We need a way to implement Apertium4 Bison grammars that are robust.

Here, Bison GRAMMAR is a LALR(1) grammar with actions between braces.

There should be a way (or a procedure) to complete handwritten rules, if possible automatically, to generate a robust parser (and translator). The idea is to take the handwritten Bison grammar ${\displaystyle G}$ and complement it with automatically-generated glue rules in such a way that conflicts are not produced (or are harmless) to produce a new grammar ${\displaystyle G'}$.

One possible way to do so is to model left-to-right restarts as follows:

• by accepting the longest possible constituent (in the original grammar ${\displaystyle G}$ that cannot be merged with the remaining output and generating some kind of translation for it
• and treating the remaining output as a complete sentence again

For instance, if the grammar ${\displaystyle G}$ is:

S : NP VP  { write(agree(nom(\$1),\$2)) };
NP : n   { write(\$1) } ;
VP : v  { write(\$1) }
| v NP  { write(acc(\$2),\$1) };

the sequence ${\displaystyle vnn}$ could not be parsed and would lead to an error. However, it could be seen as a NP followed by a VP, and a translation could be generated for both. One could augment the grammar by systematically adding some rules:

S : NP VP  { write(agree(nom(\$1),\$2)) }
| NP TryS { write(\$1); write(\$2) }      # translate an NP and skip
| VP TryS { write(\$1); write(\$2) }      # translate a VP and skip
;
TryS : S  { write(\$1); }
| ;                                  # the remaining part may be empty
NP : n   { write(\$1) } ;
VP : v  { write(\$1) }
| v NP  { write(acc(\$2),\$1) };

However, this may generate conflicts, and there should be a way to avoid them while keeping generality. In fact, many conflicts are caused by ambiguity in the grammar (that is, the sentence is already covered).

One can generate TryS rules for all left-hand-sides incrementally, but maybe order is important, so some ordering of constituents may be useful. Unary productions should be avoided.