# Semantic Enhancement
# Book: Language
# Chapter URL: https://programmer.ie/books/language/06-chapter/
# Prose fingerprint (sha256, normalized): a60e19f3152b1b950faf88adfdecf2e6bd81af54efd6d0295012936fc2fab41d
# Snapshot generated at build time by Hugo. This is the complete rendered prose
# of the chapter; it is not a transcript of the source file.

Imagine a team inheriting a ten-year-old payments service and its forty-page architecture document. One passage describes the retry policy in dense prose: three paragraphs of conditions, backoff intervals, idempotency keys, and two exceptions buried in the middle. A new engineer reads it three times and still cannot say what happens when the third retry fails during a partition. A colleague draws seven boxes and nine arrows on a whiteboard: states, transitions and the two exception paths marked distinctly. The engineer can now trace the route to failure. The contrast is plausible, but it is an illustration, not an experiment. A second explanation and more time with the material could account for some of the improvement. To credit the diagram, we would have to compare it with a well-constructed prose explanation on the same task.
Two pages later, the same instinct could fail. Another passage compares four database options across eleven dimensions, with exact throughput figures, consistency guarantees and cost bands. An attractive isometric illustration puts database cylinders above clouds but omits the figures and exceptions. It might attract attention, yet it removes the information needed to make the choice. A comparison table is a better candidate because it aligns the values while retaining their units and qualifications. If the source already contains a clear table, leaving it alone may be better still. Different structure calls for a different representation, and sometimes for no transformation.
The last chapter asked how a smaller representation could help a reader decide whether to enter a source. This one asks about the reader who is already inside it. An enhancement appears beside or over a passage, but the canonical source does not change. The reader must be able to dismiss the enhancement and recover the original wording. When does a local change help reconstruction, and what evidence would tell us that it preserved the source rather than quietly rewriting it?
FIXED PAGE ↓ PREVIEW: compress the source before entry ↓ ENHANCEMENT: modify the representation during use Three things an enhancement can do The word “enhancement” covers three different acts. They differ not only in appearance but in the evidence required before the result can be trusted. Take the retry passage.
Reveal. Make structure already expressed in the source easier to locate. Highlight the two exceptions. Mark the condition that triggers each backoff. Number steps whose order the source specifies. The words may remain unchanged, but the emphasis is still an editorial choice: highlighting the wrong exception, or assigning an order the prose left open, can mislead. Revealing is usually easier to audit than redrawing, not automatically neutral.
Re-express. Render information the source already contains in another form: a state diagram of the retry policy, a table of backoff intervals, or a timeline of its retry windows. The form is new, but each asserted state, transition, quantity, ordering and exception should be supported by the source. When the prose leaves an outcome unspecified, the rendering must preserve that uncertainty rather than fill the blank with a plausible arrow.
Add. Bring in material not contained in the passage: a definition of idempotency, a link to the paper behind the backoff scheme, or a comparison with a later design document. These can help a reader, but they are not statements from the passage. The interface should label them as external context or analysis, identify their provenance, and let the reader distinguish an established contradiction from one the system merely suspects.
Conceptual comparison using the retry-policy example. A transition absent from the source would be an addition, even inside a diagram. Open full-size diagram. Revealing and re-expressing aim to preserve the passage’s claims while changing access to them. Adding is a different operation: it extends the available evidence or interpretation. Even a correct external fact is not thereby something the original author said, and a source citation does not by itself establish that the addition is true or in scope. The interface therefore needs a visible boundary between source-derived and externally supplied or inferred material. Otherwise a helpful definition, an unverified cross-document comparison and an invented claim can all arrive under the same reassuring label of “clarification”.
The line is thinner than it looks. A diagram forces choices the prose may not have settled: which transitions exist, whether a condition applies to one path or several, and whether an outcome is known. Not every layout decision asserts a relationship; box position is sometimes merely layout. But an arrow from the third retry to a dead-letter queue does assert a transition. If the source never specified that route, the arrow is an addition disguised as re-expression. An ambiguous passage should yield an explicitly uncertain edge, a question for the reader, or no diagram. Verification has to inspect the relations and their qualifiers, not just the names inside the boxes.
What structure makes cheap How should a system choose between a table, a diagram, a timeline and plain prose? Not by a general belief that images help. Different representations make different operations cheaper for a receiver doing a particular task. A table aligns exact values and labels for comparison, provided units and missing values remain visible. A timeline makes temporal order and overlap inspectable, provided its scale and precision are honest. A diagram exposes states and directed connections, provided its arrows mean what the source says. Prose can carry scope, argument, ambiguity and qualification without forcing everything into a box. An interactive form can let the reader explore alternatives, but each selectable path may require checks of its own. The right question is what the task needs the reader to recover, and how much work and error the chosen form introduces.
