We use the dream reports from collections as gathered in the DreamBank, a project to combine the results of several scientific studies and resources over the years in one online search interface [Domhoff and Schneider 2008]. These collections of dream series vary greatly in type, size, and intended purpose. Some series consist of a longitudinal collection of dream descriptions of a single person, such as the collection “Dorothea: 53 years of dreams” consisting of around 900 dreams. Other series represent a specific group of dreamers such as adult male and female blind dreamers [Hurovitz et al. 1999]. Some collections are in English, gathered in Australia, Canada or the US. One of the collections, collected in Switzerland, contains dream reports in German. The DreamBank is an ongoing project and collections are added regularly. We use a snapshot from the DreamBank retrieved in April 2015 containing 22 thousand dreams divided in 67 different collections.

For our experiments we performed the following selection steps on the DreamBank data, where we limited ourselves to English written collections. Since some of the DreamBank collections overlap, we removed the duplicates from our sample. We also removed a part of the description in the collection “College women from the late 1940” that contained answers to specific questions, and we only kept the dream description itself. We applied an automatic language identification step [Lui and Baldwin 2012][1] that filtered out a handful of other dreams (for example dream #0694 of the Barb Sanders collection [Domhoff 2006] where she dreams about a Spanish conversation). Next, the data was tokenized automatically,[2] leading to a sample of 21,598 dream descriptions containing a total of 4.3 million word tokens. Dream descriptions contain an average of 56 words, with a population standard deviation of 38.5.

We noted that some collections in the DreamBank are much larger than others, and that dream descriptions of certain persons (e.g. Barb Sanders) are relatively prominent in the DreamBank content. We decided to create a sample of the DreamBank where we limit the amount of dream reports per individual dreamer to a random selection of at most one hundred dreams. This produced a sample of 6,998 dream descriptions, comprising 1.3 million tokens in total with an average of 65 words and a standard deviation of 43.7 per dream description, similar to the larger sample. We used both the large and the small sample in our experiments.

We show an example of a tokenized dream description with 97 tokens:

To discover what the typical linguistic attributes of dream reports are, we need a comparable set of contrasting reports that is as similar to dream reports as possible, both in structure and in content. Comparing dream reports to a collection of news paper articles or personal letters will lead to obvious findings such as: dreams do not report on political debates and the weather forecast, and will not end in ‘yours sincerely’. This is not the type of differences that we are interested in. We therefore aimed to find a collection of personally written recollections of true daily life events. Recall that dreams are known to reflect daily life events and activities for at least 75–80% of the cases.

Comparable collections of personal stories recollecting true events, not just fantasies or fiction, are difficult to find when looking for existing curated corpus collections. For this reason we resorted to collections of web texts to build our own corpus.

The first part of the contrasting data consists of personal stories. The stories are crawled from Prosebox,[3] an online community to share journals and personal stories. Just prior to this research, OpenDiary, a community where users could post diary entries, was taken down. Many of these users moved to Prosebox, and especially older posts are mostly diary entries or journals.

We collected all public posts that were available at the end of March 2015. As a result, we crawled 130 thousand posts with over 67 million tokens. We applied the same filtering pipeline to the Prosebox posts as was applied to the dreams; that is, we applied a language filter where we only kept the posts which were identified as English; second, we tokenized the posts. Since the number of tokens is much larger, we downsampled the corpus to a similar number of tokens as the DreamBank samples, i.e. the large sample and the smaller limited sample, containing 4.3 million words and 1.3 million words respectively, with an average of 64 and 63 words and standard deviations 78.8 and 94.3, respectively. In other words, we kept the average document size virtually equal to that of the DreamBank samples; the Prosebox data does exhibit a larger variance in size. We show an excerpt of a Prosebox text here:

The second part of the contrast data consists of Reddit posts. Reddit is a website where users can submit content of almost every kind. The site uses a community system, where each community is called a subreddit. We collected posts from a number of subreddits where the posts are texts about daily and personal experiences such as communities named offmychest, diaries, relationships, shortscarystories, lifeinapost, anxiety, and self. Prior to this research, the complete Reddit corpus was not available.[4] In total we crawled 122 thousand posts with 54 million words with an average post length of 71 words for the 1.3 million words sample, and 72 words for the 4.3 million words sample, respectively with standard deviations of 61.7 and 67.8. We show an example Reddit story here:

As a first analysis of the text collection, we set out to train machine-learning classifiers to distinguish dream reports from personal stories automatically. Both the extent to which the classifiers succeed, and the features they use to make their decision, can lead to insights in the differences between dream reports and other narratives. Terms or groups of terms that are identified by classifiers as strong indications that a text is a dream report or not are apparently typical for their respective class.

