Community-Contributed Media Collections: Knowledge at Our Fingertips - Community-Built Databases: Research and Development - page 41

Some approaches have been addressing the user interactivity in an effort to im-
prove the relevance of the results. For example, the Koru search engine [ 48 ] allows the
user to automatically expand queries with semantically related terms through an
interactive interface to extract relevant documents. The interface is composed of
three panels: (a) the query topic panel , which provides users with a summary based
on a ranking of significant topics extracted from the query, (b) the query results ,which
presents the outcome of the query in the form of a series of document surrogates, and
(c) the document tray , which allows users to collect multiple documents they wish to
peruse. Real users were asked to experiment with the system to identify the improve-
ments offered over traditional keyword search. The main advantages reported by
testing users were the capability of lending assistance to almost every query and the
improved relevance of the documents returned.
Other approaches have been devoted to improve efficiency, in terms of query
processing scalability, and proficiency in entity recognition. Bast et al. [ 49 ] present
the ESTER modular system for highly efficient combined full-text and ontology
search. It is based on graph-pattern queries, expressed in the SPARQL language,
and on an entity recognizer. The entity recognizer combines a supervised technique
with a disambiguation step to identify concepts in the query and in the documents.
In addition, the system includes a user interface which suggests a semantic comple-
tion based on the typed keywords, and the display of properties of a desired entity.
The interface is designed in such a way as to offer all the features of a SPARQL-
based query engine, with the added benefit of being intuitive for inexpert users. For
example, when a user has typed “Beatles musician”, the system will give instant
feedback that there is semantic information on musicians, and it will execute, in
addition to an ordinary full-text query, a query searching for instances of that class
(in the context of the other parts of the query), showing the best hits for either query.
Good performance in terms of scalability and entity recognition has been achieved
by the proposed system.
2.6.3 Video Content Interpretation
User-contributed video collections like YouTube present new opportunities and
novel challenges to mine large amounts of videos and extract knowledge useful to
categorize and organize their content. Following the taxonomy depicted in Fig. 2.2 ,
Sect. 2.6.3.1 presents a classification technique based on the visual features of video
clips, while Sect. 2.6.3.2 describes a novel approach to synchronize and organize a
set of video clips related to a concert event.
2.6.3.1 Concept-Based Classification
Automatic indexing of video content has already received significant interest as an
alternative to manual annotation and aims at deriving meaningful descriptors from
the video data itself. Since such data is of sensory origin (image, sound, video),
techniques from digital signal processing and computer vision are employed to