START Conference Manager    

ASIST 2012 Annual Meeting 
Baltimore, MD, October 26-30, 2012 


A Tag-Based Framework for Extracting Spoken Surrogates
Hyun Hee Kim

Monday, 6:30pm


Summary

This study investigated the usefulness of a user-assigned tag method, a maximum marginal relevance (MMR) model of reducing redundancy, a pitch method, and a sentence location approach for speech summarization. Based on our investigation results, we proposed a modified MMR (m-MMR) model for automatically summarizing video speech transcripts that not only reduces redundancy, but also utilizes social tags, title words, and sentence positional information. We used title words as well as tags in our m-MMR model not only because tag sets only cover 70% of title word sets, but also because tags pose sparsity problems. This study demonstrated that our proposed method outperformed an m-MMR-based content keyword method that select key sentences using content keywords instead of tags and title words. Elaborating on our results, we discussed the implications of our findings as an effective theoretical basis for the development of a hybrid automatic speech-summarization system that utilizes acoustic/prosodic features as well as textual and structural ones.