Study One investigates whether very strong collocations exhibit priming by comparing the time taken to recognise target nouns when they follow primes with which they are frequently collocated (as attested in BNC data) with the time taken to recognise them when they follow words with which they do not commonly co-occur.
Materials
On the basis of frequency data from the BNC, two sets of 40 prime-target word pairs were created. One set (the ‘collocate condition’) comprised pre-modifier primes (adjectives and noun-modifiers) and target nouns which appear directly adjacent to each other in the corpus with sufficient frequency to qualify as strong collocations. The second set (the ‘unrelated condition’) was created by re-arranging these primes and targets to form new pairs which are not found directly adjacent to each other in the corpus. Since priming seems most likely to take place when the appearance of the first word strongly predicts the appearance of the second, mutual information – which measures just this predictability – was used to gauge strength of collocation. Though we saw in Section 4.4 that conditional probability provides a slightly more accurate prediction of normative word association than MI, the latter measure was preferred as the primary criterion for item selection because it is so widely used in the corpus- linguistic literature. An analysis based on MI is therefore likely to be more meaningful to the corpus linguistic community as a whole.
Since mutual information can become unreliable for low frequency items, pairs were only included if they were attested at least ten times in ten different texts within the BNC. These criteria should also help to ensure that all collocations selected are known to the majority of speakers, rather than reflecting the usage of specialist groups. All prime-target pairs in the collocate condition had a mutual information score of at least 8.16 (max = 12.41, Mdn = 9.28) and a t-score of at least 3.49 (max = 42.66, Mdn = 10.78). Only words of between four and seven letters in length were included in the lists, and no very low frequency words were included. All primes are attested in the BNC between 1,000 and 20,000 times; all targets are attested between 1,000 and 37,000 times.
For each prime word, two non-words of between four and seven letters were also generated using the ARC Nonword Database (Rastle, Harrington, & Coltheat, 2002). Each priming word was paired with two nonwords. This provided 160 prime-target pairs in all: 80 word-word pairs, of which 40 were unrelated and 40 collocating pairs, with each target and each prime word appearing once in each condition; and 80 word- nonword pairs, with each prime appearing twice. This list was divided into two blocks, in which each target appeared once and each prime appeared four times – once in the related, once in the collocate and twice in the nonword conditions. The two blocks are shown in Appendix Ai.
Participants
22 students at the University of Nottingham. All were native speakers of British English.
Procedure
Participants were tested individually in a quiet room. Presentation of the stimuli and recording of the reaction times were controlled by Psychology Software Tools’ E-
Prime software and items were displayed on a CRT monitor. On each trial, a fixation
point (‘+’) was presented, centred on the screen, for 2,000ms. This was replaced with the priming word, which was presented in lowercase letters for 600ms. The prime was then immediately replaced by the target, also in lowercase letters. The target stayed on the screen until the participant made a response or for 2,000ms, whichever was sooner.
word and ‘N’ if it was a nonword. They were told to make this decision as quickly as possible. Reaction times were measured from target onset to response. Participants received 30 practice trials, sampled from across the conditions. There was then a break, at which point participants were invited to begin the trial proper in their own time. They were then presented with one block of 80 trials before a second break and then the second block of 80 trials. Order of presentation within blocks was pseudo- random. An initially randomised order was adjusted to ensure that for each of the four conditions (prime-related target; prime-unrelated target; two x prime-non-word) there was an equal number of primes appearing for the first, second, third and fourth time in the block. This was to offset any possible effect of repetition priming between primes. The order of presentation of the two blocks was counterbalanced across participants.
Results and discussion
Reaction times of less that 250ms and greater and 1,250ms (2.98% of the total) were excluded from analysis. Mean accuracy by participant, collapsed across conditions was 96%. There was no effect of condition upon accuracy.
Reaction times for correct responses (which were not normally distributed within conditions) were faster for collocating (Mdn = 546ms) than for unrelated targets (Mdn = 567ms). This difference was significant both in the analysis by participants (T = 28.00, p (one-tailed) < .001, r = -.68) and in the analysis by items (U = 573.50, p (one-tailed) < .05, r = -.24). Collocating primes appear therefore to substantially facilitate the recognition of the target words.
As was discussed above, it is already well known that normatively associated words show a significant priming effect. I have argued that, though normative association and priming are linked, the existing evidence for the former cannot count as evidence for the latter since, on the one hand, many associates are not collocations and, on the other hand, only the most salient of collocations are prominent in word association norms. The present study overcomes the first of these problems: all of the items used here are collocations. The issue of whether priming works for collocations which are not sufficiently salient to appear prominently in word association norms has not so far been addressed, however. The collocating items used in this study were selected purely on the basis of corpus evidence, without reference to association norms. A
post-hoc search for these collocations in the Edinburgh Association Thesaurus (EAT) <http://www.eat.rl.ac.uk/> found that 21 of the 40 collocating pairs in our study are attested as associates, with 10 of these being supplied by at least 5% of respondents. 13 collocations are not attested as associates, and 6 pairs are not classifiable because the prime is not listed in the thesaurus (these classifications are indicated in Appendix Ai). None of the non-collocating pairs were found to be associates.
It should be noted that a pair’s failure to appear in the Edinburgh norms does not show that the words are not strong associates for any of our participants. The Edinburgh database was not collected from the same population of speakers as took part in this experiment, and – as was discussed above – the methodology used to elicit this norming set is likely to have found only the strongest associates. However, it seems likely that word pairs not appearing here will at least not be amongst the most salient associates for the majority of speakers of English. The putatively ‘non-associated’ pairings were: death-penalty deep-sigh foreign-affairs huge-amounts peace-talks private-sector wild-flowers east-coast fatal-error liquid-assets officer-corps silent-movies silk-shirt
In light of this information, we can re-analyse the data to ask whether priming is restricted to collocations which are sufficiently salient to feature in the association norms. To this end, items were re-organised into three groups: unrelated pairs,
as associates. Median reaction times for correct responses in each condition are shown averaged across participants in Table 8.
Table 8: median lexical decision times for accurate responses
unrelated associated collocates non-associated collocates
median RT (ms) 567 534 512
Significant facilitation was found in comparison to the unrelated condition for both associated (T = 30.00, p (one-tailed) < .001, r = .64) and non-associated collocations (T = 43.00, p (one-tailed) < .005, r = .58). However, probably as a result of the small number of items involved, neither of these advantages is statistically significant in an analysis by items (unrelated vs. associated collocates: U = 326.50, p (one-tailed) > .05,
r = -0.18; unrelated vs. non-associated collocates: U = 206.00, p (one-tailed) > .05, r =
-0.15). While the small number of items – and the imperfect nature of the EAT as a check for psychological association in our participant group - must weaken our conclusion somewhat then, we can tentatively conclude that priming may be taking place here even when collocations are not sufficiently salient to feature on normative association lists.