142
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
5
10
15
20
25
precision
k
painter-movement
categorization
dm
pm
pm-stem
pcm
Figure 6.13. Precision for categorization of the painters.
of
PM
and especially
DM
is still good. The results of
PCM
indicate that when the
precision of the intermediate mapping is low (35%), the use of related instances
does not improve the results. In this experiment we even observe a deterioration
of the performance. Here
DM
clearly outperforms
PM
. This can be explained by
the fact that using
PM
considerably less painter-movement pairs could be extracted.
We expected the recall of
PM
to increase when applying stemming on the names
of movements and the texts extracted [Porter, 1980]. Although the number of pairs
extracted slightly increases, the precision does not improve (Table 6.15).
6.4.3 Tagging Instances
In this subsection, we focus on two case-studies on the tagging of instances related
to the methods described in Section 6.2.3. We compare the extracted lists of tags
with ground truth extracted from a social website. No previous work is known to us
in this field. We therefore present two exploratory studies in the automatic tagging
of instances. In the first experiment, we tag the set of 224 artists and evaluate
the tagging using
Last.fm. The second experiment focusses on books, where the
results are compared with data from
LibraryThing.com.
Tagging Musical Artists
In this experiment, we focus on the
tagging of the 224 artists as done by the
Last.fm
community using the method described in Section 6.3. Using a large set of artists,
we select the 248 most frequently applied tags after the normalization procedure
11
.
11
The list of tags used can be found at http://gijsg.dse.nl/tags224.html
6.4 Experimental Results
143
We investigate whether our method is well suited to label the 224 artists and com-
pare the results with the tags as applied by the
Last.fm users.
The previous experiments showed that
PM
was the most successful alterna-
tive to identify artist similarities, while
DM
outperformed
PM
with respect to the
labeling of artists with genre names. We hence use
DM
to find the co-occurren-
ces between artist names and tags and reuse the results from
PM
to identify the
artist similarities. For fairness, the pages from
Last.fm and Audiocrobbler.com are
excluded from the search results.
Per artist in the test set, an average of 79 tags was identified using
DM
. All
tags in the test set were linked to at least one artist, however not for all tag/artist
combinations a score could be identified, as not all artists are related to one another.
We compare the computed ranking of the tags for the artists with a normalized
ranking as identified by the
Last.fm users as described in Section 6.3. For instance,
the terms
’Rocker’ and ’rock’ have the same normalized form.
We evaluate the computed rankings for the different values of
w as follows.
We first evaluate the precision and recall for the highest ranked tags and secondly
compute Spearman’s rank correlation between the computed ranking and the one
from
Last.fm.
Precision and Recall. We selected the set
S
Dostları ilə paylaş: