How can you translate a Sankey diagram into clear language without changing where a quantity came from or where it went? Start by identifying the nodes, following the ribbons in their labelled direction and reading the unit represented by ribbon width. A Sankey diagram makes flows visible by drawing connecting bands whose thickness represents an amount on the diagram’s stated scale. Translating it accurately means preserving the relationship between sources, destinations and quantities, not merely naming the coloured shapes.
Readers searching for how to read a Sankey diagram, explain Sankey flows, describe an energy-flow chart or translate a Sankey visual into plain language need to distinguish width from colour, direction from page layout, and a conserved quantity from a percentage calculated on a different base. A thick band may mean more energy, people, money, material or another measured quantity. The original caption, labels and units decide which. This guide shows how to turn that structure into accessible prose or another language without distorting the data.
The worked examples below are hypothetical learning cases. Sankey diagrams appear in energy analysis, material accounting, public data, customer journeys, spending reports and research, and the rules of interpretation can vary with the model. The practical objective is simple: make every translated claim traceable to a source node, a destination node and a stated value.
A Sankey diagram is a network of quantities
A Sankey consists of nodes that represent categories, stages, sources or destinations and links that represent quantities moving or being allocated between them. The bands may branch, merge or cross, and their drawn widths communicate relative size. Unlike an ordinary flowchart, a Sankey usually encodes quantitative magnitude in the links. The translator should say ‘the diagram shows how the measured total is divided between categories’ rather than simply ‘the arrows show what happens next’ if the numbers are the central message.
The precise interpretation depends on the data model. In an energy Sankey, links may represent energy entering processes, useful outputs and losses. In a customer-flow display, links may represent counts moving between categories at successive stages. In a financial display, they may represent expenditures. None of these should be assumed from a band alone. Read the title and legend before describing the story. A Sankey graphic is a structured account of quantities, not a general illustration of causal relations.
Identify nodes and links separately
A node is usually a labelled bar, block or category where flows enter or leave. A link is the connecting ribbon. Saying that the ‘factory’ is a flow may confuse a facility node with energy allocated to it. Likewise, an arrowhead or ribbon curve is not itself the unit of measurement. A precise translated account states that a particular quantity flows from Node A to Node B. This source–destination sentence provides a natural-language replacement for the graphical connection.
Labels can be repeated across stages. A node called ‘other’ at one step may not be the same set of records as another node called ‘other’ elsewhere. Translators should preserve stage context rather than merging categories with identical words. When names are shortened for display, consult the source data dictionary before expanding them. A single invented word in a category label can redirect a large reported flow to the wrong place in the reader’s mind.
Direction comes from connections and arrows, not intuition
Many Sankey diagrams read from left to right, but that is a layout preference rather than a universal law. Some run vertically, some loop, and some are interactive. The direction of flow must be established from the diagram’s arrows, source and target fields, or accompanying explanation. A translator who writes ‘A supplies B’ when the data actually represent B contributing to A has reversed the relation even if the same two names appear in the sentence.
Read each link as a directed proposition: from a source, through a labelled connection, to a target. If orientation is not shown, say ‘the diagram connects A and B with a ribbon labelled 40 units’ rather than choosing a direction. The decision about direction often matters more than whether the prose reproduces the chart’s aesthetic shape. An accurate translation can be visually plain as long as a reader can reconstruct which relationship the link encodes.
Width means a measured amount on a scale
The most distinctive feature of a Sankey is that link thickness is proportional to a quantity according to the chart’s construction. A ribbon twice as thick as another can represent twice the amount if the same width scale applies. But screenshots may distort dimensions, interactive diagrams may rescale sections, and visual comparisons across separate charts may not be valid. Translate reported numbers when available instead of estimating exact quantities by looking at pixels. ‘The widest visible flow’ is not the same as ‘exactly half the total’.
