User:Firespeaker/HFST bug

In 2011, a bug in how HFST handles words containing spaces was documented and resolved (apparently in r1518?), but it introduced a new bug. This page documents the new behaviour.

text.lexc

Make sure to include the space in '% ' under Multichar_Symbols.

Multichar_Symbols

% 

LEXICON Root

erke:erke # ;
erke% me:erke% me # ;
medvedev:medvedev # ;

Compiling

1. $ hfst-lexc test.lexc -o test.hfst
2. $ hfst-invert test.hfst | hfst-fst2fst -w -o test.hfst.ol

Testing

Some correctly analysed forms

• $ echo "erke" | hfst-proc test.hfst.ol

^erke/erke$

• $ echo "erke me" | hfst-proc test.hfst.ol

^erke me/erke me$

• $ echo "medvedev" | hfst-proc test.hfst.ol

^medvedev/medvedev$

The incorrectly analysed form

• $ echo "erke medvedev" | hfst-proc test.hfst.ol

^erke medvedev/*erke medvedev$

Expected output

This form is analysed correctly by a transducer identical to the one above except with the "erke me" form removed:

• $ echo "erke medvedev" | hfst-proc test2.hfst.ol

^erke/erke$ ^medvedev/medvedev$

Another test case

This one is meant to be more familiar to English-speakers :)

Multichar_Symbols

% 

LEXICON Root

word:word #;
word% form:word% form #;
formation:formation #;

• $ echo "word formation" | hfst-proc test3.hfst.ol

^word formation/*word formation$

• $ echo "formation word" | hfst-proc test3.hfst.ol

^formation/formation$ ^word/word$

Notes

• This doesn't seem to affect transducers written in other formats. E.g., the transducer that results from apertium-eng-kaz.eng.dix outputs the following:
• $ echo "right there" | apertium -d . eng-kaz-morph

^right there/right there<adv>$^./.<sent>$

• $ echo "right the" | apertium -d . eng-kaz-morph

^right/right<adj>/right<adv>/right<n><sg>$ ^the/the<det><def><sp>$^./.<sent>$

Other materials

• spectie explains the bug to firespeaker
• irc.freenode.net#hfst