English. In this paper we present the results of a gender detection experiment carried out on a corpus we built downloading dream tales from a blog. We also highlight stylistic differences and similarities concerning lexical choices between men and women. In order to carry the experiment we built a feed-forward neural network with traditional sparse n-hot encoding using the Keras open source library.
It is generally accepted that dreams are just an unconscious production, and that represent a type of non-manipulable happening. However, many people believe that dreams are premonitory of future events as well as representations and reworkings of past events. Humans tend to preserve all personal events, some of them in the form of a diary, namely the best method to tell an event and keep its aura of magic.
Until recently, dream reports were relegated to the
the pages of paper journals or revealed to
familiar people. At an earlier time, dreams are gathered
from sleep research labs, psycho-therapeutic and
in patient settings, personal dream journals and
occasionally classroom settings where “most recent
dreams” and “most vivid dreams” are collected as
in
Social media have opened millions of pages where people feel at ease to confess their thoughts, their experience and even their secret fantasies. These platforms such as Twitter, Facebook and web blogs are a good ground for computational text analysis research in social science and mental health assessment via language.
Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Diary narratives represent a field already investigated by researchers. The recent development of web communities focused on telling dreams allows researchers to access and discover new characteristics related to the language of dreams. Stylistic and linguistic features of dreams in blog reports are essential in order to detect writing style and content differences between men and women, but also enable future researches associated to the different types of personality and styles associated with mental health diagnoses and therapeutic outcomes.
The aim of this paper is to show that despite dreams are just an unconscious production, there are several stylistic differences between the reports of dreams by males and females on online blogs. The model we built is able to represent and classify all stylistic differences.
Moreover, this research represents a preliminary step in the field of dream tales which will be followed by an attempt to find stylistic differences between dream tales and other forms of self narration (i.e. travel tales).
The paper is organized as follows: in Section 2 we introduce Related Work, in Section 3 we describe the corpus we built and the blog. Methodology is described in Section 4 and Results are in Section 5. In Section 6 we present our Conclusions and we introduce Future Work. 2
Textual analysis of dream reports is still not a
completely investigated field in NLP. One of the
purposes of computational dream report analysis lies
in understanding how and why a dream narrative
differs from a waking narrative
On the basis of this approach (Bulkley, 2014) update the categories list and evaluate it on four datasets of the DreamBank corpus. It has been shown that this type of word analysis can be applied to detect the topics of dreams. In addition, this latter contribution provides evidence that it is possible to guess about a person’s life and activities, personal concerns and interests based on an individual dream collection .
Other works focus on identifying the emotions in the reports of dreams. In particular (Razav et al., 2014) use a machine learning method to assign emotion labels to dreams on a four-level negative/positive sentiment scale. In their research, dreams are represented as word vectors and dynamic features are included to represent sentiment changes in dream descriptions.
In a more accurate sentiment analysis,
In psychiatric studies, the gender variable is
identified as a predictive for psychotic behaviors and
disorders. In
Dream diaries refine the research in uncovering connections between dreams and dreamer’s sociocultural background, mental conditions and neurophysiological factors. The language of online dreams in relation to mental health conditions has yet to be analyzed, however prior laboratory research suggests that dream content may differ according to clinical conditions.
In
In
The web is full of blogs, where people can share opinions, questions and personal feelings and thoughts about their own life. Furthermore, people also share their dreams, one of the most personal hidden aspects of life.
It is very easy to find a blog in which thousands
of people share their “dream experiences”,
sometimes discovering that other people have had
similar experiences dictated by similar life styles.
We investigated a blog, called SogniLucidi, on
which every day thousands of people tell their
dreams and nightmares, mixing their nightly
fantasies with their unconscious writing style choices.
SogniLucidi, that literally can be translated in
LucidDreams took its name from a term coined by
the Dutch psychiatrist Frederik
There are many techniques that, when correctly applied, allow dreamers to obtain a “Lucid Dream” and that we report for completeness: CAT (Cycle Adjustment Technique), MILD (Mnemonic Induction of Lucid Dreaming), WBTB (Wake Back To Bed), WILD (Wake Initiated Lucid Dreams), RCT (Reality Control Test) and ITES (Induction Through External Stimulus).
The corpus we built for the investigation is balanced with gender and the number of authors analyzed is not randomly selected but represents the precise number of participants to the blog. 3.1
In this paragraph, we present the resulting statistics obtained using the NLTK module together with other statistics formulas for the analysis of the corpus we built on SogniLucidi blog. In Table 1 we report two important statistics about words: the number of tokens in texts written by men and women and word types. We can notice that there is a big difference in the number of tokens used by Males (80629) and Females (57673).
80629 12254
57673 11158 In Tables 2 and 3 we present four lists of six exclusive nouns and six exclusive verbs used by men or women. Both exclusive nouns and exclusive verbs are the most relevant for frequency for Males and Females classes. Verbs are reported in their base form. The results indicate, without interpretative effort for a human, that most relevant topics given these high frequency words are associated to activities and events that the dreamers want to happen, in settings and adventurous situations for male dreamers. Meanwhile dreamers belonging to Females class seem to set their dreams in a baleful scenario, where “transizione” (transition) and “trapasso” (transition) mean that they dream about twilight state, beyond death or they fantasize about surreal activities.
