 BOOSTED DECISION TREE CLASSIFIER 1. nerv* 2. wor* 3. anx* 4. hop* 5. you* 6. scar* 7. tomorrow 8. fun 9. war 10. your* 11. going 12. be* 13. interview other notables: 16. lov*, 21. hospital, 36. awesome, 51. yay, 89. exam*

18 All LiveJournal blog posts posts per minute: ~107 Bagged Naive Bayes classifier Anxious true positive rate: 28% Anxious false positive rate: 3.4% Boosted Decision Tree classifier Anxious true positive rate: ~30% Anxious false positive rate: ~6% Percentage of anxious posts in 10-min period Percentage of anxious posts in 10-min period Adapted Wald adjustment (lower bound on 95% CI) Adapted Wald adjustment (lower bound on 95% CI) average 60-min moving average

19 Dow Jones daily close 7-day exponential moving average 13.2K 13.0K 12.8K 12.6K 12.4K 12.2K 12.0K 11.8K 15% 11.6K 10% μ + 6σ 5% 11.5M posts Percentage anxious blog posts 0 Jan 2008 Feb Mar Apr May Jun Jan 26 Of three predictive spikes, this is the furthest from a local maximum: it appears 5 trading days later on Feb 1. The SC primary happens on this date. The Fed lowers rate 4 days before and 4 days after this date. Feb 24 This spike comes three days before the second most critical maxima over this six month period. The Dow takes nearly two months to recover. After searching newspapers near this date, it is not clear what event may have caused this spike. Consumer confidence and poor business/housing reports follow in the next two days. Mar 25 This spike is probably noise, although it does preface a steep decline. Detecting important blogs and topics may eliminate spikes like these. Conference Board s consumer confidence came out this day and could be responsible for the spike. May 16 This spike appears three days before the most important local maxima. As of June 24, the Dow has still not recovered from May 19, dropping nearly 10% to date. Michigan s consumer sentiment index came out this day, along with unexpectedly poor housing numbers. May 19 followed with many poor business reports (2.5 s.d. anxiety spikes on May 19).

20 OUR BLOG COMMENT DATASET 5 blog genres 33 top blogs 1,094 blog comments

21 Great post and I really like the video. This is extremely similar to the approach I use in writing almost anything ProBlogger Just wait until hackers exploit the print layer to this mesh stuff enough to grab root and start injecting python code Scobelizer

22 Great post and I really like the video. This is extremely similar to the approach I use in writing almost anything ProBlogger Just wait until hackers exploit the print layer to this mesh stuff enough to grab root and start injecting python code Scobelizer

23 neither Proportions of agreement 49.4% agree 39.2% disagree 11.1% Wald method p < 0.05

24 LEXICAL uni/bi/trigams TFIDF POS raw tags combo lexical SENTIMENT congressional floor rotten tomatoes LIWC SEMANTIC sim to post ESA AGREE/DISAGREE/NEITHER NAMED ENTITY organizations people

25 Features + Info Gain LIWC pos. emotion words LIWC affect words exclamations ellipsis great is tech blog cosine similarity to post great [noun] personal pronouns present tense verbs [prepos] [poss pronoun] tf-idf dot product with post coordinating conjunctions agree agree agree agree neither!disagree agree neither!disagree agree!disagree neither agree!neither agree