ADR 0006: Semantic priming is rejected for the review queue

Status: Accepted

Context

The Semantic Relatedness track (#200) will give LensWord a knowledge graph of synonyms, antonyms, collocations, and topics (#202, #203). It is tempting to spend that graph on the review queue itself: show a related word shortly before or after its target, on the theory that activating a related concept should make the target easier to recall. That is semantic priming, and this ADR records the decision not to build it, with the evidence for and against.

The evidence for rejecting it

• Magnitude is small. The Semantic Priming Project (1,661 targets, 768 subjects) measured first-associate priming at 26 ms at a 200 ms stimulus-onset asynchrony and 20 ms at 1200 ms in a lexical decision task; naming produced only 6 ms.
• It decays within seconds. Deacon et al. (1999) found the priming ceiling at roughly 2 seconds or one intervening item — an order of magnitude shorter than any realistic gap between review-queue cards.
• It requires an existing lexical representation. Priming activates a connection between two words someone already knows. A word being learned has no such representation yet, so the mechanism cannot apply to exactly the population a spaced-repetition queue spends most of its time on: new and weak items.
• It produces no retention benefit. Even twelve primes do not cause long-term semantic priming, while repetition priming was obtained in the same experimental design — the two effects are not interchangeable, and only one of them lasts.
• The scheduler hazard is the operative reason, not the small effect size on its own. FSRS reads answer accuracy as its difficulty/stability signal (ADR 0004). If priming raises in-session accuracy without raising retention — which is exactly what "no long-term benefit" means — a primed card would report a false difficulty/stability signal, inflating its interval as if it had been genuinely strengthened. This is a scheduler correctness problem, not merely an inefficient feature: FSRS's persisted stability (ADR 0004) exists specifically to reflect actual review history, and priming would feed it a number that does not correspond to demonstrated retention.

This is specifically about semantic priming. Two related phenomena are not covered by this rejection and are not conflated with it:

• Repetition priming — literally re-showing the same item — persists from hours to roughly a year, and is not the same mechanism at all. It is already what a review queue does by design.
• Social priming — behavioral effects from incidental cues — is unrelated to lexical/semantic priming and separately has a poor replication record; it is mentioned here only to avoid conflating the two "priming" literatures.

The evidence against interleaving/clustering, considered honestly

A different, related idea is worth naming so it does not get smuggled in under "relatedness": clustering similar items together (or deliberately interleaving them) to aid discrimination. The case for this is weaker than it is often presented, and this ADR records the counter-evidence rather than only the evidence that supports the decision already made:

• A meta-analysis found d = 0.73 [0.41, 1.05] on trials-to-criterion (interleaving helps you get there faster) but d = −0.24 [−0.71, 0.23] on post-tests (a confidence interval that crosses zero — no reliable post-test effect).
• Vote count across the underlying studies: 6 negative, 2 conditionally negative, 1 null, and 4 positive — not a one-sided literature.
• Ishii (2015)'s result is confounded by visual similarity between items, not semantic relatedness specifically.
• Sun & Fang (2021) found the opposite temporal profile from the studies the trials-to-criterion result is drawn from.

Read together, the trials-to-criterion effect is real but the durable learning effect is not established, and specific studies have confounds that limit how far the positive result generalizes. This is not treated as a case for building clustering/interleaving in the review queue either — it is recorded so a future proposal has to engage with the actual state of the evidence rather than a rounded-up summary of it.

Decision

LensWord will not implement semantic priming, prime-target pre-exposure, or any relatedness-based reordering of the review queue.

The knowledge graph shipping in #202/#203 is for querying relatedness on demand (a "related words" panel, diagnosis features in #180's track), not for silently changing what order cards are shown in or which card follows which.

Boundary with #180 §3 (contrast and interference)

#180 §3 rules that separating vs. interleaving similar concepts "must [be chosen] based on observed errors and learning stage rather than one global rule." A later phase of this track (Phase 3, not implemented by this ADR) proposes showing a contrasting word only at first introduction of a new word, before any observed errors exist — a learning-stage rule, not a global one, and therefore of exactly the kind #180 §3 requires rather than an exception to it. This boundary is recorded here and confirmed on #180 directly; it does not authorize Phase 3's implementation, only clears the conflict a reader of both issues would otherwise expect.

Boundary with #185's closed strategy catalog

#185 declares its intervention catalog closed, with contrast as one of its nine members. A later phase of this track (Phase 5, not implemented by this ADR) supplies the mechanism behind contrast — an adjacent contrast card, an ask-to-relate framing, and FSRS grading isolation so a contrast card's answer never mutates the target word's own review state. It does not add a new catalog member and does not decide which intervention to run; #185 keeps sole ownership of the strategy catalog and of intervention selection.

