Syntactic priming is generally understood to occur between instances of the
same syntactic construction (e.g., Bock, CogPsy 1986). Here we present a
corpus study which investigates priming effects between similar but
non-identical constructions. We find that priming occurs between NPs of
different forms, and that priming strength is linked to the similarity of the
two NPs.

Adaptation (Church, COLING 2000) measures the increase in production
frequency due to priming, and has been used to study structural
priming in corpora (Dubey et al., EMNLP 2005). As adaptation (or
"self-adaptation") measures the priming effect between two occurrences
of a rule X, we therefore introduce a new measure, "cross-adaptation",
which gauges the priming strength between two different structures X
and Y.

We measure cross-adaptation between pairs of NPs which are either (a)
coordinate sisters, (b) from the same sentence or (c) from adjacent
sentences. We use two parsed corpora, one of written text (the Brown
corpus) and one of unscripted spoken text (a section from the BNC). In
both corpora, we find that the magnitude of cross-adaptation and
self-adaptation are similar in each of the three conditions,
indicating that syntactic priming is not restricted to syntactically
identical constituents.

Next, we examine whether adaptation strength is linked to construction
similarity. For each NP, we record the lexical categories of the words
it dominates (e.g., "DT NN" for "the dog"), and create a vector of
these part-of-speech tags. The similarity of two NPs can then be
measured as the Euclidean distance between their vectors. Removing
matching pairs from the dataset, we find a strong and significant
correlation (all p < 0.0001) between NP similarity and adaptation
strength over both corpora and in all three conditions (see table
below). This result holds even with other measures of similarity
such as edit distance.

We conclude that the occurrence of a structure primes not only the
re-occurrence of that exact structure, but also the occurrence of a
set of similar structures. This suggests that priming of similar
constructions should also lead to facilitation in an experimental
setting, a prediction which remains to be tested in future work.

Condition          Correlation

BNC (spoken)       r     N
Between-Sentence   0.51  16682
Within-Sentence    0.50  15996
Within-Coordinate  0.41    233

Brown (written)    r     N
Between-Sentence   0.45 101856
Within-Sentence    0.45  85335
Within-Coordinate  0.44   1392