Christoph,

I like your ideas, especially scrolling the text and leaving the controls in place. Then when the scrolling stops, I update the controls to show the values for the lines in view. That way I can get rid of 90+% of my controls!

As for ease of use, I have been rapid prototyping and trying to find easy ways to allow the user to make these edits. To answer one question put to me above, the user must see one or two dozen lines of context to make their decisions for a single line. The decisions are join to previous line, split from previous, delete line, change text of line, identify line as a "See also", a glossary term, an index term, and identify the level of indentation (1, 2, 3 or 4). Basically I am taking semi-structured text (a book index) and adding semantic information. My parser gets most lines right, but since I am converting PDFs into text, weird things happen, the indentation gets screwed up, Unicode characters intrude, headers and footers must be discarded, etc. Thus the user must edit the semantics.

I could just post my results in an XML text editor and let them edit it that way, but I found that to be very hard for a user to do. Reading and editing XML is painful.

Also, directly editing the text won't work because I must store a list of change rules, not the final result. This is because I have two streams of data coming in that must be kept in synch: the text in a single column without indents but in reading order and the text in multiple columns (from which I get the indentation information that drives the computation of level.) The user edits the single column data, and I apply the edits also to the multi-columned data.

- Paul