2.3.1 - Baddeley’s acoustic and semantic study

The coding question

Memory coding concerns the form in which information is represented, not just how long it is retained.

If a remembered sequence is represented mainly by sound, sound-alike words should be hard to distinguish. If meaning is important, words with similar meanings should cause more confusion. Baddeley’s earlier 1966a work found a large acoustic-similarity effect in short-term serial recall and a much smaller semantic effect. The 1966b classic study asked whether the same pattern occurred when participants learned and later recalled word sequences. It contains three experiments. The 1974 Baddeley and Hitch working memory model is a later theoretical account, not the date or procedure of this study.

The 1966b materials contained four ten-word sequences. A contained sound-alike words, such as “man”, “can” and “cap”; B was its acoustically dissimilar control. C contained similar-meaning adjectives, such as “great”, “large” and “big”; D was its semantically dissimilar control. Each control was matched to its corresponding similar list for word frequency, reducing a familiarity alternative. Comparisons should be A with B and C with D, rather than treating different word classes across A and C as perfectly equivalent.

Separate groups learned different lists. The dependent measure was how many words were reproduced in the correct serial positions. The relevant word sets remained visible on cards; their arrangement was changed to prevent position on a card becoming a sequence cue. This was primarily a test of learning the order, not guessing which words belonged in the list.

The first experiment’s puzzle

The first experiment showed why an immediate learning score cannot simply be treated as pure long-term memory.

The retained sample consisted of young servicemen: 18 in A, 17 in B, 20 in C and 20 in D, after three A participants failed a listening check. Words were played by tape at one every three seconds. Participants had 40 seconds to write the sequence in order. They repeated this for four learning trials, then spent 20 minutes on a task involving immediate recall of eight-digit sequences. An unexpected retest asked for the learned ten-word order. The delay and intervening task were intended to reduce short-term retention.

The acoustically similar list was harder during learning, but unlike the other lists it did not show significant forgetting between trial four and retest. At retest there was no difference between the acoustic groups. The semantic groups did not differ significantly during learning or retest and both showed forgetting. Thus Experiment I alone did not demonstrate the eventual semantic-LTM conclusion.

Baddeley reasoned that the immediate learning trials could use both STM and LTM. For dissimilar words, STM could boost trial-four performance, then cease contributing after the delay. Sound-alike words were already difficult for STM, so their immediate scores might have depended more on what had been learned into LTM; less of a temporary advantage was left to lose. This accounts for poor initial performance with relatively little later forgetting. “No significant forgetting” is not evidence that the acoustic list gave the highest score or that memory never fades.

The controlled timings and explicit order score help replication. However, young servicemen are not representative of all learners. Nor does recalling a word order after twenty minutes capture all long-term memories, which can be visual, emotional, personal or retained over years.

Separating temporary support from learning

The next two experiments tested whether temporary rehearsal had hidden a semantic effect on long-term learning.

Experiment II used the acoustic A and semantic C lists with women from the research unit’s participant panel, described in the original as housewives. Six participants failed the hearing check. Retained group sizes for A were 20, 25 and 23, and for C were 17, 26 and 20. The three procedures were:

Each additional task used six eight-digit sequences. Y markedly lowered learning scores for the semantic list and produced little loss from trial four to retest. In contrast, the acoustic list was relatively unaffected by placing the task before recall, consistent with already having little STM support. Z helped establish that timing of interference mattered, rather than any extra activity producing the same result. However, this experiment lacked the dissimilar B and D control lists, so it could not settle the original similar-versus-dissimilar question on its own.

Experiment III restored all four lists with separate groups of 15, 20, 16 and 21 men and women from the panel. Words were shown on a slide projector for three seconds, with about two seconds for slide changeover. This reduced mishearing as an explanation. Between each presentation and test, participants recalled six eight-digit sequences presented at one digit per second. They then had one minute to reproduce the ten-word order. After four trials there were 15 minutes of self-paced digit copying, followed by retest. This differs from Experiment I’s twenty-minute intervening task.

With STM support reduced, the acoustic difference was only transient: it was significant at trial two, reversed by trial four, and absent at retest. The semantic list learned more slowly than its control, performing significantly worse at trial four and retest. Neither semantic group showed significant forgetting over the final interval. The important semantic effect was therefore poorer learning performance, not proof that similar-meaning words decayed faster during the delay.

What the evidence can establish

The study’s strongest inference comes from the changing controls across the experiments, rather than a slogan that STM is sound and LTM is meaning.

Similarity was manipulated, presentation was standardised, frequency was matched within each comparison, and Experiment III reduced hearing error and temporary rehearsal. These changes make a coding explanation more credible because they target specific alternatives. The use of separate groups avoids one group practising all four lists, but participant differences between groups could still contribute; random allocation is not established by the original report and should not be invented.

Order recall gives an objective numerical outcome and permits non-parametric comparisons. Yet it is a narrow operationalisation of LTM. Visible word cards, repeated learning and short delays differ from recalling an autobiographical event without prompts. Reliable order scoring does not guarantee ecological validity, or prove that the same effect is equally large in every type of long-term memory. The panel and service samples also constrain generalisation across ages and cultures.

The study supports different coding emphases and therefore challenges a simple account in which STM and LTM are identical processes distinguished only by duration. It does not on its own prove the full multi-store model, all components of working memory, or a particular brain location. Baddeley’s discussion allows acoustic learning through an intermediate STM contribution and says semantic coding is extensive, not exclusive. Distinguishing those claims prevents evidence being asked to establish more than its procedure tested.

The original 1966b publication, pp302–309, is the study authority. Pearson’s selected-study summary is useful support, but the original separates Experiment I’s puzzle from the final conclusion and identifies all four lists in Experiment III.