fix deinflection bug

[StefanVukovic99]

With #503, more than one inflection rule chain is now shown. Seems the deinflector is producing an incorrect chain for -te form of ichidan verbs:

Not sure if the two subcategories of ichidan verbs in this PR are aptly named, but this seems to fix the problem.

[github-actions]

✔️ No visual differences introduced by this PR.

View Playwright Report (note: open the "playwright-report" artifact)

[djahandarie]

I honestly don't quite understand what's going on in the deinflector logic to determine if this is right (and whether there's a better naming option) so hopefully toasted-nutbread can provide some insight.

[Casheeew]

lgtm

[toasted-nutbread]

I believe the reason this works is because of some shenanigans with the bitflags and the way that the Translator class uses Deinflector.rulesMatch. Specifically, this code:

const definitionRules = Deinflector.rulesToRuleFlags(databaseEntry.rules);
for (const deinflection of uniqueDeinflectionArrays[databaseEntry.index]) {
    const deinflectionRules = deinflection.rules;
    if (!partsOfSpeechFilter || Deinflector.rulesMatch(deinflectionRules, definitionRules)) {
        deinflection.databaseEntries.push(databaseEntry);
    }
}

Deinflector.rulesMatch is only checking to see if any of the bits match between the two rules passed into it, so v1d and v1p will both be matched by terms with the v1 flag.

This is seems to be why these rules are set up this way:

        ['v1',    /** @type {import('translation-internal').DeinflectionRuleFlags} */ (0b000000011)], // Verb ichidan
        ['v1d',   /** @type {import('translation-internal').DeinflectionRuleFlags} */ (0b000000010)], // Verb ichidan dictionary form
        ['v1p',   /** @type {import('translation-internal').DeinflectionRuleFlags} */ (0b000000001)], // Verb ichidan progressive or perfect

This functions as a way to just prevent this from happening:

食べて
  => ('て' => 'てる' via 'masu stem'; rules='v1')
食べてる
  => ('てる' => 'て' via 'progressive or perfect'; rules='iru')
食べて
  => ('て' => 'る' via '-te'; rules='v1')
食べる

With the new rules:

食べて
  => ('て' => 'てる' via 'masu stem'; rules='v1d')
食べてる
  => attempts: ('てる' => 'て' via 'progressive or perfect')
     However, rulesIn='v1p', which doesn't match 'v1d', so this doesn't work
<stop>

[StefanVukovic99]

Exactly.

[toasted-nutbread]

I think the overall idea is that v1d is used as a sort of termination indicator to prevent anything that has been converted from masu stem to have a る ending, that can no longer be used for the progressive or perfect conversion, since that is effectively undoing the addition of the る.

Generally I don't see a huge issue with it being done this way, I can't think of any immediate edge cases where this breaks down. The only thing is that it makes the subtleties of the logic a bit more confusing to understand I suppose, due to the idiosyncrasies of the bitflags test I mentioned above, but that's fine. I think that issue can be addressed by adjusting the structure of the deinflect.json file, which I think would also benefit other language support as well. I'll try looking into that at some point.

[djahandarie]

Thanks for breaking it down. I think I understand, and approve of the change. That said, this introduces even more complexity into the already confusing logic here, so I definitely feel a need for some rework of the deinflector overall, as well as deinflection rule names to be a little less opaque, with more clarity of intent of each part in the naming and structure.