Citation Context

..., to go from the raw, messily correlated features to a “latent” orthogonal set of components encoding the same information. While PCA has previously been applied to prosodic features, as discussed in =-=[16]-=- and the references therein, it has not before been used in language modeling. 3. The Features, the Data, and Principal Component Analysis Our feature set consisted of 76 prosodic features, taken from...

Citation Context

... finite taxonomy of dialog activities can support most search needs. We therefore chose instead to use an empirically-derived representation of dialog activities. This representation, as described in =-=[14]-=-, is derived by the application of Principal Component Analysis to 76 local prosodic features. While using the common features pitch height, pitch range, speaking rate, and volume, this feature set is...

Citation Context

...king importance vaguely somewhere in the area, but more precisely indicating important and unimportant moments. To explore this we used Principal Components Analysis (PCA), as described in detail in (=-=Ward and Vega, 2012-=-). In short, this method finds patterns of prosodic features which co-occur frequently in the data, and thus provides an unsupervised way to discover the latent structure underlying the observed regul...

Citation Context

...ime maps to a point in this space. This representation is obtained by applying Principal Component Analysis to 78 local prosodic features computed every 10ms calculated over a 6 second sliding window =-=[2]-=-. This feature set was choosen for simplicity of computation and for providing coverage of most of the prosodic aspects known to be most relevent for dialog. It resembles that used in [2], but with mo...

Citation Context

...pical dialog situations where backchannels occur, we need to start with a way to describe dialog situations. While there are many taxonomic systems to choose from, here we use a new, empirical method =-=[7]-=-. Reasoning that the local prosody is a good indicator of dialog activities and states, we started with 76 local prosodic features, consisting of pitch height, pitch range, speaking rate, and volume, ...

Citation Context

...answers to questions that varying in the speaker’s level of certainty. Similar to recent work that has applied principal components analysis (PCA) to large sets of low-level acousticprosodic features =-=[14]-=-, we identify a set of 10 principal components from a large set of word-level prosodic features. We then use the smaller set of prosodic features to learn several decision trees for each speaker and a...

Citation Context

...mation. The specific feature set was chosen for simplicity of computation, for providing coverage of most of the prosodic aspects known to be most relevant for dialog, and for symmetry, as shown in Table 1. Surprisingly, the dimensions resulting from PCA turned out to be meaningful: each of the top couple dozen turned out to align with some known aspects of dialog, such as grounding, turn-taking, seeking and expressing sympathy, degrees of novelty and interest, topic shifts and closings, emphasis, explanations, humor versus regret, personal versus impersonal topics, and facts versus opinions (Ward and Vega, 2012; Ward, 2014). Table 2 shows our tentative interpretations of the top twenty dimensions. We therefore can refer to the space defined by these dimensions as “dialog-activity space.” In a dialog every point in time maps to a point in this 76-dimensional space. 3.2. Similarity and Proximity in this Space Proximity in this vector-space model can serve as a measure of similarity. Initially we used simple Euclidian distance, over all 76 dimensions. As a preliminary exploration, we selected a few seeds and examined what positions the model found as most similar. As 4 1 this speaker talking vs. other ...

Citation Context

... solely on acoustic information has been applied to importance prediction at a very fine granularity (Ward and Richart-Ruiz, 2013). Four basic classes of prosodic features derived from PCA were used (=-=Ward and Vega, 2012-=-): volume, pitch height, pitch range and speaking rate cross various widths of time intervals. The data was labeled by annotators using an importance scale of 1 to 5, and linear regression was used to...