Across languages it has been shown consistently, that in three-site ambiguities such as (1) attachment to N2 is strongly dispreferred whereas N3 and N1 are variably acceptable depending on the particular construction as well as on the language investigated.
Various explanations for this phenomenon rely on the assumption that the three potential attachment sites differ in accessibility (e.g. Lewis, 2000, Hemforth, Konieczny, & Scheepers, 2000a,b). But differences in the salience or activation of the three sites have never been directly tested.
In a series of repetition priming experiments we probed for the three
potential referents of the relative pronoun at different positions in the
relative clause. However, in all experiments we only found recency effects
(N3 
N2 N1) not matching the attachment preferences. From this we may
conclude that salience of the antecedents does not play a role in
RC-attachment. But it may also be that the priming task does not tap into
the right kind of processes.
Therefore we recently applied a different task. Subjects were presented with unambiguous sentences like (2) on the computer screen. Number marking on the three NPs and the verb in the RC disambiguated the attachment. They then had to read them aloud. After that the sentence disappeared from the screen and subjects had to recall it immediately. Reading aloud as well as recalling were tape recorded and transcribed.
The results from the error patterns in this task are highly revealing. Not only did subjects adjust N2-attachments to N1 or N3 attachments more often than vice versa. We counted how often one of the NPs was not recalled. This was reliably more often the case for N2 than for N1 or N3, closely mirroring the attachment preferences.
When sentences have to be recalled, they have to be reconstructed. Obviously, the trace of N2 is less accessible than that any of the other NPs.