Information Technology Reference
In-Depth Information
Fig. 6.9
Average precision scores for language model parameters: relevance feedback from
hypertext to Semantic Web
performing below the baseline, although its performance increased as the window
size increased, reaching an average precision of 0.6262 with
m
=
,
<
.
05).
However, given that a window size of 10,000 covered most documents, increasing
the window size will not likely result in better performance from
lca
.The
ponte
relevance feedback performed very well, reaching a maximum MAP 0.8756 with
a window size of 300 using
inquery
weighing, and so was insignificantly different
from
inquery
(
p
3
000 (
p
0
>
.
05). Lastly, both
ponte
and
okapi
experienced a significant
decrease in performance as
m
was increased, so it appears that the window sizes
of 300 and 100 are indeed optimal. Also, as regards comparing baselines,
inquery
outperformed
cosine
(
p
0
<
.
05).
For language models, both averaged relevance models
rm
and concatenated
relevance models
tf
were investigated, with the primary parameter being
m
,the
number of non-zero probability words used in the relevance model. The parameter
m
was varied between 100, 300, 1,000, 3,000,and 10,000. Remember that the query
model
is
the relevance model for the language model-based frameworks. As is
best practice in relevance modeling, the relevance models were not smoothed, but
a number of different smoothing parameters for
0
ε
were investigated for the cross
entropy ranking function, ranging from
ε
between 0.01, 0.1, 0.2, 0.5, 0.8, 0.9, and
0.99. The results are given in Fig.
6.9
.
Search WWH ::
Custom Search