Ribbon width also does not automatically mean probability, rate, percentage or number of people. A plot of annual energy may use petajoules, while an audience transition plot may use thousands of visits. The title and unit determine the meaning. If the source gives absolute amounts and percentages together, preserve both and identify the percentage’s denominator. Replacing ’20 terajoules’ with ’20 percent’ because the band looks like one fifth of the page is a false translation.
Units should remain attached to the flow
Some Sankey figures annotate every link with its number and unit; others give a single unit in the subtitle. The translator must carry that unit into the relevant sentence, especially if the figure becomes detached from its caption. Energy might be recorded in joules or watt-hours; material flow in tonnes; money in a specified currency; population in persons or households. These quantities should not be compared as though they belonged to a common scale. A diagram can contain visually consistent ribbons and still require substantial explanation of measurement definitions.
Time basis matters as much as physical unit. ‘100 gigawatt-hours per year’ is not the same as ‘100 gigawatt-hours at one moment’, and a rate such as tonnes per hour is not the same as cumulative tonnes across a project. In translated prose, mention period, region and accounting boundary when present. A number without its time denominator can mislead readers about scale. Preserving the complete unit phrase may feel less elegant, but it protects the quantitative conclusion.
Totals and conservation: check what is being counted
Energy and material Sankeys often use balancing relationships in which input quantities are allocated among outputs, storage changes and losses. That accounting should be made explicit if the source provides it. However, not every Sankey is a strict physical conservation diagram. A customer-journey plot may omit dropouts, a survey may weight cases, and a network display may represent overlapping records. Assuming every node’s inflow must equal its outflow can be wrong when the model’s scope or data definitions differ.
To translate totals faithfully, first ask whether the system is intended as a closed accounting balance, a partial subset, a transition count or a descriptive network. Then compare the listed incoming and outgoing links under those rules. If 100 units enter a process and 60 emerge as useful output while 40 are explicitly shown as rejected heat, you can describe the allocation. If the figure only depicts 60 moving onward, you cannot invent the missing 40 as waste without additional source context.
A loss branch is still a measured destination
In energy diagrams, ribbons representing heat loss, transmission loss or unused energy can look like a flow that disappears from the main route. The quantity has not been abolished as a bookkeeping object; it has been assigned to a different category. A translator should identify the original label, such as ‘rejected energy’, rather than casually calling it ‘destroyed energy’. In physically consistent energy accounting, energy changes form or is dispersed; careless wording can misteach the underlying principle.
Other Sankeys use labels like ‘abandoned’, ‘not retained’, ‘unclassified’ or ‘unmatched’. These are category outcomes, not automatically material or energy losses. A student reading a learning-path Sankey might interpret an exit branch as students who failed, though the source may simply mean records did not continue in the observed data window. Translate the category conservatively and retain its definition. The branch’s emotional connotation should not overpower its data meaning.
Colours are labels only when the legend says so
Colour may indicate source group, destination group, process type, energy carrier, time cohort or merely visual distinction. Two charts can use identical colours for different categories. A translation that says ‘red means danger’ or ‘blue means electricity’ without source evidence imposes a meaning the legend may not contain. If the legend defines colour classes, state them. Otherwise describe colour as a visual identifier and rely on named nodes to communicate the relationship.
Accessible text should not require readers to distinguish colours to know which band goes where. Write source and destination names, values and units. For example: ‘The ribbon from renewable supply to generation is 20 units’ provides more durable information than ‘follow the green ribbon’. A translator can retain the colours in a fuller visual description while ensuring the factual relationship survives in monochrome print, audio narration and target-language text.
One link can split into several
A source node may send its quantity to multiple destinations. A faithful translation should preserve each destination and its labelled amount. If 90 hypothetical units split as 50 to heating, 30 to transport and 10 to other uses, the prose should identify all three branches and ensure their sum is consistent when the diagram expects an exhaustive distribution. Do not combine small branches into one catchall category merely because the target sentence would be shorter.
