Skip to main content

My Dissertation, Natural Language Description of Emotion: General Theoretical Issues

dissertation
emotions
language
research
Uncategorized
Author

Abe Kazemzadeh

Published

August 13, 2012

Dissertations, especially in computer science, can be very technical and hard to explain to people who are not other computer scientists.  However, I’m trying to explain my dissertation topic for people who do not have any special training, and I thought a blog would be a good place to do this because blogs are very informal.  I’m not sure how many people will actually read this, but if it’s not entirely clear to you, please write me a comment and let me know how I can improve it.  Also, if you are technically inclined, feel free to comment as well.

The field of my dissertation is in natural language processing, which is a subfield of computer science that aims to make computer programs that can “process” human language. By “process” I mean that the computer should be able to understand human language in some limited, computational sense.  Some examples of natural language processing are:

Hopefully these examples provide an overview of natural language processing, which is also known as NLP for short.  My dissertation topic fits in a couple of these categories.  First of all, it is about emotions, so it is in the affective computing subfield of NLP.  It is also about the meaning of language about emotions, so it is also in the natural language understanding or semantic analysis fields.  It also deals with a dialog agent that asks questions, so it can be seen as tangentially in the dialog system subfield.

Now that I’ve explained a bit about the general area of my dissertation topic, let me say some more about how it is different and new.  In affective computing and sentiment analysis, most previous work has focused on determining the writer’s emotion or the emotion that the writing will provoke in the reader.  I call this “expression of emotions” in language.  My work focuses on description of emotions, which are slightly different.  Description of emotion is when people talk (or write) about emotions.  The reason why this is interesting and important is that I can talk about sadness, despair, resentment, jealousy, rage, vindication, pride, ecstasy, relief, and cheerfulness without being or feeling any of these.  The previous sentence would have seriously thrown off any algorithm that is trying to determine my emotional state, or the emotion that is inspired in you, the reader.  So I picked this area, with the help of my professors and labmates, which we call Natural Language Description of Emotion, as a topic to study.  Another reason why this area is interesting and important is that animals are able to express their emotions, but humans are the only species I know of that can describe and name emotions (that may or may not be their own current emotions).  This is an important ability of humans because it allows a great deal of introspection and allows us to think about emotions hypothetically. On the downside, we humans aren’t only feeling emotions “in the moment” like animals do, but imagining about, gossiping about, and projecting emotions, and more.  Sometimes I feel like writing a self-help book when I finish my dissertation!

So now having a topic,  that is somewhat interesting and important and different from previous work, I needed a thesis statement.  Now that I have given (hopefully) a gentle introduction, I need to unleash the technical jargon.  It is not as complicated as it sounds, but I’m going to write it now to inoculate the reader:

Natural Language Descriptions of Emotion are Definite Descriptions that refer to Intersubjective Theoretical Entities.

See, that wasn’t that bad was it?  It sounds very technical but I can explain it fairly easily if you, can patiently bear with me.  I’ve already explained what natural language descriptions of emotion are.  The next thing to explain is definite descriptions.  The notion of definite descriptions was discovered by Bertrand Russell, a philosopher, logician, and mathematician, among other things.  Definite descriptions capture how people can talk about a specific, unique object.  If I see a rabbit and point to it and say “rabbit”, you know that the meaning of “rabbit” is the object that I’m pointing at.  However, we can talk about objects that are not physically present and even though they are not physically present, we can talk about them as if they were a unique object.  The form of a description that refers to a unique object or entity is what is known as a definite description.  An expression in the form of “the” + noun, that is a noun preceded by the word “the”,  is a definite description.  Proper names like “Abe Kazemzadeh” or “Bertrand Russell” are also examples.

So why are we talking about “definite descriptions” again? First, I’m claiming in the thesis statement that when we talk about and describe emotions we are actually referring to something that is specifically and uniquely describible.  We can claim that “happiness” is distinct from “elation” based on how we describe them.  Also, we can claim that the happiness I feel is the same as the happiness that you, the reader, feels, even though we have a different personal histories and perhaps some differences in our brain chemistry.  So this is the idea that I’m trying to capture when I say that natural language descriptions of emotion are definite descriptions.