Boundary with #176's queue ordering

#176 owns in-session queue order through its self-adjacency rule (a minimum number of intervening items between repeats of the same word, enforced by LearningStepScheduler). No phase in the Semantic Relatedness track reorders the review queue. If a future phase in this track ever needs to influence queue order, it must go through LearningStepScheduler rather than add a second, competing ordering component.

Phase 5 implementation boundary and pre-registration (#206)

Phase 5's contrast card is a pair presentation, not a queue item: both words are carried by one ContrastCard, the prompt asks the learner to describe the difference, and the answer path does not invoke FSRS or mutate due_at. The pair is eligible only when both words have persisted FSRS stability at or above the provisional default of 21 days. That threshold is configurable because Nation's evidence does not identify a principled cutoff; it is an experimental setting, not a claim about the correct value.

The existing semantic-relatedness opt-in and the new contrast-card sub-setting are both required, and both default off. A diagnosis planner's isolate decision suppresses a graph-derived synonym/antonym pair; a planner-selected contrast decision takes precedence over that fallback. No contrast response is recorded as a standard review observation until a later measurement phase defines how to attribute its delayed outcome.

Before measurement begins, the prediction is registered as follows: compared with matched uncontrasted pairs at one week, contrast cards are expected to improve discrimination accuracy while reducing in-session accuracy. A null or negative one-week result is a valid outcome; if the confidence interval does not support a positive delayed discrimination effect, the feature will be reported as null/negative and removed rather than retained because the theory sounds plausible. This prediction does not treat the expected in-session cost as a regression.

Consequences

Positive

• Closes off a class of feature request ("add priming/relatedness to the queue") with a citable answer instead of a re-derivation each time it is proposed.
• Protects FSRS's stability signal (ADR 0004) from a source of accuracy-without-retention noise that would have been difficult to detect after the fact, since it would look like the scheduler working correctly.
• Keeps queue ordering, intervention selection, and the strategy catalog each owned by exactly one component (LearningStepScheduler, #180's diagnosis/intervention layer, #185's catalog respectively), rather than the knowledge graph adding a fourth, overlapping opinion about what to show next.

Negative

• The knowledge graph's most intuitively appealing use — improving recall by showing related words near each other — is the one use this ADR takes off the table. Relatedness will only ever be queryable on demand, not woven into scheduling.
• A genuinely evidence-based case for interleaving (distinct from priming) is not precluded by this ADR, but would need its own ADR built on the counter-evidence recorded above, not a re-citation of only the supportive half of the literature.

Addendum: the deck-boundary rule (issue #203 TODO 6)

knowledge_graph.py's edge derivation resolves a lexical association (a synonym string, an antonym string, a topic tag) against words the same learner owns, by exact casefolded term match. A synonym string that names a word not in the learner's deck produces no edge — it is vocabulary they do not study, and an edge to it would point at nothing a query can follow.

Kept as-is, with the consequence stated explicitly: this caps the feature's recall. A learner who tags "generous" with the synonym "magnanimous" gets no edge unless "magnanimous" is also a card in their own deck. The alternative — adding unowned lexical nodes as a second node type, so an edge could point at a word the learner has not added — was considered and rejected for this phase: it would mean the graph answers questions about vocabulary nobody is studying, which is a different and larger feature (effectively a second, unbounded vocabulary source) than "relate the cards I already have," and it multiplies every downstream query (related, prerequisites, distractor selection) with a node-type check it does not currently need.

If recall becomes a real limitation later, the fix is additive — a new node type — not a change to how owned-word edges already work, so this decision does not need to be revisited to make that possible.

Sources

• Hutchison, K. A., et al. (2013). The Semantic Priming Project. Behavior Research Methods, 45(4), 1099–1114. (1,661 targets, 768 subjects; first-associate priming 26 ms at 200 ms SOA, 20 ms at 1200 ms in lexical decision; 6 ms in naming.)
• Deacon, D., Hewitt, S., Yang, C.-M., & Nagata, M. (1999). Event-related potential indices of semantic priming using masked and unmasked words: evidence that the N400 does not reflect a post-lexical process. Cognitive Brain Research, 8(3), 293–307. (Priming ceiling at ~2 seconds / one intervening item.)
• The "even twelve primes do not cause long-term semantic priming, while repetition priming was obtained in the same design" result distinguishes semantic from repetition priming's durability.
• Meta-analysis of interleaving vs. blocking: trials-to-criterion d = 0.73 [0.41, 1.05]; post-test d = −0.24 [−0.71, 0.23]; vote count 6 negative / 2 conditionally negative / 1 null / 4 positive.
• Ishii, T. (2015). Visual-similarity-confounded interleaving study (confound noted, not a clean relatedness result).
• Sun, X., & Fang, X. (2021). Interleaving study with a temporal profile opposite the trials-to-criterion literature above.
• ADR 0004 (docs/adr/0004-memory-scheduling-model.md): persisted FSRS stability and why it must reflect genuine review history.