Sample visualizations

[mdingemanse]

This is an open issue that exists to drop visualizations of conversational structure and conversational dynamics for easy cross-reference. Most of these made with non-production or alpha versions of our codebase. We use this issue to have a reference point for discussing possible features and design aspects. There is no claim or implication that we want all these or that they are all good (they are not).

1. Eyeballing chat vs. chunk regimes. Fill is mapped to talk_skew ('hotter' = someone's hogging the conversational floor). Top turns, i.e., high frequency turn formats (interjections) are plotted as ciricles. Sometimes such top turns occur in a streak, a series of consecutive tokens; these are shown here as open circles. One kind of observation this affords is that streaks tend to occur in 'chunky' regimes.

2. Comparing ASR systems. Two alternative ways of plotting human-annotated ASR-generated timing data: top lines up turns by speaker (making it easier to spot alignment and disalignment of individual turns), bottom lines up turns by dyad (making it easier to spot where rapid speaker changes lead to missing turns).

3. ASR word-level confidence. Plotting the timing and word-level confidence of ASR output. Height and fill of turns are both mapped to word-level confidence. Visually, turns that look more 'frayed' and lighter in colour are less to be trusted.

4. Eric Carle confidence plot. Tongue-in-cheek example of the same data as above, now with colourful dots marking confidence score at word level (lighter = lower confidence). Enables quick eyeballing of problematic bits.

5. Snailplot. Tongue-in-cheek example of the freedom you get when plotting in different coordinate systems. Code for creating this here.