The diagram might group tiny flows under ‘other’. That group name should be retained, along with any source definition; it should not be replaced by a guess about what the small components are. In a multilingual summary, the first sentence can report the largest branch and a later sentence the remaining branches, but the hierarchy should not become a statement that small categories are unimportant. The reader’s decision may depend precisely on an apparently thin branch.
Several sources may merge into one destination
Merging bands show multiple incoming contributions to a single target category. The target node may total them, but only according to the displayed accounting scheme. A translator must not imply that one source caused or controlled another just because their ribbons join. For example, wind and solar might both contribute to an electricity-supply node. That does not show a causal relationship between wind and solar. It shows contributions to a common destination under the model.
If incoming values are labelled, report each before quoting the combined total. Verify whether the diagram’s aggregate is calculated from displayed rounded values or from more precise underlying data. A difference of one unit can be a rounding effect rather than a data error, but it should not be silently explained away unless the source describes its conventions. Translators who write out the arithmetic help readers see both the quantity and the limitations of visual precision.
Repeated categories across stages are not always the same population
A Sankey may show categories at several time points: beginner, intermediate and advanced at the start, middle and end of a course. A learner who is ‘intermediate’ in the first period and ‘intermediate’ in the second appears in two different stage-specific nodes. Translating both as one timeless group conceals the transition being measured. The correct prose makes stage and time explicit: ‘At the first assessment, 40 learners were intermediate; by the second, 25 of that starting group remained in the intermediate category,’ if those are the actual data.
Similarly, a business funnel diagram may mix prospective customers, active customers and returning users. Read cohort definitions before saying how many people ‘moved’. Some displays count sessions or transactions rather than unique individuals. If the unit is ‘visits’, translating it as ‘people’ can overstate the number of individuals involved. A link’s path is meaningful only in combination with what the underlying records actually count.
Normalisation and percentage denominators
A Sankey may display absolute values, node-level percentages, shares of an overall total or conditional proportions. These are not equivalent. A ribbon labelled 50% could mean half of its source node, half of all observed flows, or half of a selected comparison group. Identify the denominator from the legend, tooltip or method note before converting the link into words. If the source is ambiguous, preserve the ambiguity instead of inventing the base.
Consider a node with 20 people entering from a group of 100 and an outgoing ribbon labelled 50%. It might represent ten people, but only if the percentage is relative to those twenty and the model represents one event per person. Some systems draw width according to counts while annotating a conditional probability; others encode both with tooltips. Translate the exact label and explain the difference between a visual width and a percentage statement. Otherwise readers can conclude that half the entire population followed a path that involved only a small subgroup.
Stocks, flows and changes over time
Stock refers to an amount accumulated at a point or over a defined state, while flow often describes an amount transferred during an interval or at a rate. Sankey diagrams can show transfers between stocks, but the two concepts should not be merged. A chart labelled ‘annual material throughput’ differs from an inventory statement labelled ‘materials held at year-end’. Translating either as a generic ‘amount of material’ removes an important temporal distinction.
Words such as input, output, accumulation, storage and change require the model’s accounting boundary. Energy sent to storage might be an output from one process but an input to another in the diagram. Avoid moralising a branch as ‘saved’ or ‘wasted’ based only on its route. If the figure gives a balance equation, reproduce the terms with their signs and units, and say which side of the boundary each quantity occupies. A clear reader can then follow the allocation without needing the original coloured bands.
Different stages and repeated values can create double counting
An interactive Sankey may display the same underlying unit as it moves through several stages. Summing every ribbon on the screen can count that unit multiple times. For example, 100 items may move from source to processing and then 100 items from processing to distribution. Adding both link labels to report ‘200 unique items’ would be wrong. A translation should distinguish throughput along a path from unique items entering the system, and should only produce a combined total when the data model supports it.
