Sentiment Analysis

Wouter van Atteveldt
Glasgow Text Analysis, 2016-11-17

Sentiment Analysis

Sentiment Analysis: problems

“The man who leaked cell-phone coverage of Saddam Hussein's execution was arrested”

  • Language is subjective, ambiguous, creative
  • What does positive/negative mean?
    • e.g. Osgood ea 1957: evaluation, potency, activity
  • Who is positive/negative about what?
    • Sentiment Attribution

Sentiment Analysis resources

  • Lexicon (dictionary)
  • Annotated texts
  • Tools / models

Lexical Sentiment Analysis

  • Get list of positive / negative terms
  • Count occurrences in text
  • Summarize to sentiment score
  • Possible improvements
    • Word-window approach (tomorrow)
    • Deal with negation, intensification

Lexical Sentiment Analysis in R

  • Nothing new here!
  • Directly count words in DTM:
lex = list(pos=c("good", "nice", "great"), neg=c("bad","stupid", "crooked"))
library(slam)
npos = row_sums(dtm[, colnames(dtm) %in% lex$pos])
nneg = row_sums(dtm[, colnames(dtm) %in% lex$neg])
sent = data.frame(id=rownames(dtm), npos=npos, nneg=nneg)
sent$subj = sent$npos + sent$nneg
sent$sent = ifelse(sent$subj == 0, 0, 
                   (sent$npos-sent$nneg) / (sent$subj))
sent = merge(meta, sent)
Head(sent)
id date medium year month week npos nneg subj sent
162317736 2016-02-01 The New York Times 2016-01-01 2016-02-01 2016-02-01 0 0 0 0
162317809 2016-01-25 The New York Times 2016-01-01 2016-01-01 2016-01-25 0 0 0 0
171932192 2016-07-04 The New York Times 2016-01-01 2016-07-01 2016-07-04 0 0 0 0
171932219 2016-07-03 The New York Times 2016-01-01 2016-07-01 2016-06-27 1 0 1 1
171932226 2016-07-03 The New York Times 2016-01-01 2016-07-01 2016-06-27 0 0 0 0
171932227 2016-07-03 The New York Times 2016-01-01 2016-07-01 2016-06-27 0 0 0 0

Lexical Sentiment Analysis in R

a = aggregate(sent[c("sent", "subj")],
              sent[c("week", "medium")], sum)
library(ggplot2)
ggplot(a, aes(x=week, y=sent, colour=medium)) + 
  geom_line()

plot of chunk unnamed-chunk-5

Lexical Sentiment Analysis: Alternatives

Apply directly to tokenlist:

tokens$sent =0
tokens$sent[tokens$lemma %in% lex$pos] = 1
Head(tokens[tokens$sent > 0,])
id sentence offset word lemma POS POS1 ner sent
230726 171932219 50 3857 good good JJ G O 1
316450 171932322 2 193 good good JJ G O 1
431087 171932429 22 3577 Good good JJ G O 1
547654 171932557 17 1848 good good JJ G O 1
547812 171932557 25 2589 great great JJ G O 1
560047 171932568 13 2011 good good JJ G O 1

Lexical Sentiment Analysis: Quanteda

Use quanteda::apply

library(quanteda)
library(corpustools)
dfm = dtm.to.dfm(dtm)
dfm = applyDictionary(dfm, lex)
head(dfm)

Document-feature matrix of: 999 documents, 2 features (0% sparse).
(showing first 6 documents and first 2 features)
           features
docs        pos neg
  162317736   0   0
  162317809   0   0
  171932192   0   0
  171932219   1   0
  171932226   0   0
  171932227   0   0

Acquiring a lexicon

Parsing a sentiment lexicon

Parsing a sentiment lexicon

lex = readRDS("lexicon.rds")
dict = list(
  pos = lex$word1[lex$priorpolarity == "positive"],
  neg = lex$word1[lex$priorpolarity == "negative"],
  trump = "Trump",
  clinton = c("Hillary", "Clinton"))

Proximity-based sentiment analysis

  • Political texts contain multiple statements
  • Apply words to actors in close proximity
    • Within sentence/paragraph -> create dtm/dfm at sentence level
    • Within N words -> use token list
  • However, “Trump calls Clinton crooked”

Sentence-level sentiment:

tokens$doc = paste(tokens$id, tokens$sentence, sep="_")
dtm = dtm.create(tokens$doc, tokens$lemma, minfreq = 10)
x = sapply(dict, function(x) 
  row_sums(dtm[, colnames(dtm) %in% x]))
Head(x)
pos neg trump clinton
162317736_1 2 2 0 0
162317736_2 2 4 0 0
162317736_3 0 1 0 0
162317736_4 1 2 0 0
162317736_5 3 1 0 0
162317736_6 0 4 0 0

Proximity-based sentiment: semnet

devtools::install_github("kasperwelbers/semnet")

library(semnet)
tokens$concept = NA
for(c in names(dict))
  tokens$concept[tokens$lemma %in% dict[[c]]] = c

hits = windowedCoOccurenceNetwork(location=tokens$offset, 
    term=tokens$concept, context=tokens$id,
    window.size=40, output.per.context = T)
hits = subset(hits, x %in% c("clinton", "trump") 
              & y %in% c("pos", "neg"))
hits$sent = ifelse(hits$y == "pos", 1, -1)
tapply(hits$sent, droplevels(hits$x), mean)

   clinton      trump 
0.10921228 0.02649982

Sentiment Analysis: difficulty

  • Liu: “Although necessary, having an opinion lexicon is far from sufficient for accurate sentiment analysis” … “sentiment analysis tasks are very challenging.”
  • “[sentistrength] has human-level accuracy for short social web texts in English, except political texts.”
  • Subjective language is
    • Creative
    • Ambiguous
    • Subjective
    • Content-sensitive
  • Political communication is (often) nuanced, complicated

Improving sentiment analysis

  • Domain adaptation
    • Get term statistics
    • Merge with lexicon
    • Manually check top-X frequent words
  • Targeted sentiment anlaysis (next section)
  • Crowd sourcing (e.g. Haselmayer in press)
  • Machine learning (see handout)

Hands-on session III

  • Sentiment Analysis of election campaign
    • (or your own data …)
  • What is the overall sentiment?
    • Development over time, per medium
  • What is sentiment for different actors?
  • Try out different lexica
  • Try to adapt lexicon to domain

Conclusion

  • Frequency-based text analysis
    • Corpus analysis and topic modeling
    • Natural language procesessing
    • Sentiment Analysis
  • Lots of resources out there
    • See e.g. vanatteveldt.com/glasgow-r
  • Questions?