In text classification, a machine learning classifier is fed with labeled documents from which it learns to model the characteristics of the given labels. Its labeling performance is tested by applying the classifier to a held-out set of documents. For this experiment, we used the sets of 4.3 million words for both the dream data and contrasting data.[5]

We tokenized all documents with the Stanford Tokenizer.[6] The word tokens were standardized to lowercase. We extracted word unigrams, bigrams, and trigrams as features. To avoid bias from explicit markers of dream reports, we removed any features that contained one of the following words: dream, dreamer, dreamt, dreamed, dreams, awake, awaken, woke.

We compared the performance of three different classification algorithms: Support Vector Machines (SVM), Naive Bayes, and Balanced Winnow. We used the libsvm [Hsu et al. 2003] implementation of SVM, with linear kernel and setting the C parameter to 1.0. We applied Naive Bayes by using the Multinomial Naive Bayes implementation in Scikit-learn [Pedregosa et al. 2011].[7] For Balanced Winnow, we made use of the Linguistic Classification System [Koster et al. 2003]. The α and β parameters were set to 1.05 and 0.95 respectively. The major threshold (Θ+) and the minor threshold (Θ−) were set to 2.5 and 0.5. The number of iterations was bound to one.

We evaluated the performance of the three approaches by means of ten-fold cross-validation. To avoid author bias, the reports by the same author were kept together in either the test set or the train set during each fold. During each training phase, the 7,500 most frequent features were selected and presented as binary values.

The classification results, micro-averaged over examples, are given in Table 1. All three approaches yield a precision and recall of 0.97, which indicates that the dream and non-dream reports can easily be distinguished with a small remaining margin of error. Table 1 also displays the exact number of documents that were correctly classified. The Balanced Winnow classifier has a slightly higher number of correct classifications than Naive Bayes and SVM.

The Balanced Winnow classifier returns an interpretable model of the features that the classifier used internally to make its predictions. Upon analysis of the 30 most indicative features of the dream and non-dream classes, we obtained the following insights about the two types of texts:

To discover what type of topics are typical for dream reports, we employed an unsupervised method that is currently popular for discovering latent themes or topics in large document collections. Latent Dirichlet Allocation (LDA) [Blei 2012] is a probabilistic generative algorithm that aims to give a broad overview of the topics that occur in a collection of documents. Topics are defined as a distribution over a fixed vocabulary (in practice, a topic consists of a set of semantically related words). Topic modeling is an unsupervised process solely based on word occurrences in documents. LDA assumes that documents are created based on an underlying topic distribution and each document is generated from a mixture of these topics that each have a different proportion in the document. LDA uses an iterative process to estimate this underlying distribution based on the observed words in the text.

We ran experiments with LDA on the full DreamBank sample of 22,046 dreams. We filtered the dream texts to exclude all function words and punctuation marks and only kept those content words that were automatically part-of-speech tagged by the Stanford parser as nouns, verbs and adjectives. All words have been converted to lower case. Such explicit filtering step ensures that the generated LDA topics contain only content words.

For these experiments we use the LDA implementation provided in the Mallet toolkit [McCallum 2002]. We ran LDA with 2,000 iterations with Gibbs sampling and 50 topics. The produced LDA model was used to annotate each document with its most relevant topics; namely those topics that cover at least 10 percent of the document. Documents have three such topics on average.

Setting the number of topics parameter is a rather arbitrary choice. We ran experiments with 100, 200 and 400 topics as well and studied the output. When raising the value of this parameter, more fine-grained topic descriptions are produced. These detailed topics are still understandable and coherent topics, but, as can be expected, they tend to have a lower coverage in the document. As we aim to look at significant differences between topic distributions in different sample sets and to compute g-tests (log-likelihood tests) [Rayson and Garside 2000], we choose to keep the number of topics fixed at 50.

When people are asked what is typical about dreams, they will often mention weirdness as a typical property of dreams. This might be due to the fact that many dreams are forgotten the next morning while weird or impressive dreams tend to stick in people’s memory [Bulkeley and Hartmann 2011]. This recollection could be attributed to the bizarreness effect, the inclination to remember bizarre things better than ordinary things [McDaniel et al. 1995]. Domhoff shows in [Domhoff 2007] that bizarreness does occur in dreams, but it is not as manifest as people tend to believe.

