UNESCO & FILLM
HOME | ABOUT | NEWS | MEMBERSHIP | EVENTS | PUBLICATIONS | BLOGS | CONTACT
MENU BLOGS

Blogs

Bloggers

Archive

Tags
Is Sentiment Analysis Really About Sentiment? A Call for a Multidisciplinary Approach

| 28 August 2014 | Sebastian Feller


A couple of months ago I was asked to advise a research team on the topic of sentiment analysis. Interesting topic, I thought. I had actually taken a class on sentiment analysis on Coursera a while back and thus had a rough idea on what this whole sentiment business was all about. To be frank, I could not remember much from the Coursera class other than that there was lots of statistics involved. Countless formulas for calculating all kinds of probabilities and eventualities. What I did remember very clearly though was my rather negative sentiment regarding how sentiment analysis is supposed to work. Ironic, isnít it? Oh, letís not even go there. Irony is one of the many terrible things that sentiment analysts (or however you might want to call yourself in this business) would like to ban for good. No more ironic talk, for nobody, please. Why? Well, we leave that for later.

Anyhow, so I got myself into the sentiment analysis business and tried to make a meaningful contribution. I got invited to a team meeting. The first thing I said was: Donít come with statistics. People donít calculate meaning, why should machines do? Okay, I do admit that statistics has never been my favorite subject and lots of it seems to lie outside of my intellectual well-being. Anyways, let me point out some more objective reasons for hating statistics (Oh my, please donít run this through your sentiment analysis tool. I am afraid you will diagnose me with depression. Okay, letís delete ďhatingĒ and replace it by ďnot likingĒ). First, statistics is about quantities but language is of a purely qualitative nature. Yes, you can count words in a text and calculate all kinds of scores like mutual information and collocation scores. However, what words really mean is by no means captured by these scores. Second, statistics assumes randomness. Statistical tests tell us if what we observe is something that we would expect from chance or not. But isnít this a rather funny assumption towards language use? Does chance really play a role in how people use language? I would say thereís a fairly high chance that not (Wait: performative contradiction! Another candidate for the sentiment analysis censorship.).

The next thing I said was: How do you expect to understand sentiment from looking at the text only? In language use, we donít express meaning only through words. How about the face? (If you could look at my face right now, you would get a fairly good idea of how I feel about this kind of trust-the-text mentality.) People make all kinds of faces while they speak and in doing so they nuance the meaning of the words they use, especially in terms of sentiment. Paul Ekman made a fortune on this insight. How about our hands? There is a whole branch of psychology and linguistics devoted to the study of gesture. All this seems to have little to no relevance to sentiment analysis. Interesting, to say it mildly.

I could go on and on and on like this, but I believe you got the point by now. Sentiment analysis (at least to the extent that I have encountered it so far) needs to be revamped. Experts from disciplines like computer science, linguistics, psychology, communication studies, and intercultural studies, to name a few, have to work on this together, if a machine should ever be able to understand sentiment in a meaningful way. You donít go to the butcher to buy potatoes. Applying the same logic, you should not expect the computer scientist to analyze and represent meaning. Sentiment analysis is a horribly complex task and it needs a ďhorriblyĒ complex team to get the job done. What I can say so far is that we have made a start. After a couple of team meetings, I feel that the idea has eventually come across and the team is now opening up to more colleagues from different disciplines. I will keep you posted on how we are doing.

Oh, by the way, what does LIWC say about my sentiment? I submitted this post to LIWC online. Here the results. Go figure!

LIWC Results

LIWC Dimension Your
Data
Personal
Texts
Formal
Texts
Self-references (I, me, my) 5.04 11.4 4.2
Social words 7.19 9.5 8.0
Positive emotions 1.87 2.7 2.6
Negative emotions 1.15 2.6 1.6
Overall cognitive words 10.79 7.8 5.4
Articles (a, an, the) 6.47 5.0 7.2
Big words (> 6 letters) 22.16 13.1 19.6

The text you submitted was 695 words in lenght.

Comments: fellers(a)ihpc.a-star.edu.sg

Tags: sentiment analysis, statistics, irony, LIWC, face, machine, multidisciplinary

ABOUT FILLM

About Us


Mission and Strategy

Constitution

Committee

Former Officers

FILLM Logo
NEWS

News


Archive
MEMBERSHIP

Membership


Members
EVENTS

Forthcoming

FILLM Events

Member Events

PUBLICATIONS

Recent


Book Series

Congress Proceedings

FILLM Newsletter

Others
BLOGS

Latest Posts


Bloggers


Archive

Tags
CONTACT

Webmaster


FILLM Committee

Members