The other reason why I introduce the idea of definite descriptions into my dissertation is more technical, so I won’t go over that in too much detail, but apart from the details it’s not that hard to understand.  From a technical point of view, if we want to reduce statements in human language to logic, we need to consider words as logical variables, and in logic variables must refer to objects or sets of objects.  However, in natural language we have words that refer to descriptions rather than objects.  The technical aspect of definite descriptions allows variables to stand for descriptions rather than objects or sets of objects.  The whole business of reducing language to logic is important because computers are essentially logic machines and to get them to understand anything, you need to reduce it to logic.

This is a good point to introduce the last part of the thesis statement “intersubjective theoretical objects”.  The idea of theoretical objects was discovered by Rudolf Carnap, a philosopher of science.  He was involved with trying to describe things logically, unambiguously, and in a way that was empirically verifiable.  The way he proposed doing this, was to only talk about things that are physically observable.  If you try to do this, you’ll realize that is is very impractical, but it was a provocative idea.  If you can only talk about physical objects, then everything you say can be scientifically tested.  However, even scientists talk about imaginary, unobservable things: for example, the idea of an atom was essentially a such a fairy tale for more than two thousand years, from the time that Democritus, the ancient Greek, proposed it, to the time that it was physically studied in laboratories in the late 1800’s and early 1900’s.  So actually, these unobservable objects (at least before a certain point that they become observable), are actually pretty important to science because they are like the cutting-edge hypotheses that need to be tested.  These type of objects that are not observable were called “theoretical objects (or entities)” by Carnap.

In saying that natural language descriptions of emotion are definite descriptions that refer to theoretical objects, what I mean is that we may not (currently) be able to observe and test our language about emotions, but at a certain point, we may be able to (for example if there is some kind of brain physiology that gets discovered).  Currently, we don’t know if happiness, anger, sadness, etc. are unique physiological states, but we are able to talk about them as if they were.

I actually think that some emotion descriptions will never be reduced to an observable (e.g. physiological) basis, and that’s why I say that natural language descriptions refer to intersubjective theoretical entities. What I mean by this is that the experience of an emotion may be subjective to a given person, but that the meaning of the words describing the emotion are understood across subjects, i.e. intersubjectively.  Furthermore, even if we could observe the subject’s internal state, it wouldn’t matter for some emotion descriptions because the meaning of the descriptions are determined by the conventional meaning of language (i.e., being brought up in a linguistic community), and  not determined by  some internal state.

To better illustrate intersubjectivity, compare the words “red”, “rabbit”, and “bachelor” (I took these examples from a book called Word And Object by Willard Van Orman Quine) .  “Red” refers directly to sensory experience, so it is subjective, but we understand about how eyes work so we know that there is some biological similarities in how we experience “red”.  Also, the “edges” of the meaning of “red” are gradual: if you slowly vary the shade/hue, you might slowly make something into not red, but it will be gradual, not abrupt.  “Rabbit” refers to a physical object, so it’s core meaning is not subjective, although people might have different subjective associations about rabbits, e.g., if they have had a pet.  “Bachelor” is unobservable and subjective from a sensory point of view.  For a given person, they will be able to determine who is a bachelor given their social relations.  However, for a random person outside of their social relations, they will not be able to say whether they are bachelors or not (for a female, they will be able to say for sure that she is not, but for a random male, they can only estimate, say given age, what is the likelihood of being a bachelor).  However, if not restricted to sensory data, a person can simply ask if an unknown male human is married or not.  The core meaning of the word “bachelor” is not established by observation but rather by language. This language-established-agreement-in-meaning is what intersubjectivity is.   In saying that descriptions of emotion are intersubjective, what I’m saying is that what we describe as an emotion is as much determined by our language as it is by our physical observations.  So of the three words, “red”, “rabbit”, and “bachelor”, natural language descriptions of emotion are most similar to “bachelor” in terms of the type of subjectivity they contain.

So now you should understand the thesis statement, the main claim of my dissertation.  This part is the theoretical part.  There are two more broad parts: the computational part, and the experimental part.  The computational part involves making a model that specifies the details of the thesis statement into a computer program.  The experimental part involves testing the program in the wild, as it interacts with people.  I’ll write about these other aspects of my dissertation shortly…


Originally published on abewrites.

clustrmaps.html