20 likes | 154 Vues
This paper presents a novel unsupervised algorithm, Hierarchical Voting Experts (HVE), for segmenting hierarchically structured sequences. Building on the Voting Experts framework, we extend its application to general domains, demonstrating its effectiveness in unsupervised segmentation tasks. Our findings reveal that entropy metrics similar to those in previous studies serve as valid statistical cues. HVE has shown promise in segmenting various types of data, including Morse code and spoken audio, proving to be surprisingly effective despite its simplicity. Future work includes applying HVE to audio tokenization.
E N D
Hierarchical Voting Experts: An Unsupervised Algorithm for Hierarchical Sequence Segmentation Matt Miller and Alex Stoytchev, Developmental Robotics Lab, Iowa State University [Saffran, Aslin et. al. 1996] Baby’s task: tupirogolabubedakupadoditupirotupirogolabutupiro… tupiro*golabu*bedaku*padoti*tupiro*tupiro*golabu*tupiro… [Cohen and Adams, 2001] Voting Experts’ task: itwasabrightcolddayinaprilandtheclockswere… High Boundary Entropy Low Internal Entropy Our Work i t w a s a b r i g h t … • Extend Voting Experts to more general domains • Use VE for unsupervised segmentation of hierarchically structured sequences • Improve accuracy of segmentation by using • “top down” information Votes: 4 1 6 1 3 1 7 1 i t w a s * a * b r i g h t … itwas*a*bright*cold*day*in*april*andthe*clockswere… Interesting because: Entropy metrics very similar to “Statistical cues” of Saffran, Aslin et. al. General model – useful for more than text Surprisingly effective, given its simplicity Miller and Stoytchev, ICDL 2008
Hierarchical Voting Experts: An Unsupervised Algorithm for Hierarchical Sequence Segmentation Experiments Other codes: • Demonstrate that HVE works • Explore the domain of applicability • Morse code • ASCII Octal • Random Code 3rd Voting Expert: • Improves Accuracy • Top-Down Information Other Experiments: • No Time “Future” Work • Tokenize an audio stream and apply HVE to find breaks • Use artificially generated and spoken audio • Results on “baby talk” and audio CD data are promising Miller and Stoytchev, ICDL 2008