Citation Context

...feature windows, for a hypothetical occurrence of the word lot . 5 Method As our corpus we used Switchboard, a large corpus of smalltalk between strangers over the telephone recorded in two channels (=-=Godfrey et al., 1992-=-). We collected datapoints from both sides of 20 dialogs, totalling almost two hours, taking a sample every 10 milliseconds. This gave us 600,000 datapoints. For each datapoint we computed 76 prosodic...

Citation Context

...f what prosody contributes to dialog. Going further, one might also use temporal features (Ward et al., 2011), features of gaze, gesture, and words, perhaps in a suitable vector-space representation (=-=Bengio et al., 2003-=-). Better feature weighting could also be useful for refining the ranking of the dimensions: while our method treated one standard deviation of variance in one feature as equally salient as one standa...

Citation Context

...rizing glottalflow waveforms (Pfitzinger, 2008), identifying the key dimensions of variation in pitch contours using Functional Data Analysys (Gubian et al., 2010), and for purely practical purposes (=-=Lee and Narayanan, 2005-=-; Jurafsky et al., 2012). In our own laboratory, Justin McManus applied PCA to 4 left-context, single-speaker prosodic features, and identified the first PC with a continuum from silence to cheerful s...

Citation Context

...hat the dimensions of dialog state expressed by prosody do not aligne with those expressed by words, and perhaps confirm that words can correlate with social and dialog functions in unsuspected ways (=-=Tausczik and Pennebaker, 2010-=-). We next listened to some of some of these datapoints in context. First we listened to a few lowvalued ones and came up with informal hypotheses about what they had in common. We then listened to mo...

Citation Context

...ts relate to these dimensions; for example, to take the set of utterances labeled wh-questions in NXT Switchboard and examine where they are located in the “dialog space” defined by these dimensions (=-=Calhoun et al., 2010-=-; Ward et al., 2012 submitted). Acknowledgments This work was supported in part by NSF Award IIS0914868. We thank Olac Fuentes for suggesting PCA, Justin McManus for the prototype analysis, Shreyas Ka...

Citation Context

...sizes were fixed, not aligned with utterances, words, nor syllables. The specific features we computed were chosen for convenience, based on a basic set previously found useful for language modeling (=-=Ward et al., 2011-=-). These were 1. a speaking-rate measure, over 325 millisecond windows, 2. volume, over 50 ms windows, 3. pitch height, over 150 ms windows, and 4. pitch range, over 225 ms windows. All were speaker-n...

Citation Context

...ucture, question type, recovery from misunderstandings, uncertainty, and so on. Finally, it would be interesting to explore which of these dimensions of state actually matter most for dialog success (=-=Tetreault and Litman, 2006-=-). In addition to the identification of specific dimensions of dialog in casual conversations, this paper contributes a new method: that of using PCA over low-level, observable features to identify im...

Citation Context

... used. Clustering has previously been applied as a way to categorize user intentiontypes and goals, using lexical-semantic features and neighboring-turn features as inputs (Lefevre and de Mori, 2007; =-=Lee et al., 2009-=-), among other methods (Gasic and Young, 2011). Boyer et al. used Hidden Markov Models to identify dialog “modes” that relate to common sequences of dialog-acts (Boyer et al., 2009). There is a also a...

Citation Context

...gely from tradition, rooted in the concerns of precursor fields such as linguistics and artificial intelligence, and refined for elegance and utility (Traum and Larsson, 2003; McGlashan et al., 2010; =-=Bunt, 2011-=-). We feel that these perspectives may be helpfully complemented by bottom-up, empirical investigations of dialog state, as a way to discover new facets of dialog state and as a way to discover which ...

Citation Context

...f the prosodic and other vocal parameters relevant to emotional dimensions (Goudbeek and Scherer, 2010) or levels of vocal effort (Charfuelan and Schröeder, 2011), categorizing glottalflow waveforms (=-=Pfitzinger, 2008-=-), identifying the key dimensions of variation in pitch contours using Functional Data Analysys (Gubian et al., 2010), and for purely practical purposes (Lee and Narayanan, 2005; Jurafsky et al., 2012...

Citation Context

...g and thus speech recognition, as an improvement on dialog-act descriptions of state or descriptions in terms of raw, non-independent prosodic features (Shriberg and Stolcke, 2004; Ward et al., 2011; =-=Stoyanchev and Stent, 2012-=-). Initial results of conditioning on 25 dimensions gave a 26.8% perplexity reduction (Ward and Vega, 2012 submitted). These dimensions could also be used for other purposes, including a more-like-thi...