Community-Contributed Media Collections: Knowledge at Our Fingertips - Community-Built Databases: Research and Development - page 35

in two steps. From a photo with user-defined tags, an ordered list of candidate tags is
derived for each of the user-defined tag, based on co-occurrence. Then, the lists of
candidate tags are aggregated and classified to generate a ranked list of recommended
tags. The co-occurrence between two tags is computed as the number of photos
where both tags are used in the annotation. The obtained value is normalized with
respect to the overall frequency of the two tags individually. Two measures have
been proposed to normalize the tag co-occurrence: symmetric and asymmetric
measures. The first, according to the Jaccard coefficient [ 33 ], is defined as the size
of the intersection (co-occurrence of the two tags) divided by the size of the union of
the two tags (sum of the frequencies of two tags). This measure can be exploited to
identify equivalent tags (i.e., tags with similar meaning). By contrast, the asymmetric
measure evaluates the probability of finding tag t j in annotations, under the condition
that these annotations also contain tag t i . For each user-defined tag, an ordered list of
candidate tags is derived from the collective knowledge (i.e., user-generated content
created by Flickr users). The larger the collective knowledge, the more relevant and
useful the list of candidate tags. Given diverse lists of candidate tags, they are merged
in a single ranked list by means of two strategies: voting and summing. The first one
computes a score for each candidate tag, and the ranked list of recommended tags is
obtained by sorting the tags according to the number of votes. On the other hand, the
summing strategy computes for each candidate tag the sum of all co-occurrence
values between the considered tag and the user-defined tags.
To evaluate the recommendation system proposed in [ 31 ], 331 photos with at
least one user-defined tag have been analyzed. 131 photos were used as a training
set, while the remaining 200 photos were the actual test set. Experimental results
show the effectiveness of the proposed recommendation system in selecting rele-
vant tags. For almost 70% of the photos, the system suggests a good recommenda-
tion at the first position of the ranked list, and for 94% a good recommendation is
provided among the top 5 ranked tags.
2.5.2 Video Annotation
To extend the accessibility of video materials and enhance video querying, manual
or automatic annotation is needed. Since manual annotations often reflect personal
perspective, videos may be tagged very differently by different users. However, the
study reported in [ 25 ] suggested that user interaction with multimedia resources
within social networks could help generate more consistent and less ambiguous
tagging semantics for video content. In particular, it has been observed that when
multiple users are allowed to label content over a long period of time, stable tags
tend to emerge [ 34 ]. This can be thought of as a form of user consensus built by
letting users interactively correct tags, similarly to wiki pages, thus providing more
reliable metadata. This form of “collaborative tagging” (see taxonomy in Fig. 2.2 )
has been investigated by inferring semantics for the content from user behavior,
both explicitly (i.e., through direct user input) and implicitly (i.e., by monitoring