Mytheme Computation
Despite working on the mytheme project for several years, it has proved persistently difficult to isolate and work with mythemes computationally. Part of the difficulty lies in the elusiveness of the mytheme itself. It is not simply a word, motif, sentence, or event, but a higher-order unit of meaning: a second-order signification that lifts off from sentence-level meaning while remaining materially carried by the sentence.
Nonetheless, the value in developing a contemporary computational approach is considerable. To work with mythemes is not merely to classify stories or detect recurring motifs. It is to approach myth as a system of deep-level thinking. Properly identified, the mytheme points to structures of relation, opposition, mediation, and transformation that organise thought beneath the surface of narrative (and everyday language more generally). More than this, because mythemes are transformational units, we can also understand them as generative. They allow one thought to be turned into another; they represent structures of thought (not just the recordings of thought). The computational mytheme project therefore proposes not only an analytical method, but a method for modelling the conditions of thought and creativity itself.
§
Having recently hit up against a brick wall, the following is an attempt to reboot. It’s fair to say, what we are trying to do is really difficult so it’s inevitable to have these moments. Sebastian Mallaby’s ‘The Infinity Machine’, which provides an in-depth portrait of the work of DeepMind, is a reminder of the necessary twists and turns in developing AI, especially when you are tackling things at a root level.
Working with vastly less resource and focusing on just the mytheme (or more broadly, a structuralist approach to language), the ambition is to contribute to more searching questions about thought and next-level architecture of AI language models.
Rebooting Mytheme Machine
Let’s start with a working definition:
A mytheme is the smallest narratively operative unit of symbolic meaning in a myth: not a word, motif, sentence, or event in itself, but the relational meaning carried by a passage when it contributes to a wider pattern of transformation, opposition, mediation, or inversion. In computational terms, this means a mytheme should not be treated as a literal textual fragment. A sentence may express a mytheme, but the mytheme is the abstract structure that the sentence carries.
For example, a sentence such as: ‘The child is abandoned in the forest,’ does not simply contain the words child, abandoned, and forest. It may carry a secondary meaning such as: vulnerable human life is expelled from the social world into the natural world. Or, more abstractly: culture expels its own fragile offspring into nature.
That secondary (or abstract) meaning may reappear elsewhere in very different wording:
‘The infant was placed in a basket and set upon the river.’
‘The youngest son was sent away from the village.’
‘The girl was left at the edge of the wilderness.’
These are not the same sentences. They do not necessarily share vocabulary. But they may carry a structurally comparable mytheme: the vulnerable figure is displaced from protection into danger, or the social order externalises what it cannot contain. So the computational problem is not simply one of keyword matching. It is a problem of semantic abstraction.
Working definition for coding
For practical coding purposes, let’s define a mytheme as follows:
A mytheme is a repeatable symbolic-relation unit, expressed through one or more narrative moments, in which particular actors, objects, places, or actions are organised into an abstract relation such as opposition, transformation, mediation, substitution, contradiction, or reversal.
This definition has several important consequences:
Mytheme is not identical to the sentence. The sentence is evidence. The mytheme is an interpretation of what the sentence does within a mythic structure.
Mytheme is not merely a motif. ‘Forest’, ‘king’, ‘animal’, ‘wound’, ‘journey’, and ‘marriage’ are motifs. A mytheme is more relational: human enters forest, king loses authority, animal reveals truth, wound marks hidden identity, journey converts ignorance into knowledge, marriage resolves a social contradiction.
Mytheme is not a topic. A passage may be about ‘death’, ‘kinship’, or ‘animals’, but the mytheme concerns the structural role that topic plays: death may be punishment, passage, sacrifice, renewal, concealment, or exchange.
Mytheme is repeatable at the level of structure rather than wording. Two passages can express the same mytheme while using different vocabulary, characters, settings, and surface actions.
Computational Formulation
We might formulate the task as:
Given a narrative passage, identify the abstract symbolic relation carried by that passage, and determine whether comparable relations recur elsewhere in the corpus
This suggests a pipeline that moves from surface text to progressively more abstract representations, follows:
1. Segment the narrative into candidate units
The first step is to break stories into units that are likely to carry meaningful narrative action. These may be sentences, clauses, short paragraphs, or event-like segments.
A useful unit is not necessarily a grammatical sentence, but a narrative proposition, such as:
Actor A does X to Actor B in setting C, producing consequence Y
For instance: ‘The fox tricks the crow into dropping the cheese’, could be represented as:
fox / deceives / crow / causing loss of valued object
Already, this is more useful than the sentence itself.
2. Extract narrative roles
Each unit can then be coded according to roles:
| Surface element | Abstract role |
| Fox | trickster / animal mediator |
| Crow | deceived possessor |
| Cheese | valued object |
| Flattery | deceptive speech |
| Dropping | loss / transfer |
| Eating | appropriation |
This helps shift from words to functions. The important point is that ‘fox’ is not always simply fox. In one story, the fox may function as trickster; in another, as helper; in another, as liminal mediator between wildness and intelligence.
3. Identify the relational structure
A mytheme emerges when the passage is recoded as a relation:
deceptive speech causes the transfer of value from the vain possessor to the cunning outsider
This is much closer to a mytheme than ‘fox tricks crow’, because it can be compared with other stories where there is no fox, no crow, and no cheese.
For example:
‘The stranger praises the king until he gives away his treasure.’
This may share the same structure:
flattery converts symbolic superiority into material loss
4. Encode secondary meaning
The computational challenge is to produce an intermediate layer between literal text and full interpretation. We might call this mytheme representation.
A useful representation might include:
| Surface event | The fox tricks the crow into dropping the cheese. |
| Narrative roles | trickster animal; vain possessor; valued object; deceptive speech; loss |
| Abstract relation | cunning speech defeats vain possession |
| Symbolic opposition | intelligence / vanity speech / possession wild cunning / social display |
| Transformation | possession becomes loss praise becomes deception voice becomes vulnerability |
| Mytheme representation (candiate) | Deceptive recognition causes the vain subject to lose what it possesses. |
This format makes the passage computationally usable without pretending that the machine has ‘understood’ the myth in a human sense.
Mytheme as a second-order signification
A mytheme might be described as a second-order signifying unit (this borrows from Roland Barthes’ well-known account of the ‘sign’ in Mythologies).
At the first level, the sentence says:
A fox tricks a crow.
At the second level, the sentence may signify:
cunning overcomes vanity
speech transforms possession into loss
animal intelligence exposes social foolishness
the desire for recognition makes the subject vulnerable
This means the same sentence can potentially carry more than one mytheme, depending on the wider system of relations in the story. So computationally, we can avoid assuming that each sentence has one fixed mytheme. Instead, each passage may produce a set of candidate mythemes, each with a confidence score and evidence.
Coding schema
A practical coding schema could include the following fields:
| Field | Question |
| Text segment | What is the actual passage? |
| Surface action | What happens literally? |
| Actors / objects | Who or what is involved? |
| Narrative roles | What functions do they play? |
| Symbolic domains | Which domains are activated? Nature, culture, kinship, death, food, law, animality, divinity, etc. |
| Oppositions | What contrasts are present? |
| Transformation | What changes from beginning to end? |
| Mediation | Is something crossing or resolving an opposition? |
| Inversion | Is an expected relation reversed? |
| Mytheme Representation | What abstract symbolic relation is carried? |
| Comparable segments | Where else does this relation recur? |
This schema is important because it prevents the computational process from collapsing into either keyword extraction or overly free interpretation.
Computational methods
Different methods can be used at different levels.
Embeddings can help identify passages that are semantically similar even when they do not share vocabulary. But embeddings alone may detect broad similarity rather than mythemic structure. They might cluster ‘abandoned child’, ‘exiled son’, and ‘girl left in forest’, but they may not explain the symbolic relation.
Semantic role labelling can help extract who does what to whom. This is useful for identifying event structure.
Dependency parsing can help identify relations between actors, verbs, and objects, although it will not itself identify symbolic meaning.
LLM-assisted coding can be used to propose candidate mytheme representations, especially if the prompt asks the model to distinguish literal action from symbolic relation.
Graph modelling can represent relations among actors, domains, objects, transformations, and oppositions.
Clustering can then be applied not to raw sentences alone, but to enriched representations of mythemes.
Important: Do not embed only the original sentence. Embed the interpretive coding layer as well.
For example, instead of embedding only, ‘The child is abandoned in the forest,’ we embed something like:
Literal action: child abandoned in forest.
Narrative role: vulnerable dependent expelled from social protection.
Symbolic domains: kinship, wilderness, danger, initiation
Opposition: culture/nature, protection/exposure, belonging/exile.
Transformation: child moves from social inclusion to liminal danger.
Mytheme: the vulnerable figure is expelled from culture into nature.
This gives computational methods a much better object to work with.
A more precise working definition
A stronger version of the definition might be:
A mytheme is an abstractable unit of narrative-symbolic relation, inferred from a textual segment, whose significance lies not in its surface wording but in its role within a system of oppositions, transformations, and mediations across one or more myths.
This definition gives us the computational target. We are not looking for repeated words. We are not even looking simply for repeated events. We are looking for repeated symbolic relations.
Practical rule of thumb
A candidate mytheme should be expressible in a sentence of the following kind:
X is transformed into Y through Z.
A mediates between B and C.
A contradiction between B and C is displaced onto D.
A figure associated with X crosses into the domain of Y.
An expected relation is inverted: X becomes not-X.
A social problem is expressed through a natural image.
A kinship contradiction is resolved through sacrifice, exile, marriage, birth, or metamorphosis.
So a useful mytheme is not:
‘The girl enters the forest’
But:
A socially vulnerable figure crosses from culture into nature, where danger becomes the condition of transformation.
Or:
Entry into wilderness converts social exclusion into initiatory knowledge.
Summary
For the purposes of coding, we might use the following definition:
A mytheme is a repeatable, abstract relation carried by a narrative unit, in which surface elements such as characters, actions, objects, and places encode a secondary symbolic meaning. It is identified by translating a passage from literal content into relational structure: actors become functions, events become transformations, and motifs become positions within symbolic oppositions.
The computational task is therefore to build a pipeline that moves:
text segment → event structure → narrative roles → symbolic domains → oppositions/transformations → candidate mytheme → cross-story comparison.