Deference

The direction of deference, with the source to back it up.

A rule-based Japanese linter says nothing about 「お持ちします」. As a string it is fine. But if the reader is the one carrying something, that form does not raise them — it raises whoever the act is directed to, which here is the writer.

Deference decides from whose action it is, who it is directed to, and whether the message goes inside or outside your company, then cites the page of the Council for Cultural Affairs’ Keigo no Shishin (敬語の指針, 2007) the judgement rests on.

Real output, computed ahead of time

This page is a static Space, so the model does not run in your browser. Everything below is genuine output from the published model, recorded in advance. To check your own text, see Run it yourself.

Message


      

        needs context
        surface form
        variation — not reported
      

      

    

    

      
Notes from the guidelines

      

    

  






  
Where the difference is, and where it is not

  

    On error types that cannot be decided without context, textlint and a
    surface-rule set both score 0.0%. On types that are visible in the
    surface pattern, the rule set also reaches 100% — and no advantage is
    claimed there.
  

  

    Detection rate by error type. On the four context-dependent types, textlint and the surface rule set are at zero while Deference reaches 86 to 100 percent. On the four surface-visible types, the rule set and Deference are both at 100 percent.
  

  

  

    Detection rate against over-flagging rate. Deference sits near 98 percent detection with about 1.5 percent over-flagging; the surface rule set is at 43 percent detection; textlint is at 1 percent detection with 9.7 percent over-flagging.
    

      Upper-left is better. textlint detects 1% while over-flagging 9.7% — its
      punctuation and sentence-length rules fire on perfectly correct text.
    

  






  
Variation is not error

  

    Painting every divided usage red destroys trust in a checker. The guidelines
    say as much themselves:
  

  

    男女の違いや世代の違いなどによって画一的に考える態度は避けるべきである。
    (One should avoid treating usage uniformly by gender or
    generation.) — Ch.1 Sec.2-2, p.8
  

  

    53 such cases are held apart and reported as none of them by default:
    doubled honorifics the report calls established by custom
    (お伺いする), forms it calls equally usable
    (ご利用いただく ↔ ご利用くださる), and usages whose
    tolerance it says varies by individual
    (させていただく).
  






  
It does not tell you your Japanese is wrong

  

    Output stays informational — how the guidelines organise the matter, not a
    verdict. Tests enforce this in both languages, checking that phrases like
    “is wrong” or 「誤用」 never reach the user.
    Every run closes with:
  

  

    Honorifics are chosen as a form of self-expression (Keigo no Shishin,
    Ch.1 Sec.1-3). What is shown here is how the guidelines organise the matter;
    it is not a judgement on the writer’s Japanese.
  






  
Run it yourself


pip install git+https://github.com/NagaYu/deference

deference check mail.txt --audience external
deference check mail.txt --audience internal   # the judgement changes


from deference import Deference, MailContext, Audience

result = Deference().check(open("mail.txt").read(), MailContext(audience=Audience.EXTERNAL))
for f in result.reportable: print(f.span.text, f.message, f.citation.render("en"))

  

    The model is 277.5M parameters and runs on CPU in about 26 ms per
    message after int8 quantisation. It works without the trained checkpoint too,
    falling back to the rule engine.