Shall we define "decent enough alternate description" or "probably bad alternate text"?

[gautierchomel]

Issue #18 refers to "decent enough alternate description". To define that "decent enough" without relying on complex LLM technologies, we can proceed by eliminating images with probably bad alternate text.

The abelab project analysis referred to "Images to fix" and calculated Missing or bad textual alternative for non decorative graphical resources as follows: content images – content images with alt-text (more than one word and not equal to filename) – contents images decorative

So a first set of rules to define "probably bad alt text" could be :

• alt is not equal to image filename
• alt contains more than one word

The second will certainly trigger false positives as the alt can be meant to articulate with text around as in

<img [...] alt="schema"/>
<p>Representation of the engine to help maintenance operations</p>

[HadrienGardeur]

The second will certainly trigger false positives as the alt can be meant to articulate with text around as in

This will also trigger gaiji, where an image is used to replaced a single character.

[HadrienGardeur]

On a more general note: the manifest command is quite different from epubcheck or ACE because it doesn't return validation messages but an RWPM.

We can inject these extracted image descriptions (using alt or various ARIA roles) into the manifest, but it's not really designed to "flag" bad ones. Additional values in properties could be added with that goal in mind as long as we continue returning a manifest but that's a new concept for this command.