Bizarreness can emerge in different forms in dreams. The most prominent type of oddity seems to be caused by the discontinuity of events and sudden switches of scenes. In the study of Reinsel et al. [Reinsel et al. 1992] bizarreness was measured by counting three different type of occurrences: discontinuous events, improbable combinations, and improbable entities. Discontinuous events were found to be the most contributing factor in bizarreness (around 60% of the counts) while improbable entities were much less present (only around 8%). This is in line with previous studies on large volumes of dream reports that show that the amount of bizarreness attributed to strange entities is relatively low; most characters in dreams are known persons [Domhoff and Schneider 2008]. Domhoff [Domhoff 2003, 132] investigated metamorphosis, a specific type of discontinuous events, via a search in DreamBank.net and found only 50 mentions of metamorphosis in the whole DreamBank. Transformations turn out to be highly infrequent in dreams. [Schweickert and Xi 2010] study metamorphosis as a typical dream phenomenon and focus on the relation between change in form to change in inner states. No evidence was found that form change is connected to a change in mental state. This was a small-scale study on a set of 65 dreams from 21 persons.

Interestingly, bizarre thoughts do not only occur in our sleep but can also occur when awake. It has been shown that people in a relaxed undisturbed awake condition produce dream-like reports when asked to recall that most recent thoughts in the same way subjects are asked after being awakened [Reinsel et al. 1992]. These awake fantasies are very similar to dream reports, including bizarreness.

In our study we aimed at using a quantitative approximation of bizarreness and applying this metric to the DreamBank texts. We focus on the discontinuity of events in dreams and try to quantify this by looking at textual coherence in the dream reports. We hypothesize that dreams are less coherent in their discourse structure than personal stories. We measure two different aspects of discourse structure, namely discourse marker frequency and entity-based text coherence, using the smaller balanced sample sets of 1.3 million words.

Discourse analysis is a broad and multi-disciplinary field that studies language in use beyond the sentence level [Trappes-Lomax 2004]. Automatic discourse parsing is still in its early development phase, as was illustrated by 2015’s CoNLL shared task [Xue et al. 2015] on shallow discourse parsing where the best system achieved a overall F-score of 24%.

In this initial experiment we only focus on discourse marker occurrences and measure whether there is a difference between discourse marker frequency in the dream data and the personal stories. Discourse markers are words or phrases that explicitly signal discourse structure and describe how two sentences or phrases are related to each other. For example, for example indicates that the current sentence exemplifies something that was mentioned in the previous sentence. Typical discourse markers are but, since, while, even though, and because.

We used a list of 60 markers[9] based on annotations from the Penn Discourse Treebank [Prasad et al. 2008] that was used in the CoNLL shared task. In total, about 40 thousand occurrences of these discourse connectives were counted in the DreamBank data, and about 50 thousand in both Reddit and Prosebox, which means that there are about 20% fewer discourse markers counted in the DreamBank data than in the contrasting data. In Figure 1 we show the distribution of 28 of the 60 markers in the balanced DreamBank, Reddit and Prosebox data sets, having a frequency of occurrence over 250. For most markers we can observe similar distributions, or a slightly lower count for the dreams. One marker however occurs substantially more often with the dreams than with Prosebox or Reddit: “then”. This is a typical discourse connective that is used in sequential narration. Conditional discourse markers like “if” and “if then” and causal markers like “because”, “so” and “since” occur less frequent in the dream sample.

In a second experiment we study entity-based coherence. Mentioned entities and chains of referring expressions in a text are core indicators of text coherence. We assume that discontinuity in dreams is expressed in sudden shifts in scenes and events, and we expect that these are linked to shifts in mentioned entities. On the basis of this assumption, we tried to measure discourse coherence by applying an existing automatic model to detect entity-based coherence.

We used the Brown Coherence Toolkit v1.0 [Elsner and Charniak 2008]. The authors of this toolkit present an extension of the entity-grid coherence model proposed by Barzilay and Lapata [Barzilay and Lapata 2008]. An entity grid represents the entity mentions in a document in a textual matrix where each row represents an entity and the column represent the syntactic roles of the entities (subject, object, other). This matrix is used to predict which role each entity will have in the next sentence.

To detect the entities in the text we used the extended entity grid based on the Wall Street Journal corpus that was automatically pre-processed with OpenNLP software, available in the Brown Coherence Toolkit.

We applied the model to each of the balanced dream and personal stories data sets and measure its performance with a binary discrimination test as was previously done in the work of Elsner and Charniak [Elsner and Charniak 2008]. The binary discrimination test tests the model’s ability to distinguish between a human-authored document in its original order, and a random permutation of that document. The test reads any number of documents and performs the test on each one, using 20 random permutations.

The results of this test are shown in Table 5. All achieved results are substantially lower than the scores reported by Elsner and Charniak, who report an F-score of 86%, when training on Wall Street Journal (WSJ) newspaper text and testing on a held-out set from the same corpus. As WSJ consist of financial news paper articles, we can expect a drop in performance when switching to a completely different textual genre of dreams and personal stories. Nevertheless, the scores in Table 5 indeed suggest that the dream text is less coherent than more formal edited text in terms of coherence relations, but also as compared to Prosebox and Reddit, which are remarkably similar.