The emotional profile of words. Tyler Schnoebelen Stanford University http://www.stanford.edu/~tylers/emotions . Welcome!. If you’re reading the presentation online, please make sure to check out the notes fields—they have content that supplements/explains the content of the main slides.
Tyler Schnoebelen Stanford University http://www.stanford.edu/~tylers/emotions
Welcome! If you’re reading the presentation online, please make sure to check out the notes fields—they have content that supplements/explains the content of the main slides. Any questions or comments? Please send them to me! TylerS at Stanford dot edu.
Some claims Linguistics is about understanding human beings. To understand human beings is to understand the variety and complexity of emotional experiences they have. Linguistics can offer a lot by showing how linguistic resources are used in creating and coping with these experiences.
Goals Immediate goal: What are the conceptual tools and actual methods to analyze language in terms of emotion? Bigger goals: What is the structure of the affective lexicon in English? What about other languages?
Classifying Classification structures our body of knowledge. Especially indispensable when it’s not clear where to begin.
Typical sentiment analysis
Approach, assumptions, hypotheses Words are not just positive/negative Sentences are not just objective/subjective, either Words are better characterized by the types of emotional work they do Items that occur in similar situations are alike Though not in ways we understand yet Placement in the network dictates which semantic shifts are possible
The Experience Project 31,675 “confession” stories Tagged by readers with 27,187 tags I understand 11,277 You rock 3,781 Sorry, hugs 3,733 Teehee 3,545 Wow, just wow 916
An example Are you missing me? Do you wish you could reach out and touch my face. Are you tossing and turning, or pacing back and forth? You know where I am and that I am alone. But what you may not know is how much I need you.
Monroe et al 2009 Log-odds with Bayesian priors for a given word, in a given topic, do Democrats/Republics use it a lot more taking into consideration how they use words OUTSIDE that topic
Other categories “Sorry, hugs” “I understand” feel hate wish cancer died sorry “You rock” amazing angel beautiful
But is there structure? There are 111 uses of omg There are 2,332,769 tokens total, so omgmakes up 4.75e-05 of the data There are 456,454 words marked for “hugs”, so if everything was random we’d expect: 456,454 * 4.75e-05 = 21.72 occurrences of omg in the “hugs” category Actually, there are 14—smaller than we’d expect Is it significantly smaller? We use a g-test:
Restrictions I restrict myself to words that have 10 or more instances overall Observed/Expected scores that are unusual .5 standard deviation outside the mean for at least one category of a word Are statistically significant by g-test POS tagging keeps adjectives, adverbs, and discourse items (wow, omg, !, yup) 1,297 words
Constraints We’ll set wow, just wow aside due to sparseness of the data and focus on the other four categories. If we reduce down to just over-representation, under-representation and “as expected”, there are logically 3^4=81 patterns Well, “-1, -1, -1, -1” and ‘1, 1, 1, 1” aren’t actually possible And the technique describe above gets rid of “0, 0, 0, 0”. But if all things were equally, we would still have 78 patterns
We don’t Instead of 78 patterns, there are 42 Nor are words evenly divided
IMDB—Internet Movie DataBase 45,772 movies 1.36 million reviews 29 genres 1-10 stars (And “no rating”)
Structure of stars
Probabilities Similar to what we did with the Experience Project, but an extra layer: Expected=p(genre)*p(star-rating)*p(word)*total corpus size Restricted to 12 genres that don’t overlap Combined with Experience Project data
Words can be infected… Semantic dynamics, changes of meaning E.g., semantic contempt can creep in to things from their use (Kaplan 1999)
3 types of emotional meaning Valence—positive/pleasant vs. negative/unpleasant “happy” vs. “sad” “psyched” vs. “disappointed” Me-ness Primarily about the speaker, identity work Do ideophones go here, perhaps? Us-ness Relational—moving interlocutors closer or further away Most intense emotional reactions will happen here
Us-ness Examples: Pitch in Zapotec (Sicoli 2007) Breathiness in Japanese (Campbell 2004) Kinyarwanda diminutives Diphthongs in Czech Palatalization in Yiddish and Basque Impersonating your listener
Us-ness Every context has a valence, it’s quickly set and assessed Positive valence + us-ness: Intimacy Negative valence + us-ness: Pejoration Teasing and joking can bring us closer—or push us apart.
What does this mean for fieldwork? Most African languages don’t have enormous corpora like English But the “observed vs. expected” thinking will keep you attuned to many different features, even if your counts are low. And the patterns in Well-Studied-Languages-with-Giant-Corpora can serve as a template What are the linguistic resources (lexical items as well as pitch, syntax, etc) that are used to convey these meanings? How do valence, me-ness, and us-ness get conveyed? What is the distribution of items in different categories? Are the categories themselves different?