This matters in public budgets and educational progression diagrams. A monetary amount may be allocated to a department and then distributed from that department to projects. It is not two separate sums of new money. Likewise, one student may be represented in multiple transitions without being a duplicate person in the underlying cohort. Good narration follows the path rather than counting every appearance as an independent participant.
Source definitions can matter more than graphic design
An energy Sankey might include ‘primary energy’, ‘final energy’ and ‘useful energy’. These terms belong to an accounting framework and should not be flattened into three synonyms for energy. A marketing Sankey might define ‘converted’ differently from ‘purchased’. A translation that changes category definitions will produce accurate-looking sentences with incorrect totals. Read footnotes and methodological notes before choosing target-language labels.
Some visual tools abbreviate labels to fit narrow blocks. If a source node says ‘Elec. gen.’, verify whether it means electricity generation or another term within that data set before expanding it. Terminology management is particularly useful for repeated Sankey reports: make a glossary of each node type, unit and calculation rule, then use the same choices across translated figures. Consistency allows readers to compare periods without wondering whether a changed word represents a changed category.
Interactivity changes what is available to the reader
Online Sankey diagrams often reveal exact values only in hover tooltips, show a small subset until a node is expanded or allow readers to filter cohorts. A static screenshot may capture one filtered state rather than the whole data set. Translators should preserve the filter settings, selection criteria and source date when describing a captured view. Otherwise the resulting prose may claim to represent the entire system when it describes only a narrow slice.
If the source is interactive and the target is plain text, the translation may need a compact accessible table for selected paths. Choose the table’s scope explicitly: major flows only, all flows above a stated threshold or the complete set of links. Do not imply completeness if details are omitted for brevity. Document what cannot be recovered from the screenshot, especially hidden tooltips and hover-only definitions. A good translation helps readers understand both the data and the boundaries of the supplied representation.
A text-first method for accessibility
Describe the figure’s scope and unit first. Then name the principal source nodes, destination nodes and largest labelled paths, using the source’s direction. After that, describe branches, merges and smaller outcomes that matter to the question. If the chart is complex, group paths by starting node and keep a consistent sentence template: source, amount, destination. This lets a screen-reader user build the graph step by step without listening to a chaotic recital of colours and curves.
An accessible version can include a separate table with Source, Destination, Value and Unit columns. Every row should correspond to one link and retain any filtering or time basis. If the data contain identical node labels at different stages, add stage identifiers. Avoid replacing exact values with adjectives like enormous or tiny when quantities are available. Natural language should add context, not take precision away from the reader.
How translation between languages changes the risk
Languages organise ‘from’, ‘to’, ‘through’, ‘allocated to’ and ‘remaining after’ differently. Some use case marking or particles rather than the English word order. A literal word-for-word translation may therefore reverse the direction of a link even when all category nouns are correct. The translator should explicitly identify the source and destination roles, then construct a natural target-language clause that preserves those roles.
Quantity words also differ. A term glossed as ‘share’ may indicate a portion of a total, a financial share or an act of distribution. ‘Flow’ may refer to water movement, cash movement or a statistical transition. A technical glossary should capture what the quantity represents in this specific diagram. Cross-linguistic learning becomes more robust when students explain a link in their own words before selecting a target-language phrase. That short reasoning step exposes directional misunderstandings early.
What automatic summarisation tends to get wrong
Image-to-text systems can recognise prominent labels and estimate which bands look wide, but they may guess the direction of crossing links, conflate repeated nodes or turn visual proximity into causality. An AI summary might report an unlabelled ‘loss’ branch because many energy Sankeys contain such branches. That is not permitted if the source says something different. Automated extraction should first recover a structured list of actual links, values and node identifiers; summary writing comes after verification.
Ask a model to state which claims are taken directly from printed values and which are interpretations based on the visible layout. Check each link against the source. Do not let the model supply missing totals, currencies or denominators. For high-stakes financial, engineering or public-policy use, a reviewer familiar with the underlying accounting model must confirm that the text preserves the data. Translation fluency is not a substitute for quantitative audit.
