Provided by Cognitive Sciences ePrint Archive
Automatic Dream Sentiment Analysis
David Nadeau1, 3, Catherine Sabourin2, Joseph De Koninck2, Stan Matwin1 and Peter D. Turney3
1School of Information Technology and Engineering
2School of Psychology
University of Ottawa
Ottawa, Canada
3Institute for Information Technology
National Research Council Canada
Gatineau and Ottawa, Canada
{david.nadeau, peter.turney}@nrc-cnrc.gc.ca, {jdekonin, stan, csabo069}@uottawa.ca
Abstract
In this position paper, we propose a first step toward
automatic analysis of sentiments in dreams. 100 dreams
were sampled from a dream bank created for a normative
study of dreams. Two human judges assigned a score to
describe dream sentiments. We ran four baseline algorithms
in an attempt to automate the rating of sentiments in dreams.
Particularly, we compared the General Inquirer (GI) tool,
the Linguistic Inquiry and Word Count (LIWC), a weighted
version of the GI lexicon and of the HM lexicon and a
standard bag-of-words. We show that machine learning
allows automating the human judgment with accuracy
superior to majority class choice.
Introduction
Research in psychology shows that emotion is a prominent
feature of dreams [2], [6], [11]. Typically, the level of
emotions, or sentiments, is assessed in dreams by content
analysis made by human judges using scales of various
levels, or by dreamers themselves. In this work, we show
how to automatically obtain equivalent measures. We used
a value from 0 to 3 to estimate both the positive and
negative content of dreams, as applied by independent
judges and we compared it to an automatic analysis.
The granularity of our scale (4 levels) was chosen to reflect
the variety of sentiment experience and to maintain
simplicity. One envisioned application of this measurement
is the assessment of the stress experienced by the dreamer.
Previous work aiming at drawing a link between negative
sentiments in dreams and dreamer stress relied on content
analysis of written dreams [1].
A more general application of automatically analyzing
dream sentiments would be the mining of large dream
banks and discovery of unsuspected data about sentiments
in dreams of individual of different age, social status, etc.
From a machine learning perspective, the task of dream
sentiment analysis is expressed as a classification problem
with labels {0, 1, 2, 3}. The goal of this work is to create a
system that can reliably replace human in analyzing
sentiments in dreams.
The next three sections go as follow: first, the dream
corpus is detailed, then our experiments in automatic
dream sentiment analysis are presented and, finally, related
works are discussed.
Dream Bank
Dreams were gathered from a dream bank created during a
normative study conducted at the Sleep Research
Laboratory of the University of Ottawa (UofO). The ethics
committee of UofO has approved this normative study as
well as the use of the dream bank for future studies.
Volunteers were informed that their dreams could be used
in other studies on dreams and they all gave their consent.
Their participation mainly consist of completing a brief
dream diary at home during a maximum of three weeks,
and to write down all the dreams they remembered when
waking up, until a maximum of four dreams. A sample of
100 dreams, from 29 individuals of varied age and sex, was
used in this study.
Manual Sentiment Analysis
The second author of this paper annotated the 100 dreams
with two scores ranging from 0-3. One score is for the
positive orientation of the dream and the other one is for its
negative orientation. The third author of this paper
independently annotated 26 dreams. With this second
annotation, we calculated the inter-judge agreement shown
in Table 1. We also report the mean squared error (MSE)
on the agreement. MSE is presented and discussed in the
result section. Judges based their rating on example dream
passages like in Table 2.
Scale Inter-judge agreement_______MSE
Positive 57.7% 0.54
Negative 80.8%0.19
Table 1: Inter-judge agreement on 26 dreams.
At this point, we dropped the positive scale. The reason is
twofold. First, the agreement between annotators is too low