Technology

BraveNewTask: UTwente Searching and Hyperlinking

Description
This is a short presentation to report our experiments and results of brave new searching and hyperlinking tasks for the MediaEval Benchmark Initiative 2012 in Pisa, Italy. The searching task involves finding target video segments based on a short natural language sentences query and hyperlinking task involves finding links from the target video segments to other related video segments in the collection using a set of anchor segments in the videos that correspond to the textual search queries. To find the starting points in video, we only used speech transcripts and metadata as evidence source, however, other visual features (for e.g., faces, shots and keyframes) might also affect results for a query. We indexed speech transcripts and metadata, furthermore, the speech transcripts were indexed at speech segment level and at sentence level to improve the likelihood of finding jump-in-points. For linking video segments, we computed k-nearest neighbors of video segments using euclidean distance
Categories
Published
of 7
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  • 1. Searching Hypelinking LIMAS Evaluation Conclusion UTwente does Brave New Tasks for MediaEval 2012: Searching and Hyperlinking 5 October, Pisa, Italy Danish Nadeem Robin Aly Roeland Ordelman Team UTwente Enschede, NetherlandsBrave New Tasks : Searching and Hyperlinking 1/7 Danish Nadeem Robin Aly Roeland Ordelman
  • 2. Searching Hypelinking LIMAS Evaluation ConclusionSearching Searching for jump-in points involves: • Choice of evidence sources: ASR text for sentences and segments plus meta data. • Ranking: Lucene scores • Jump-in point selection: the beginning of each segment. Ranking function: • Normalized scores of each component • Linear, equally weighted combination to compute the final score.Brave New Tasks : Searching and Hyperlinking 2/7 Danish Nadeem Robin Aly Roeland Ordelman
  • 3. Searching Hypelinking LIMAS Evaluation ConclusionHyperlinking Hyperlinking video segments: • Target links should be topically similar to the source • We experiment with concept representations Concepts representation: • Segments are represented by 508 concepts • For a given source segement, k-nearest target segments are retrieved using Euclidean distance Example: segment1 .. segment2 .. segment3 .. source ..Brave New Tasks : Searching and Hyperlinking 3/7 Danish Nadeem Robin Aly Roeland Ordelman
  • 4. Searching Hypelinking LIMAS Evaluation Conclusion LIMAS architecture (Chen et al., 2012)Brave New Tasks : Searching and Hyperlinking 4/7 Danish Nadeem Robin Aly Roeland Ordelman
  • 5. Searching Hypelinking LIMAS Evaluation ConclusionResult of submitted runs Runs Window size MRR mGAP MASP 60 sec 0.156 0.122 0.085 LIMSIsegment 30 sec 0.155 0.088 0.085 10 sec 0.093 0.033 0.050 60 sec 0.074 0.054 0.111 LIMSIsentence 30 sec 0.073 0.035 0.112 10 sec 0.034 0.002 0.076 60 sec 0.213 0.161 0.124 LIUMsentence 30 sec 0.204 0.131 0.129 10 sec 0.136 0.081 0.122 • LIUM sentences showed best performance for all three measures. • LIMSI segment types performed better on MRR and mGAP measures. • LIMSI sentences showed better performance on MASP.Brave New Tasks : Searching and Hyperlinking 5/7 Danish Nadeem Robin Aly Roeland Ordelman
  • 6. Searching Hypelinking LIMAS Evaluation ConclusionConclusions • Searching • Speech segments have better evidence for retrieval. • Sentences are better for retrieval of jump-in points. • Hyperlinking • Results awaited [?]Brave New Tasks : Searching and Hyperlinking 6/7 Danish Nadeem Robin Aly Roeland Ordelman
  • 7. Searching Hypelinking LIMAS Evaluation Conclusion Thank you!Brave New Tasks : Searching and Hyperlinking 7/7 Danish Nadeem Robin Aly Roeland Ordelman
  • We Need Your Support
    Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

    Thanks to everyone for your continued support.

    No, Thanks