Males destinazione (destination) esplosione (explosion) foresta (wood) lenzuola (linens) spiaggia (beach) terrazze (terraces)
Females balzo (bound)
luce (light) nuvola (cloud) piscina (swimming pool) transizione (transition) trapasso (transition) The training corpus consists in dream text descriptions written by two groups of authors: 28 Female authors.
The corpus is balanced and labelled with gender. Gender annotation has been done manually and based on the name of the users, their profile photos and description. For each author, a total of fifteen texts about dreams are provided. Authors are coded with an alpha-numeric author-ID. For each author, the last fifteen texts about dreams have been retrieved from the personal web diary’s timeline. As a result, the time frame of the dream reports might vary from days to months, depending on how frequently users report their dreams on the blog. To train our classification model, we exploited the descriptions of dreams only and not the comments (both comments of the authors and comments of other members of the SogniLucidi blog). 4.1
For preprocessing we used the Python library BeautifulSoup along with same regex procedures. We performed the following preprocessing steps:
Detecting stop-words by document frequency and removing. Only n-grams that occurred in all documents has been considered a stop-word and ignored. 4.2
Feature selection is a very critical step in any model. For feature selection we use the sklearn utilities SelectKbest. It selects the n-best feature based on a given criterion. In our experiments, the features are selected on the f classif criteria. This function perform an ANOVA test, a type of hypothesis test, on each feature on its own and assign that feature a p-value. The SelectKbest rank the features by that p-value and keep only the nbest features. The feature set for the dream dataset benefits from word trigrams in addition to other ngrams. In our final model, we use the following ngrams features: Word unigrams, bigrams and trigrams.
Word level n-grams used the following parameters:
Minimum document frequency = 2. Terms with a document frequency lower than would be ignored; Term frequency-inverse document frequency (tf-idf) weighting; Maximum document frequency = 1.0 or rather terms that occur in all documents would be ignored.
We built a neural network to perform the gender detection issue. We decided to run a feed-forward neural network with traditional sparse one-hot encoding with the Keras open source library. After a parameters selection, the model obtained the best performance with an Adam optimizer and a learning rate of 0.32, feeding it with a batch size of seventy and training for thirty epochs. Moreover, the input layer of sixty-five neurons with an initialization using a norm kernel. Then, a RELU activation function was applied, followed by a dropout layer. During optimization, we found that a relatively big dropout rate of 0.5 outperformed the smaller dropout rates. The output layer is a single neuron, followed by a linear activation function. The feature set provided to the model was an nhot encoding of the uni-, bi- and trigrams. 5
In this section we describe the results on the training data and the test data. The data we used was split into training and test data. The training set contains a known output and the model learns on this data in order to be generalized to other data later on. We have the test set (or subset) in order to test our model+ prediction on this subset. We calculated accuracy scores on the training data, both on validation set (Dev set) of 0.3 and Test set of 0.2. The performances (both for Dev test and Test set) are shown in Table 5 in terms of Accuracy, Precision and F1 Score. We obtained roughly the same results for Accuracy in Dev set and the Test set, 0.794 and 0.775, respectively.
Finally, in order to compare our approach, we considered two other baseline models namely Multinomial Naive Bayes (MNB) and Linear Support Vector Machine (SVM) besides the feed-forward
Dev set 0.796 0.937 0.803 To assess the performance of the model, the Root Mean Square Error (RMSE) was computed. RSME measures the distance of the predicted value to the true value. It is a measure of error, so the lower is the score, the better is the performance. We show RMSE results in Table 7.
Dev set 0.233
0.224
Using classification accuracy alone when evaluating the performance of the classification algorithm could be misleading, especially if the dataset- as in our case - is limited in size or is unbalanced or contains more than two classes. Hence, a confusion matrix is used to evaluate the results of the experiments. The confusion matrix M is a N- dimensional matrix, where N is the number of classes, that summarizes the classification performance of a classifier with respect to Test set and Dev set, both as in our case. Each column of the matrix represents predicted classifications and each row represents actual defined classifications. As shown in Table 8, during the validation phase, the classifier made a total of two hundred-sixteen predictions, while during the test phase the classifier made a total of two hundred-fourteen predictions. Out of two hundred-sixteen cases in validation, the classifier predicted “Females” forty-four times and sixty-four “Males”. Actually, sixty people in the sample belong to “Females” class and forty-eight
45 19
After this intermediate phase and after having tuned the parameters in order to optimize the model on the previous results, the classifier made a total of two hundred-fourteen predictions during the test phase. Out of two hundred-fourteen predictions, the model predicted “Females” fortythree times and sixty-four “Males”. Indeed, fiftynine people belong to “Females” class and, as predicted during the validation phase, forty-eight to “Males” class. We report gender prediction results on test data in the confusion matrix in Table 9.
In this paper we have shown our results on gender detection in dream diaries and writing styles differences and similarities between males and females in dream tales. First we explored the vocabulary of dream descriptions for both the genreclass by listing some of the representative words for each genre. Then, we evaluated our gender detection model on the dream reports dataset. The model succeeded in obtaining good results managing to distinguish a good part of dreams made by men or women. This research represents our preliminary step in the field, toward subsequent studies, in which we are trying to detect stylistic differences between dream tales and personal descriptive narratives, such as travel tales and other forms of self-narration.
This project has been partially supported by the PON Ricerca e Innovazione 2014/20 and the POR Campania FSE 2014/2020 funds.