Our Dreams, Our Selves: Automatic Interpretation of Dream Reports

Sleep scientists have shown that dreaming helps people improve their waking lives, and they have done so by developing sophisticated content analysis scales. Over the years, sleep scientists have developed hundreds of increasingly sophisticated ways of coding dreams. One of the the best validated and most widely used scale is the Hall and Van de Castle’s. The Hall–Van de Castle dream coding system consists of ten categories (and their sub-categories) of elements appearing in dreams, together with detailed rules to recognize and measure those elements from written reports. The system became a standard reference for quantitative dream analysis, thanks to its objective approach that facilitates reproducibility and high inter-coder reliability. In practice, the ten categories are not all of equal importance in capturing the psycho-pathological aspects of a dream’s content. Dream scientists determined that the three categories of Characters, Social Interactions and Emotions are the most valuable ones and are usually more informative than all the remaining ones combined. These are described as follows:

Characters. People, animals and imaginary figures who appear in the dream; 
Interactions. Interactions among characters of three types: friendly, sexual and aggressive;
Emotions. Markers of positive or negative emotions in the dream.

Each of these three categories is quantified by a set of metrics. For example, the Characters category includes the percentages of male, animal, and imaginary characters that appears in the dream.

As most dream content analysis scales, the Hall–Van de Castle dream coding system  is complex and, as such, require human intervention. As a result, annotations have been mostly done manually, which is time-consuming, does not scale. To tackle this challenge, we designed a Natural Language Processing (NLP) tool that automatically scores dream reports by operationalizing the Hall–Van de Castle coding system. We validated the tool’s effectiveness on hand-annotated dream reports and applied it to a set of 20k+ dream reports from dreambank.net. This dataset contains those algorithmic annotations.

Methods
Original data was gathered from dreambank.net. We used NLP tools to annotate the dreams according to the Hall-Van de Castle dream coding system.

For details about the coding system, see: https://dreams.ucsc.edu/Coding/

For details on how the annotation algorithm works see research paper associated with this dataset: http://dx.doi.org/10.1098/rsos.192080