A reproducible route from diagram to paragraph
First transcribe the diagram title, data period, unit and filter state. Second identify node names and any stage distinctions. Third make a source–target list from each visible link with its number. Fourth note how width and colour are defined. Fifth check the totals in the context of the model; do not impose conservation where the source does not. Sixth draft prose from the link list, prioritising the reader’s question without suppressing relevant counterexamples.
Finally verify every amount, percentage base, direction, label and time unit. Compare your text with the original network and, where possible, the underlying structured data rather than only the screenshot. Read the target alone: would a new reader be able to follow where the quantity goes? Does any phrase make a connection sound causal when it is only an allocation? Does the paragraph compare links that use different denominators? Only after those checks is the translated explanation ready to publish.
Twelve worked examples: translating the quantities and the paths
Worked case: one input and three destinations
An imagined energy Sankey shows 100 units entering a facility. The outgoing labels show 55 units of useful output, 30 units of rejected heat and 15 units attributed to other losses under the model. A faithful description lists those categories and amounts, identifies the accounting period and says the chart allocates the full input among them. It does not say that 45 units vanished from physics; it says the figure assigns those units to the named non-useful branches. If a source diagram omitted one branch, the translator could not invent it merely to force a neat balance. The exercise teaches the difference between reported accounting and inferred completeness.
Worked case: direction runs upward
A Sankey is rotated so its principal quantities flow from bottom to top. The target writer assumes all Sankeys read left to right and describes the upper node as the starting source. This reverses the story. The correct translation uses arrowheads or data fields to determine where each link begins and ends, then describes the unusual orientation explicitly. ‘Flows proceed upward from the listed inputs to the final categories’ is clearer than silently imposing a familiar reading order. The lesson applies to vertical diagrams in reports, posters and mobile interfaces.
Worked case: two widths but different units
One page shows an annual energy Sankey in terajoules and a separate funding Sankey in millions of currency units. Their thickest bands look similar. An inaccurate caption says the two charts show equivalent amounts. Width has meaning inside each diagram’s scale, not across unrelated units. A correct translation identifies what each quantity measures and avoids cross-chart numerical comparisons unsupported by a common definition. The exercise can be extended by asking readers to state the period and system boundary for both figures. Comparison needs a valid basis, not matching shapes.
Worked case: the meaning of ‘Other’
A category labelled ‘Other’ receives a thin ribbon. A translator expands that to ‘miscellaneous waste’. But the source data dictionary defines Other as several legitimate uses not individually shown. The invented word waste changes the meaning and can shape a reader’s judgement. The right translation retains the authorised category label, describes its value and points to the data definition if available. When a label is vague, clarification comes from evidence rather than storytelling. This is particularly important when public resources or people’s choices are being categorised.
Worked case: a percentage with the wrong denominator
A group of 40 people enters a second-stage node from an initial cohort of 200. Of those 40, 50% move to a next-stage category. The outgoing quantity is 20 people, which is 10% of the original cohort. If the translation says ‘half of all 200 advanced’, it confuses a conditional share with an overall share. The corrected paragraph identifies the denominator: half of the 40 people in that second-stage node. The visual band’s width and its tooltip must be interpreted together. Numbers can remain unchanged while the sentence changes their statistical meaning.
Worked case: the same resource appears twice
A resource Sankey tracks 70 units from production to distribution and the same 70 units from distribution to consumption. A reader sums both links and reports 140 units produced. That double counts two stages of the same flow. A precise translation says 70 units passed from production to distribution and onward to consumption. It does not treat each link as new input. This example demonstrates how narrative verbs such as proceeds, continues and is then allocated help preserve the continuity of a quantity across multiple stages.
