Automated dream analysis is made of this …

Sebastian’s note #1: large-scale study of dream reports

The Knight’s Dream, 1655, by Antonio de Pereda

Study: Fogli, A., Maria Aiello, L. and Quercia, D., 2020. Our dreams, our selves: automatic analysis of dream reports. Royal Society open science, 7(8), p.192080.

General hypotheses in dream research literature

• Continuity hypothesis: Most dreams are a continuation of what is happening in everyday life. 80% of dreams are experienced from a first-person perspective and usually involve ordinary scenarios.
• “Nocturnal therapist” theory: 80% of dreams involve negative emotions, regardless of everyday experiences. Dreams help a dreamer identify worries and concerns.
• Simulation hypothesis: Threads perceived in everyday life lead to threat simulation in dreams.
• Social simulation hypothesis: Dreaming helps simulate social skills and strengthens relationships.
• Dreams as problem solving: Mind attempts to solve a problem by looking at it from unusual perspectives.

Data

Most data in dream research has been manually coded and comes from small clinical samples.

Data sets used by Fogli et al.

• DreamBank.net: 38,000 dream descriptions from 1960 to 2015, gathered from various sources, including research studies. Free-text descriptions of the dreamers.
• Full set: 24,000 reports from DreamBank.net are written in English and composed of at least 50 words.
• No conditions set: 23,000 reports from Full set, only reports individuals for which no medical condition was reported in the free-text description
• Blidness set: 500 dreams reported by people affected by complete vision less
• War veteran set: 600 reports recorded from 1971 to 2017 by a Vietnam war veteran
• Izzy set: 4,300 reports recorded by a woman (from age 12 to 25) passionate about collecting her dreams
• Normative set: 1,000 dream reports from 100 male and 100 female US students, hand-coded by Hall and Van de Castle
• Annotated set: 1,700 hand-coded dream reports from 50 people

Coding

Hall & Van de Castle Scale

• Winget & Kramer: reviewed 150 dream rating/content analysis scales.
• Calvin Hall & Robert Van de Castle’s (H&VdC) scale: best validated and most widely used
• Gradually developed by Hall in the 1940s, expanded with the help of Van de Castle.
• Book: “The content analyis of dreams”
• William Domhoff: used H&VdC scale to analyze thousands of dream reports, provided strong support for continuity hypothesis
• 10 categories of elements appearing dreams: (1) characters, (2) interactions among characters, (3) emotions experiences by characters, (4) actions performed by characters and their sensoryexperience, (5) striving: success/failure in carrying out activities, (6) (mis)fortunes: happening to characters, (7) physical surroundings or objects, (8) descriptions of objects, people and actions, (9) food and eating, (10) characters or elements belonging to the dreamer’s past
• Three categories are the most valuable ones: characters, social interactions and emotions.
• William Domhoff combined counts based on the H&VdC scale into simple measures that correlate with dreamers’ everyday experiences.
• Cohen’s h is used in literature to compare a given dream report’s values against typical values.

Coding in Fogli et al.

• Characters: people, animals and imaginary figures; their gender and whether they were dead.
• Interactions: friendly, sexual and aggressive
• Emotions: positive or negative

Automated dream analysis

Natural language processing

• 2 knowledge bases: Wikidata (55 million data items) and WordNet. In WordNet, English words are categorized grouped into sets of synonyms.
• Expanded common English contractions (“I’m” -> “I am”)
• Text is broked down into its phrasal categories (e.g., noun phrases, verb phrases, etc.) or lexical categories (e.g., nouns, vergs, etc.), resulting in a parse tree.
• Lemmatization: e.g., “dreaming” -> “dream”

Character names

• A list of people names was gathered from the WordNet words in noun.person category which are subclass of Person or instance of Person in Wikidata.
• A list of animal names was gathered analogously using noun.animal and Animal.
• Fictional characters were identified using Wikidata items Fictional Human, Mythical Creature and Fictional Creature.
• Tool does not resolve pronouns to identify characters.

Character properties

• Tool classifies characters according to two dimensions: sex (of people characters) and “being dead”.
• Names are resolved using Wikidata.
• Identification of dead characters is perfomed based on a list of 20 death-related keywords. Keywords are matched with nodes in the parse tree. For each matching node, the distance between that node and each of the other nodes is computed. Character at nearest node is marked as dead.

Interaction between characters

• Verbs in WordNet categories of verb.contact and verb.communication were filtered down to three sets: aggression verbs (361), friendliness verbs (70), sexual interaction words (70).
• Nouns preceding and following interaction verbs are matched with sets of characters to identify interaction pairs.

Emotions

• The Emolex, an emotion dictionary containing 689 terms, is used to identify positive and negativ emotions.

Evaluation

• Data sets used to evaluate the tool: Annotated set and Normative set
• Precision, recall and F1 scores computed to measure extent to which automatically identified characters, interactions and emotions matched hand-coded ground truths by dream experts. The performance is best for the category of all characters (F1 score of 0.75, almost perfect precision) and worst for sexual interactions (F score of 0.17).
• The tool computes indicators (usually in the range between 0 and 1) such as the percentage of male/animal/imaginary, friendly/sexual/aggressive interactions per character and the percentage of negative emotions. The average deviation between automatically computed indices and ground truths is 0.24.

Hypotheses tested by Fogli et al.

Sex differences

• Aggressive behavior is more frequent and more intensive in men.
• Female friends talk about their emotions more frequently.
• Hypothesis 1 (H1): Female’s dreams are characterized by emotions (rather than interactions), and by limited levels of aggression.
• Manual inspection of small set of dream reports: men dream about physical aggression more often.

Adolescence

• Social anxiety increases (in part, due to exposure to new social experiences)
• Increased conflicts with parents
• Hormonal changes
• H2: Adolescents’ dreams are characterized by negative emotions, followed by sexual interactions in early adult life.

War

• 11–17% of American troops returning from Afghanisten or Iraq have suffered from PTSD.
• H3: A war verean’s dreams are characterized by negative emotions and aggression.
• H3 was found to be true for Vietnam veterans: frequent dream experiences associated with guilt and violence.

Blindness

• Blind people’s daily lives are similar to general people. However, they might require interaciton with carers who are often women.
• H4: Blind people’s dreams feature real-life carers (women).

Violence

• High levels of violent crime in the 1960s, followed by steady decline.
• H5: High levels of societal aggressions are associated with dreams characterized by aggression.
• H5 was confirmed in analysis of dreams following the September 11 terrorist attacks: dramatic increase in dreaming about explosions, death and fires.

Results

• H1: Male dream reports show higher markers of aggression, e.g. a higher nmber of aggressions per character.
• H1: Female reports contained more positive emotions and friendly interactions.
• H2: Izzy’s reports show an increase in negative emotions and aggressions during adolescence.
• H3: The war veteran’s reports are characterized by aggression, more male characters and less sexual interactions than reports in the no-conditions set.
• H4: Reports by blind people tended to contain more female and imaginary characters.
• H5: Indicators of aggression were highest for dream reports the 1960s and then steadily decreased, matching statistics of violent crimes in the US.

Limitations

• Demographic bias: mainly well-educates individuals from the US
• H2 and H3 each were tested using the reports of a single individual.
• Important individual characteristics of dreamers were missing.
• Inaccuracy of the NLP tools
• Data silent on causality
• Recollection bias: individuals reported what they remembered to have dreamed.

Please clap and follow if you would like to read more about scientific approaches to dream analysis!