Presentation on theme: "There’s more to sentiments than you think! Guest lecture in Muhammad Abdul-Mageed’s S604: Social Media Mining, Indiana University, March 6, 2013 By David."— Presentation transcript:
There’s more to sentiments than you think! Guest lecture in Muhammad Abdul-Mageed’s S604: Social Media Mining, Indiana University, March 6, 2013 By David Heise Rudy Professor of Sociology Emeritus, IU
Sentiment analysis assesses people’s emotionality from texts that they produce. It answers the question: When talking about X, are people happy or unhappy? The trouble is…
Emotions cover a lot more territory than just happy and unhappy! So if you assess just happiness and unhappiness, you’re missing a lot of what people are feeling and expressing in their texts. Pleasant vs. unpleasant; stillness vs. arousal; dominant vs. vulnerable.
There are three affective dimensions: Badness versus Goodness (unpleasant-pleasant) Weakness versus Powerfulness (vulnerable-dominant) Quiescence versus Activation (stillness-arousal) Research projects in 30 different cultures have verified these dimensions. Osgood, Charles E., W. H. May, and M. S. Miron. 1975. Cross-Cultural Universals of Affective Meaning. Urbana, IL: University of Illinois Press. Heise, D. R. 2010. Surveying Cultures: Discovering Shared Conceptions and Sentiments. Hoboken, NJ, Wiley Interscience. Scholl, Wolfgang. 2013. The socio-emotional basis of human interaction and communication: How we construct our social world. Social Science Information, 52(1), 3–33. The three dimensions are a universal basis for how people affectively respond to all kinds of stimuli, including other people and their behaviors.
Measuring affective response The affective meanings of words, phrases, pictures, sounds, or smells are measured with computerized scales, one for each dimension.
Measurement details Each stimulus gets a three-number profile, corresponding to measurements on Evaluation, Potency, Activity (EPA). Measurements range from -4 (bad, powerless, inactive) to +4 (good, powerful, active) Stimuli typically are rated by 30-40 respondents, and averages over all respondents constitute the final EPA profile.
Example EPA measurements of words by IU male students in 2002-2004 PeopleBehaviorsModifiers WordEPA EPA EPA champion126.96.36.199rescuing188.8.131.52adventurous184.108.40.206 grandparent2.71.5praying with1.81.7-1.3peaceful2.81.5-1.8 child1.4-0.82.1lusting for1.00.21.0feminine0.6-0.40.0 senior citizen1.2-0.9-1.8obeying0.6-0.5cautious1.2-0.2-0.7 politician-0.91.8 fighting-220.127.116.11enraged-1.21.01.8 executioner-2.01.9-0.7incarcerating-0.90.7-0.6dogmatic-0.50.3-0.6 prostitute-1.9-2.00.7babbling to 1.4panicked-1.71.2 victim-1.3-2.4-1.6stealing from-2.9-0.8-1.6cowardly-2.1 -2.0 ±1 = slightly, ±2 = quite, ±3 = extremely
GoodBad champion, grandparent, child, senior citizen, rescuing, praying with, lusting for, obeying, adventurous, peaceful, feminine, cautious politician, executioner, prostitute, victim, fighting, incarcerating, babbling to, stealing from, enraged, dogmatic, panicked, cowardly PowerfulPowerless champion, grandparent, politician, executioner, rescuing, praying with, fighting, incarcerating, adventurous, peaceful, enraged, dogmatic child, senior citizen, prostitute, victim, lusting for, obeying, babbling to, stealing from, feminine, cautious, panicked, cowardly LivelyQuiet champion, child, politician, prostitute, rescuing, lusting for, fighting, babbling to, adventurous, feminine, enraged, panicked grandparent, senior citizen, executioner, victim, praying with, obeying, incarcerating, stealing from, peaceful, cautious, dogmatic, cowardly These words can be used as indicators of good (favorable) versus bad (unfavorable) in sentiment analyses of social media. The same words can be used as indicators of power versus weakness. And the same words can be used as indicators of activation versus passivity.
Affect control theory uses the three dimensions and a mathematized theory to predict people’s behaviors. Setting up program Interact. Predicting behaviors. “ACT is one of the most encompassing and precise social-psychological theories, translating … qualitative, phenomenological approaches … into an exact quantitative system with point predictions that deliver astonishing plausible results” (Scholl, 2013: 21) Interact is at http://www.indiana.edu/~socpsy/ACT/interact/JavaInteract.html
With three dimensions and a mathematized theory that people maintain their sentiments about self and other, we can predict behaviors, emotions, attributions, and labelings. You can’t do all that with just one dimension! With one dimension, we would know that the father should do some friendly act toward his daughter. But should he rescue her, pray with her, lust for her, obey her, or what? With one dimension, we would know that the father should do some friendly act toward his daughter. But should he rescue her, pray with her, lust for her, obey her, or what?
WHERE DO YOU GET THE EPA DATA? Suppose you want to do three-dimensional sentiment analyses.
Getting EPAs for words in Interact. Available Cultures: Indiana 2002-4 Texas 1998 North Carolina 1978 Ontario Canada 1980-6 Ontario Canada 2001 N. Ireland 1977 Germany 1989 Japan 1989-2002 China-PRC 1991 Deutsch Germany 1989 Deutsch Germany 2007 institutions malesfemales Use of data noted here is free for researchers. Fees apply for commercial applications.
Or download an Excel file of 5001 words rated in English. At: http://www.indiana.edu/~socpsy/public_files/EnglishWords_EPAs.xlsx
Have any 3-D sentiment analyses been done? Some related research. Heise, D. R. (1966). "Sensitization of verbal response- dispositions by n Affiliation and n Achievement." Journal of Verbal Learning and Verbal Behavior 5: 522-525. Computed the distance between EPA profiles of 1,000 words and EPA profiles representing need-for-affiliation and need-for-achievement. Result: each word scored for each need. 137 Ss wrote stories in response to four pictures. Need-scores of words in an S’s stories were averaged to score the individual on each need. Results. n-affiliation: Text-based scores correlated 0.64 with standard method of scoring needs. n-achievement: Text-based scores correlated 0.36 with standard method, and 0.30 with an experimental measure of the need. Significance: text analyses, p <.001; experiment, p <.05. Anderson, C. W., and G. E. McMaster, 1982. Computer assisted modeling of affective tone in written documents. Computers and the Humanities 16: 1-9. Literary tension = words’ Activity minus Evaluation. Plot of tension in Joyce Carol Oates’ Wonderland, chapter where son discovers his father has murdered whole family. Same pattern of bad-activation characterizes the dramatic peak in children’s stories like Peter Rabbit. That was my dissertation! sees murders
The End! Recent ACT publications by Heise. Heise, David R. (2007). Expressive Order: Confirming Sentiments in Social Actions. New York: Springer. – Affect control theory, its research program, and its math model. Heise, David R. (2010). Surveying Cultures: Discovering Shared Conceptions and Sentiments. Hoboken, NJ: Wiley Interscience. – Reviews research on EPA, and methodology of EPA measurement. MacKinnon, Neil J., & Heise, David R. (2010). Self, Identity, and Social Institutions. New York: Palgrave. – Expansion of affect control theory to the topics in the title. Heise, David R. (2013). Modeling interactions in small groups, Social Psychology Quarterly. March issue. – Application of affect control theory to small-group interactions. This PowerPoint presentation can be downloaded from: http://www.indiana.edu/~socpsy/public_files/DimensionsOfSentimentAnalysis.pptx