Worked case: a colour-coded path without labels
An accessible summary says ‘the orange ribbon becomes green’, but the legend is absent from the cropped screenshot. Colour alone may not establish a category change; visual styles can be chosen for readability. A responsible translation describes the labelled nodes and any visible values while noting that the colour scheme’s meaning is not supplied. If a source legend is later recovered, that can support a fuller account. Until then, no colour-based claim is warranted. The exercise shows why accessibility descriptions often improve accuracy for all readers.
Worked case: the apparent loss is an unobserved group
A learner-journey Sankey shows fewer records reaching the final stage than appearing at the start. A writer calls the difference ‘dropouts’. Yet the dataset only includes records with follow-up responses, so missing respondents may explain part of the gap. The target version should use the source’s category, such as ‘not observed at follow-up’, instead of imposing a behavioural interpretation. A careful report can discuss possible explanations separately if evidence is available. The data’s observation window, not the band shape, determines what the missing flow means.
Worked case: rounded labels do not quite add up
Three outgoing ribbons display 33%, 33% and 33%, while the caption says the full input is allocated among the three outcomes. The apparent 99% total may arise from rounding. Rather than silently changing one label to 34%, the translator should reproduce the displayed percentages and, if the source confirms it, note that rounding can explain the difference. If the source has exact underlying data, those values can be checked. The core principle is that a translator should not repair arithmetic by inventing a changed datum without authority.
Worked case: a node called electricity
An energy-flow figure has nodes for primary fuels, generation, electricity and end uses. A hurried sentence says ‘electricity caused the fuel consumption’. The plotted direction may show fuels supplied to generation and electricity delivered onward. The correct explanation follows those directed allocations. It also preserves losses and accounting boundaries if shown. A Sankey link does not by itself prove a causal mechanism, but it does encode a source-to-target allocation. The target language must be careful not to trade that precise relation for a dramatic causal claim.
Worked case: currency conversion in a budget flow
A public spending Sankey labels flows in thousands of a stated currency. A translator converts an amount to another currency using an assumed exchange rate but does not state the rate or date, creating a misleading appearance of precision. The simpler accurate approach is to retain the original currency and numeric unit. If a conversion is genuinely requested, specify the rate, date and approximation separately and preserve the source amount. Translating a currency symbol into a more familiar symbol without conversion is an especially serious error because the visible numeral remains plausible.
Worked case: a static export of an interactive chart
A web interface lets users filter a Sankey by year and demographic group. A screenshot captures one year and subgroup, but a detached caption says ‘all users follow these paths’. The proper translation records the active filters and treats the screenshot as a conditional view of the data. Links hidden by the interface should not be imagined. If the target audience needs the wider picture, direct them to the full source or request additional data. The ability to describe only what was actually displayed is an important part of trustworthy quantitative communication.
A practical reader self-test
Can you name the diagram’s quantity and unit, reconstruct one source–target link without looking at its colour, and explain why a width comparison is valid only on the appropriate scale? Can you identify a stage repeated along a path without counting the same underlying unit twice? Can you explain where the denominator of a displayed percentage comes from? If so, you are reading the graphic’s information architecture rather than just its appearance.
Now take one diagram and write a short explanation for a reader who cannot see it. Have another person use your explanation to draw a simple table of sources, destinations and values. If the reconstruction differs, inspect the lost relationship. This is an effective practice method for translation, education and accessibility because the result is tested against the underlying structure rather than the smoothness of the prose.
Further routes through eduKateSG
The distinct learning job of this guide is the language transfer of quantitative directed flows. It complements general chart translation, flowchart logic and technical language learning, but it does not replace their specialist explanations. For broader translation decisions, visit eduKateSG’s Master Art of Translation. For contextual word choice and cross-linguistic transfer, use the vocabulary learning resources and compare terms such as flow, rate, stock, source and destination in meaningful sentences.
A good final sentence about a Sankey should always answer three questions: what moves or is allocated, how much, and between which named points? When a reader can answer those questions from the translated text without introducing any additional assumption, the diagram has successfully moved into language.