We describe Dublin City University (DCU)'s participation in the Hyperlinking sub-task of the Search and Hyperlinking of Television Content task at MediaEval 2013. Two methods of video hyperlinking construction are reported: i) using spoken data annotation results to achieve the ranked hyperlink list, ii) linking and merging meaningful named entities in video segments to create hyperlinks. The details of algorithm design and evaluation are presented.
This paper presents Dublin City University (DCU)'s participation in the Hyperlinking sub-task of Search and Hyperlinking of Television Content task at MediaEval 2013. The paper is organized as follows: Section 2 describes our automatic hyperlinking strategies, Section 3 gives experimental results, and Section 4 concludes the paper. 2.1
In this subsection we describe the principles underlying
our approach to the hyperlinking task. The elements
involved in the hyperlinking framework correspond to the query
anchor, the target segment, and the hyperlink. The query
anchors, as the input to the hyperlinking framework, are
de ned in [
This strategy determines the hyperlinks based on two
quality measures: the video-level and the segment-level. The
video-level measure aims to determine the relevance between
the video containing the query anchor and other videos
containing potential target segments based on the text
transcripts. DBpedia Spotlight1, implementing text annotation
by supervised learning through DBpedia Ontology2, was
used to extract a set of terms to represent the textual
content of each video. The method used to annotate terms in
DBpedia Spotlight is based on a TF*ICF model [
The segment-level similarity uses Apache Lucene 3.6.23
to determine the relevance between the query anchor and
the potential target segments. The Lucene standard
analyzer was used with the default stop word list4 to index
ASR transcripts and manual subtitles. The search input
query contained all the spoken data contained in the query
anchor. The score calculation mechanism uses a
combination of a Boolean AND function lter and ranking using the
Vector Space Model [
Score = w1Sv + w2Sl
Score = SvSl (1) (2) where Sv is the video-level similarity score, while Sl is the segment-level similarity score. We use a simple linear fusion mechanism to merge the two scores, where the weights w1 and w2 are set to 0.5 respectively. 2.3
This strategy links named entities contained in query anchors and the potential target segments, and then merges these entities to construct hyperlinks. Apache OpenNLP5 1https://github.com/dbpedia-spotlight 2http://dbpedia.org/Ontology 3http://lucene.apache.org/ 4https://lucene.apache.org/core/3 6 2/api/core/ org/apache/lucene/analysis/StopAnalyzer.html 5http://opennlp.apache.org/
1 0.2944 0.7000 0.6567 0.5450
2
0.2935
0.7067
0.6633
0.5383
was used to tag words in the ASR transcripts and
subtitles. All noun words tagged as NN, NP, and NNP were
selected as named entities. To describe and link the named
entities, a vector space model was constructed by
predicting the surrounding words given the current word. We use
word2vec6 to implement a supervised learning mechanism
using a Neural Net Language Model to create the vector
model of named entities. We use the ASR transcripts of
videos gathered from the blip10000 collection [
S = 2(Vi T Vj ) jVij + jVj j where Vi T Vj are the total number of words contained in both Vi and Vj . jVij is the length of the word vector i. All named entities located at the potential target segments are merged using Equation 4 to generate the nal score to obtain the ranked hyperlink list.
Score =
P0<i<k Si
N where Si is the score of an entity in a potential target segment, and N is the total number of named entities in the current segment.
A total of four formal runs were submitted to the Search and Hyperlinking task in MediaEval 2013, described in Table 2. Table 3 shows the Mean Average Precision (MAP) (4) value of each run. This indicates that our hyperlinking strategy based on spoken data annotation performs better. Table 1 shows P@N and MAP value of Run 3. MAP and P@N benchmark have received a good result in most runs except Topic (Anchor) 27, which describes Shakespeare and Global Theatre. A total of two other videos are related to Shakespeare and Global Theatre, while the content is presented in terms of a cartoon. The lack of visual elements leads to hyperlinks to cartoon segments, while real users will notice the unrelatedness between TV shows and cartoons. 4.
This paper presented details of DCU's participation in the TV Data Hyperlinking task of MediaEval 2013. The evaluation shows that annotating spoken data to construct hyperlinks achieves better results. In our future work, we will examine the use of visual cues to improve hyperlinking performance. 5.
This work is funded by the European Commission's Seventh Framework Programme (FP7) as part of the AXES project (ICT-269980).