A Controlled Experiment on the Process Used by Developers During Internet-Scale Code Search - Finding Source Code on the Web for Remix and Reuse - page 62

Fig. 4.3 Distribution of queries per session

combined various tokens from the original into one and left the rest unchanged, or
it led to the adding of a filter.
A reformulation (R) refinement had a new search string that contained some
tokens in common with the original but was neither a generalization nor a special-
ization. It consisted of moving a token into a filter and vice versa or changing a filter
within a search engine.
A new query (N) had no tokens in common with the original.
Spelling (SP) refinement was any query where spelling errors were corrected,
defined as a Levenshtein Distance (LD) [ 11 ] between 1 and 3. Changes with larger
LDs were placed in one of the other categories.
A No Change (NC) refinement was any query that had all tokens in common
with the original.
There were a total of 286 queries issued by the subjects in our study, which led
to 144 refinements. The number of each type of query refinement is as given in
Tab le 4.3 . We analyzed the sequences of query revisions occurring at query n and
its immediate next query, numbered n

-squared test of two-way
contingency tables to identify which sequences of refinements were significant. The
test would indicate if certain combinations of query revisions were more or less
likely to occur from other combinations.

+ 1. We used a χ

Generalization Specialization Reformulation New Spelling No change
13 61 57 4 7 2

Table 4.3: Number of query refinements by type