Automatic Dream Sentiment Analysis

We parsed each dream using Balie and count each time the
token canonic version exactly match an entry of the GI
lexicon or the HM lexicon. We choose to use the following
features for both lexicon (GI and HM):

1. the sum of positive weights

2. the sum of negative weights

3. the average of positive weights

4. the average of negative weights

5. the maximal positive weight

6. the maximal negative weight

7. the negative orientation level {0,1,2,3}

For each pair (1-2, 3-4 and 5-6), we also add a feature for
the difference and a feature for the log ratio.

The Bag-of-Words

We experiment a Bag-of-words (Bow) approach as a fourth
strategy to classify dreams. The Bow approach consists in
using as feature every unique word appearing in any
dream. Our dream sample is composed of 2758 unique
tokens that turns out to be 2758 features. A particular
dream (a textual document) is represented by a Boolean
vector of length 2758 for which the value of element j is 1
if the token j appears in the document, and 0 otherwise.
This technique is often used in text classification. It allows
linking a class (ex.: 1, on the negative scale) to some
specific words (ex.: dark, cold, night, etc.)

Results

Two metrics are required in our experiments. First, we
calculate classifiers accuracy - the sum of correct guesses
over the total number of guesses - i.e. their performance at
exactly finding the right label (e.g., human rates 3,
machine guess 3). Second, we calculate the mean squared
error of classifier - the average of the squares of the
differences between the human labels and the machine
predictions. This metric is low when a classifier guesses
near the human (e.g., human rates 3, machine guesses 2)
and becomes high if the classifier is far from human
judgment (e.g., human rates 3, machine guesses 0).

In Table 3, we report the accuracy percentage (ACC) and
mean squared error (MSE) of every strategy. Results are
for stratified 10-fold cross-validations.

Table 3: Accuracy and mean squared error of various
strategies on analysis of dream negative sentiments.

The baseline accuracy is given by a classifier that always
guesses the majority class. In our dataset, 33% of dreams
were rated with label “2” and this is the majority class.
Guessing always “2” results in 33% accuracy. The baseline
mean squared error is given by a classifier that always
guesses the average of classes. The average of all classes is
1.37 in our dataset. It results in a mean squared error of
0.993.

Features from the General Inquirer outperform other
strategies accuracy (highest number of correct guesses) and
mean squared error (lowest difference with human
judgment when incorrectly guessing). LIWC is considered
as good as GI since there is no statistically significant
difference between both resources.

We tried many different supervised learning algorithms,
but the best result was linear regression. Standard
classification algorithms have the downside of resulting in
bad mean squared errors. In Table 3, the last column
(BOW) is for a Naι've Bayes algorithm known to perform
well in text classification. Even if the accuracy is not the
lowest, the mean squared error is the worst.

Discussion

The best features to automate dream sentiment analysis are
from the GI tool [10] and the LIWC tool [9]. We believe
this constitutes a significant first step in this field. Even if
50% of accuracy may appear to be a poor score, it is
statistically better than the baseline accuracy (majority
class guessing) with 95% confidence.

The MSE of 0.577 for an accuracy of 50% means that most
errors have a difference of 1 on the scale (e.g.: human rates
3, machine guesses 2). As a matter of comparison, if every
error was for a difference of 1, it would result in a MSE of
0.5 (50 dreams out of 100 with an error of 1 or -1 = 50 time
a squared error of 1 out of 100 = mean squared error of
0.5). If every error was for a difference of 2, the MSE
would be 2.

Related Works

Dream Analysis in Psychology

Sentiment analysis is an important component for the
studies of dreams since emotions are considered by many
as responsible for structuring the content of dreams [4],
[8]. Recent findings from brain imaging studies have
shown an increased activation of limbic and paralimbic
areas during Rapid-Eye Movement (REM) sleep [7].
Dreams being strongly associated with this sleep phase,
this may account for the emotional intensity of dreams [2].
However, further studies are still needed to better
understand the origin as well as the potential role of the
emotionality of dreams.

Until now, most of the recent studies on dreams use the
classical scales of Hall and Van de Castle [3], which are
considered as being the most detailed and complete coding
system available for scoring dreams [2]. It comprises
various scales measuring both positive and negative
content, such as the presence of friendly or aggressive
interactions, emotions, good fortunes or misfortunes, and
successes or failures. However, this system is time
consuming and depends on the rater’s judgment. It is of

<<   <   1   2   3   4