stefankreitmayer
commented
5 years ago Temporary solution (for the lab study)
- Keep the 5 lanes (to prevent label overlap)
- Use a simple clustering algorithm, e.g. nearest-neighbour single-linkage clustering, to make 2-4 groups.
- As a proximity measure, use on co-occurrence in the resource. Co-occurrence can be measured either by entities in chunks or by mentions in the text - whichever works best. Entities is probably a safer choice because it's less data, therefore less computation. Being nice to the CPU (on the frontend) is more important in this case than accuracy.
- Put each cluster on adjacent lanes.
- For the horizontal positions: start with the raw position (as currently)
- To finalise the x positions: consider moving individual bubbles left or right to adjust the amount of overlap
- To finalise the y positions: consider swapping clusters vertically to put the biggest cluster near the centre.
Tasks for now
- [x] implement nearest-neighbour clustering in Elm, using TDD
- [x] write a function that takes the old bubble positions and returns the new bubble positions, using the clustering algorithm, with co-occurrence as the distance metric
Next week
- [x] Replace the co-occurrence metric with semantic proximity from the Wikigraph API @sahanbull
This came up in the pilots. Users assume semantic meaning in the positioning of bubbles. Particularly, close-together bubbles are assumed to be closely related.
As a user When I see a bubblogram In order to know which concepts are closely related to each other I want the distance between the bubbles to reflect their semantic relation And it would be nice if subtopics were positioned below overarching terms
Notes