This is an old observation in the study of diagrams. Larkin and Simon (1987) compared informationally equivalent diagrammatic and sentential representations through models of information processing. Their central distinction was between having information available and making an operation on that information efficient. Spatial grouping can reduce search and make an inference more direct. This is an account of computational affordances under specified tasks and operators, not a general experimental finding that readers learn more from pictures. For this book it supplies a mechanism worth testing; reader comprehension, delayed recall and transfer still require separate measurements.
So the question for any passage is never “which medium is best?” but “which relationships does this representation make cheap to perceive, and are those the relationships this passage is about?” The defaults that follow are starting points to be tested, not laws:
a process with topology → perhaps a diagram a comparison across options → perhaps a table a chronology with intervals → perhaps a timeline a hierarchy → perhaps a tree a spatial arrangement → perhaps a visual a qualified argument → probably leave it alone Note the asymmetry. Five kinds of structure suggest a candidate form; a qualified argument suggests caution. That is deliberate, but not because prose is universally the highest-bandwidth medium. Formal tables, equations and diagrams can express conditions precisely too. The risk is that an attractive conversion may omit the qualifier, settle an ambiguity or draw causation where the source reported association. The conservative baseline is the unchanged passage, and no alternative should replace it merely because the system can generate one. How often that baseline wins is an experimental question, not a result to announce in advance.
The multimedia principles from Chapter 3 now suggest checks rather than automatic verdicts. Keep labels and explanations near the elements they refer to, or their separation may increase integration effort. Do not assume that repeating every sentence in a diagram helps; unnecessary duplication can raise load, although alternative text and accessible equivalents can be essential rather than redundant. Remove decoration that does no work for the task. Use signalling to reveal a source-supported distinction, never to make an invented relationship look authoritative. These are design hypotheses and constraints to test with readers, not substitutes for measuring reconstruction.
Choosing, and the right to decline A system that enhances has to make two decisions for each passage: whether a different form is worth trying, and whether the generated instance is safe and useful enough to present. First identify the task and the relationships and qualifiers that must survive. Then propose a form, check the output against the passage, and compare its expected benefit with attention, generation and checking costs. If the source is ambiguous or verification fails, ask, show a clearly marked provisional view, or leave the source unchanged. In outline:
source passage ↓ what is it doing? (process, comparison, chronology, hierarchy, argument) ↓ candidate forms ↓ expected benefit against cost and distortion risk ↓ reveal · re-express · add · or leave it as it is Two separations keep this honest.
Choosing the form and executing it are different skills. A system can select an appropriate diagram and then draw a wrong one: the form choice may be sensible while the instance fails fidelity. Conversely, a correct diagram may offer no benefit over already clear prose. Evaluation needs three distinct results: whether the choice fits the task; whether the instance faithfully carries the required claims, relations and exceptions; and whether a receiver actually performs better with it. A weak generator must not condemn the whole medium, and a polished output must not conceal a failed selection.
Fit is not preference. As Chapters 2 and 3 established, confidence and liking do not demonstrate faithful reconstruction. Measure task answers, relation and exception errors, time, effort and preference separately. Where the intended gain is durable learning rather than a one-off lookup, measure delayed recall or transfer too. A view can save time for a reader who knows the notation and impose extra work on one who does not; the tested population and task belong beside any reported gain.
Leaving a passage unchanged is a first-class outcome, not a failure to generate. But abstaining everywhere would also fail the goal. A good system must recognise both when a transformation is worth its verification cost and when the original already serves the reader best.
Generation is weaker than it looks A useful warning comes from DiagramEval (Liang and You, EMNLP 2025). The study evaluates LLM-generated research-paper diagrams against reference diagrams, representing textual elements as nodes and their connections as directed paths. Node alignment and path alignment are assessed separately. In its tested models and corpus, path-alignment scores have little correlation with CLIPScore, an image-text similarity metric that can reward superficial visual agreement while overlooking relationships. That matters here because a familiar-looking diagram can still describe the wrong process. But the paper does not test our payments-policy passage, our readers or the benefit of showing them an enhancement; nor does matching a reference diagram by itself prove fidelity to the source. It motivates relation-level verification, not a blanket claim that every generator draws arrows badly.
For a system choosing among forms, the lesson is to treat generation risk as a cost and check the failure mode of the chosen representation. A table can be checked cell by cell, but it can also swap columns, lose units or silently fill a missing value; a diagram can preserve all the boxes and reverse one decisive arrow. An automated graph score, a visual similarity score and a model’s own approval are not interchangeable with checking claims against the original passage. Verification should distinguish element identity, relation direction, attached conditions and unsupported additions, with unresolved cases visible to the reader.
When enhancement helps and when it hurts Four kinds of passage pull a system in different directions, including one where doing nothing is right. They also cover the failures worth watching for.
