>wants to reduce complexity
>tries to do so by using an imprecise, context-dependent language designed for person-to-person communication to instruct a machine
???
Selenium is its own can of worms, but it absolutely sounds like you're using the wrong tool for the job here. The problem stopping people from writing browser-based tests is not that people can't understand specific syntaxes or DSLs, it's actually the opposite: people don't have a good, reliable tool to implement browser-based testing in a predictable and specific way that does what a user would intuitively expect.
Selenium fails here because it has to manage interactions between browsers, because selectors are hard to get right on the first try and continually break as the page's format changes, because JavaScript can do literally anything to a page and that is really hard to anticipate and address reliably from a third-party testing framework like Selenium, especially if components are changing the DOM frequently, etc., because Selenium is subject to random breakage at the WebDriver layer that hangs up your (often long-running) script, and so on.
Whatever the right answers to a next-gen Selenium are, attempting to guess the user's meaning based on Real English by something that is itself an imperfect developing technology like NLP is pretty obviously not the correct toolkit to provide that. Remember, a huge amount of the frustration on Selenium comes from not having the utilities needed to specify your intention once and for all -- the ambiguities of plain English will not help.
If your thing works, it will have to end up as a keyword based DSL like SQL. SQL is usually not so scary to newcomers because a simple statement is pretty accessible, not having any weird symbols or confusing boilerplate, but SQL has a rigid structure and it's parsed in conventional, non-ambiguous ways. "BrowserTestQL" (BTQL) would need to be similar, like "FILL FORM my_form WITH SAMPLE VALUES FROM visa_card_network;"
The biggest piece that's missing in Selenium is probably a new, consistent element hashing selector format; each element on the page should have a machine-generated selector assigned under the covers and that selector should never change for as long as the human is likely to consider it the "same element". The human should then use those identifiers to specify the elements targeted. I don't know how to do that.
The second biggest piece that's missing from Selenium is a consistent, stable WebDriver platform that almost never errors out mid-script; this may involve in some type of compile-time checking against the page's structure or something (which I know is hard/possibly impossible because of JS and everything else).