We describe the runs for our participation in the Search sub-task of the Search and Hyperlinking Task at MediaEval 2013. Our experiments investigate the a ect of using information about speech segment boundaries and pauses on the e ectiveness of retrieving jump-in points within the retrieved segments. We segment all three available types of transcripts (automatic ones provided by LIMSI/Vocapia and LIUM, and manual subtitles provided by BBC) into xed-length time units, and present the resulting runs using the original segment starts and using the potential jump-in points. Our method for adjustment of the jump-in points achieves higher scores for all LIMSI/Vocapia, LIUM, and subtitles based runs.
The constant growth in the size and variability of digital multimedia content being stored requires the development of techniques that not only identify les containing relevant content, but also bring the user as close as possible to the beginning of the relevant passage within this le to maximize the e ciency of information access. This starting point, referred to as the jump-in point, cannot simply be related to the locations of the words of interest being spoken, since the user may need to listen to the whole utterance in which the words were used, or slightly bigger passages, in order to get the idea of the context. Thus we assume that these jump-in points should occur at the beginning of the speech segments or utterances, and might be expressed by a pause in the speech signal. This idea underlies our experimental setup. We create one retrieval run for each xed-length segmentation unit, but present it in two ways for further comparison: with the initial boundaries of the segments, and with adjusted jump-in points.
The Search and Hyperlinking Task at MediaEval 2013
uses television broadcast data provided by BBC, and
enhanced with varying additional content such as automatic
speech recognition (ASR) transcripts [
The task was evaluated using three metrics: mean
reciprocal rank (MRR) which scores the rank of the retrieved
segment containing relevant content, mean generalized
average precision (mGAP) which combines the rank of the
relevant segment and distance to the ideal jump-in point at
the start of the relevant content within the segment [
As the les in the collection vary in style and length, we decided to segment all the content into xed length units. For these experiments we chose three values for segment length: 60, 90, and 180 seconds. These time units were the same for all types of transcripts. However, the transcripts do not always cover the spoken content in the same way: the ASR system might recognise some noise as words, or humans who create the manual subtitle transcripts might consider certain parts of the video non relevant for transcription. This explains the di erence in the number of documents for diverse types of transcripts and time units given in Table 1.
At the segmentation stage we stored the information about
potential jump-in points within each segment in a separate
le. The LIMSI/Vocapia transcript contains speech
segments boundaries predicted by their system [
LIMSI { LIMSI + LIMSI* {
LIUM { LIUM { Subtitles { Subtitles { ditionally we created an uno cial run that uses the pauses in the LIMSI transcript in order to be able to make a better comparison with the other types of transcript.
We do not have access to details of the ASR transcription systems. However we can distinguish them by the size of the vocabulary they used for this collection which contain 36,815, 57,259, and 98,332 entries for LIMSI/Vocapia, LIUM, and subtitles respectively.
For indexing and retrieval experiments we used the
opensource Terrier Information Retrieval platform1 [
formation contained in all types of transcripts.
This work was supported by Science Foundation Ireland (Grant 07/CE/I1142) as part of the Centre for Next Generation Localisation (CNGL) project at DCU.