Prose that may already be sufficient. The argument and its reasoning are explicit. Leaving it alone is the baseline; an enhancement has to demonstrate additional benefit. An important exception inside prose. A conclusion holds except under a stated condition: a scope limit or a version-specific caveat. A light highlight may be safer than redrawing it. Any diagram must attach the exception to the correct relation. A comparison with exact values. Options differ across figures, units, assumptions and missing observations. A table is a strong candidate when it preserves all of them, not an automatic winner over a well-authored source table. A process with topology. States, transitions, cycles and exception paths: the payments retry policy. This is a good place to test whether a faithful diagram reduces decision time or errors against an equally careful textual baseline. Across these cases, failures occur at different boundaries: misclassifying the passage, selecting an unsuitable form, omitting a required exception, reversing a relation, losing units, introducing an unsupported inference, presenting an unverified addition as source material, or making the result more effortful to navigate. These should be recorded separately. A system that diagrams everything fails by over-transformation; one that changes nothing may fail by missed opportunity. A diagram whose arrow is wrong has failed fidelity even if readers complete the task quickly. A faithful diagram that slows them down has not established a usefulness gain.
A test the Semantic Browser should be able to fail The payments retry passage gives us a small proposed experiment, not an accomplished result. Freeze several passages with different structures, including an already clear one, a qualified exception, a numerical comparison and a process with a branching failure path. For each passage, prepare a checked list of claims, relations, conditions, units and unknowns that must survive. Keep the same source version across conditions and have a person audit the list before generation.
Compare unchanged prose, a carefully authored alternative, AI-selected and AI-generated enhancement, and AI with permission to decline. Hold the reader’s task fixed and use counterbalanced assignment so that seeing the same answer repeatedly does not teach the test. Separate evaluation into two gates. First, assess each candidate’s source fidelity: missing obligations, inverted edges, changed quantities, unsupported additions and whether its provenance still reaches the passage. Then measure receiver outcomes on the candidates that qualify: correct answers about the third retry, identification of both exceptions, time and effort. Record preference independently. For learning tasks, add a delayed or transfer question.
The baseline matters. Beating a dense paragraph after three rereads is not enough; the enhancement must be compared with a well-edited non-generative explanation that a competent author could have provided. A helpful view with an invented exception fails the fidelity gate. A flawless diagram that does not improve reader performance may be attractive, but has not earned its generation and checking cost. And an apparent gain from one reader on one passage is a case observation, not evidence that the router should diagram processes in general.
This chapter proposes that test; it reports no book-specific outcome from it. The later preservation chapter makes the obligations more precise, and the Semantic Browser must eventually run the comparison rather than merely promise it.
Responding to the task, not building a model of the person One boundary needs stating now, because it will matter a great deal later. The proposed enhancer conditions on the source and a current task the reader has declared, or can inspect and correct. It need not retain a profile, infer a stable ability or turn a temporary question into a durable preference. That makes it task-adaptive, not personalised. Task context can still be wrong or sensitive: a guessed task must not be silently promoted to an instruction, and permission to inspect a document does not grant permission to send it to an arbitrary model or service. Durable person models, broader context admission and adaptation across sessions belong to later parts of the book.
task-adaptive representation ≠ personalised representation Where this leaves us Enhancement can help a reader inside a source in three distinguishable ways. Revealing changes the emphasis on source material. Re-expressing proposes a different form for source-supported structure. Adding supplies external context or new analysis and must be labelled and justified separately. None is harmless by definition. The first two owe the reader fidelity to claims, relations, conditions and uncertainty; the third owes independent provenance and an appropriate evidential status. Form choice is a task hypothesis, not a guarantee of benefit, and the unchanged source remains an available, sometimes superior, representation. The proposed test asks whether anything better was actually earned.
Notice the remaining problem. If every requested view independently interprets the same passage, Monday’s state diagram can preserve an exception that Friday’s table omits. Both can still be checked against the source; the problem is that the work is repeated and inconsistencies are easy to miss. Keeping one derived interpretation might make the obligation easier to inspect across views, but it would also concentrate mistakes: a wrong intermediate record could spread the same error everywhere. The next chapter investigates that trade, asking whether a traceable semantic record between source and rendering can improve consistency without pretending to become the source of truth.
References Larkin, J.H. & Simon, H.A. (1987). Why a diagram is (sometimes) worth ten thousand words. Cognitive Science, 11(1), 65–100. Liang, C. & You, J. (2025). Evaluating LLM-Generated Diagrams as Graphs. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 12678–12690. DOI 10.18653/v1/2025.emnlp-main.640. (Introduces the DiagramEval metric; the published title differs from the preprint title.) Mayer, R.E. (2021). Multimedia Learning (3rd ed.). Cambridge: Cambridge University Press. (Source of the coherence, redundancy, contiguity and signalling principles used here as engineering constraints.)