Tags on individual unchanged LU's are spread across the whole matching rule about apertium-separable

[unhammer]

Here's a minimal lsx rule that recreates the situation (the actual rules have pardefs to capture full adverbials of different kinds, I just expanded them to exactly what matches):

    <e>
      <i><w/><s n="pr"/><t/><d/></i>
      <i><w/><s n="adj"/><t/><d/></i>
      <i><w/><s n="n"/><t/><d/></i>
      <p><l>foreligge</l>          <r>ligge</r></p>  <i><t/><d/></i>
      <i><w/><s n="adv"/><t/><d/></i>
      <i><w/><s n="adj"/><t/><d/></i>
      <i><w/><s n="n"/><t/><d/></i>
      <p><l></l>                   <r>fore<s n="adv"/><d/></r></p>
    </e>

The part before the verb is there to create a longer match, since one might want different idioms in different contexts. But you'd get the same behaviour with a simpler rule and marking a different word:

    <e>
      <p><l>foreligge</l>          <r>ligge</r></p>  <i><t/><d/></i>
      <i><w/><s n="adv"/><t/><d/></i>
      <i><w/><s n="adj"/><t/><d/></i>
      <i><w/><s n="n"/><t/><d/></i>
      <p><l></l>                   <r>fore<s n="adv"/><d/></r></p>
    </e>

echo 'Hos personer med atopisk eksem foreligger ofte <a class="crossref" href="https://sml.snl.no/arv">arvelige</a> faktorer' | apertium -u -f html-noent -d . nob-nno_e
Hos personar <a class="crossref" href="https://sml.snl.no/arv">med atopisk eksem ligg ofte arvelege faktorar føre</a>

This is exactly how lsx is meant to be used, so if we want it not to spread tags that way I guess we have to do something to lsx. @mr-martian any ideas?

[mr-martian]

It would be possible (annoying with pardefs but possible) to detect which LUs are unchanged and mark those as such somehow in the FST, and then any with such a mark don't get wblank spreading. The only question then is how to represent this in the FST itself.

[unhammer]

Would it be easier to detect it in the compiled FST and do some fst-postprocessing? Ie find paths where it's all identities before the <d/>

Marking while pardefs as identity-only wouldn't help if there's a pardef that's got one <e><i>a</i></e> and one <e><p><l>a</l><r>b</p></e> – only when matching the first one do we want to avoid tag-spreading.

(Alternatively, is this something that could be more easily checked at runtime?)

[mr-martian]

Runtime might not actually be a bad idea. We could check if any of the output words are identical to any of the input words, and if they are, then they only get the wblank of that input word. This only becomes a problem if it turns out to be noticeably slower or if the input contains multiple identical words. Arguably there's an issue if the rule deletes a word and creates another one that happens to be identical, but I think we can just count that as moving it.