Text generation

[maxkfranz]

Generation of text from the model may be interesting for:

• indexing by search engines, and
• user-visible summaries (e.g. tweets, explorer view).

Potential technologies:

• EleutherAI
• aitextgen
• Randomised templates

[jvwong]

In doing some mock ups for #937, I was thinking about an Explorer, document-level, redesign and had the below.

Point is

1. there's more than 1 place to draw raw information from
2. the information we use should emphasize our niche (interactions) rather than regurgitating the article, which actually hurts search and discoverability

[maxkfranz]

Re. templating:

A activates B via X in CONTEXT.

A short template summary could be shown in search results:

• Article title
• Figure thumbnail
• Template summary

The interaction text is the most salient thing to include, but it's not that interesting because it's out of context. By including an author's one-line context, the search results could be more compelling.

[maxkfranz]

For later: Most articles contain one or two interactions, so they fit the template well. Nice and short. Outliers contain many interactions. Eventually, it may be nice to be able to include the most salient interaction or two, and that could be in the text summary rather than umpteen interactions.

Alternatively, we could prompt the user to make the context a one-line, all-inclusive summary:

• In that case, it's up to them to mention the most interesting interaction, if any.
• Only the context is the text summary.
• Text generation could be used for other interactions, but it would only be used for search indexing rather than display.

[jvwong]

Related user feedback:

Is “USP1 interacts with FANCD2 via deubiquitination” equivalent to “USP1 deubiquitinates FANCD2”? My first impression was that the former statement would mean USP1 only interacts with deubiquitinated FANCD2.

[jvwong]

See https://github.com/PathwayCommons/factoid/discussions/1195