What I’m hoping for You’ll get interested in adding emotion to your own fieldwork You’ll tell me what you’ve already observed (Especially about diminutives, by the way)
Thanks for listening! http://www.stanford.edu/~tylers/emotions
A world experienced without any affect would be a pallid, meaningless world. We would know that things happened, but we could not care whether they did or not. (Tomkins, 1995, p. 88)
Some questions Don’t worry about assigning a particular emotion—broadly speaking is a positive or negative emotion being communicated? How explicit are the emotions? If the utterance isn’t as explicit as “I am happy”, does it indicate behavior or disposition (the difference between laughing at something and liking something)? Is the emotion directed (the girl liked the teacher) or more mood-like (the girl was happy)? With ideophones, I suspect the object is often what the emotion is directed at. How intense is it? (Like vs. love vs. adore)
Affective linguistic resources We can adopt many of the characteristics Potts (2007) uses to describe expressives Epithets like the jerk, attributive adjectives like damn, honorifics, wow, affective demonstratives (that woman), etc. I’m ultimately interested in linguistic resources outside words and phrases (think pitch, speech rate, voice quality), but for now, let’s stick to words What are affective linguistic resources like? They have an immediate and powerful impact on the context. They are performative. They are revealing of the perspective from which the utterance is made, and they can have a dramatic impact on how current and future utterances are perceived. People can’t easily articulate their meaning. They are volatile. They are indispensible to language.
Not so crucial
Word choice matters Evidence suggests that people’s physical and mental health are correlated with the words they use Gottschalk & Glaser (1969); Rosenberg & Tucker (1978); Stiles (1992) “Word use is a meaningful marker and occasional mediator of natural social and personality processes.” (Pennebaker et al 2003: 548)
Way fewer Over-representation of hugs and teehee buddy, offered, answered, shock, ran, ? Under-representation of hugs and rock excuses, license Over-representation of teehee, under hugs and rock gross, sexual, mary, creepy, h, computer
Fewer than expected Under hugs and teehee expectations, soulmate, friendship, beliefs, passion, shame, graduate, belive, experiences Under rock and teehee depressing, paranoid, ignoring, depressed, disappointing, pills, terribly, terrified, financially, afford, betrayed, commit, disorder, dying, emotional, failure, fault, freak, hurts, husband, jobs, loser, mental, ocd, overweight, suicidal, unable, unhappy, worse, considered Over teehee, under rock bus, eight, embarrassing, idiot, boyfriends, course, cousins, disgusting, dry, freakin, gf, laundry, lazy, myspace, roommate, she’ll, shirt, smells, stuck, wtf, yell, blamed, dirty, hook, kissing, photos, police, stole, threw, upstairs Over rock, teehee anonymous, lame, blocked, ears, flowers, pm, rules, across, bag, blog, box, cars, dance, dress, extra, faces, gonna, hotel, information, lil, rich, shopping, train, welcome, yeah, air, coffee, fire, glass, war Over hugs, teehee, under rock military
More than expected Over rock and teehee, under hugs and understand awesome, lol, ha, anon, ya, honey, p, ladies, spell, makeup, www, soft, bra, pink, rain, com, http, obama Obviously, there’s some collocation stuff happening.
Quick review Patterns of use Compare observed vs. expected Cluster like-with-like (and “more-like”) Conceptual baggage and indexical fields Maps, overlays Indeterminacy Shifts and changes
3 basic techniques
Why don’t people go around diminishing other people?
Laws of power It is better to have more power than less. Corollary: Acts that narrow the power differential are friendly when performed by the more powerful. When the powerful reaffirm or widen the power differential, the act can be seen as hostile. Corollary: Acts that directly reaffirm the power differential are friendly when performed by the less powerful. When the less powerful narrow the power differential, the act can be seen as hostile. Friendliness brings interlocutors together, hostility pushes them apart. Corollary: "Friendliness can be suspicious"—if an act lowers a speaker or raises the listener, it may be interpreted with skepticism. Corollary: Hostility (and suspicion of hostility) from the less powerful leads to sanction/punishment from the more powerful. Corollary: Hostility (and suspicion of hostility) from the more powerful leads to resistance from the less powerful.
“Tradition tends to protect itself”
Some examples Minimize self/requests: Might help you feel (a little) better about yourself too. I think you just answered your own question a (little) bit: you stay around those people because you feel like it's work to meet someone new and life wears you out. Minimize others: Two (little) Chinese women made this beautiful arrangement, but it was huge and it fell! She has done nothing but love me and my child but despite this I have been screwing a hot (little) girl, young enough to be my daughter. Minimize interlocutor: Oh you think you're big enough to make me, you little wimp...Come on, come over here and make me, I dare you… You little fruitcake, you little fruitcake, I said you are a fruitcake.