Greek stemmer returns an error

[subnix]

We use snowball with manticoresearch and have encountered the error (manticoresoftware/manticoresearch#2888) with the Greek stemmer:

$ echo -n "ισαισα" | ./stemwords -l el
Out of memory

Enabling debug output in runtime/utilities.c produces the following error message:

faulty slice operation:
{|ισα[}'

Tested with master (b72b71f), v2.2.0 (48a67a2), v2.1.0 (4764395), v2.0.0 (c70ed64).

[ojwb]

It's happening in "steps3" - this patch seems a plausible fix, but I've only done minimal testing so far and not checked against the paper to confirm if this is what's actually intended here:

diff --git a/algorithms/greek.sbl b/algorithms/greek.sbl
index df856ac9..06854119 100644
--- a/algorithms/greek.sbl
+++ b/algorithms/greek.sbl
@@ -214,7 +214,7 @@ backwardmode (
       '{y}{s}{a}' '{y}{s}{e}{s}' '{y}{s}{e}' '{y}{s}{a}{m}{e}' '{y}{s}{a}{t}{e}' '{y}{s}{a}{n}' '{y}{s}{a}{n}{e}' (
         delete
         unset test1
-        ('{y}{s}{a}' atlimit <- '{y}{s}') or
+        (['{y}{s}{a}'] atlimit <- '{y}{s}') or
         ([] substring atlimit among (
           '{a}{n}{a}{m}{p}{a}' '{a}{th}{r}{o}' '{e}{m}{p}{a}' '{e}{s}{e}' '{e}{s}{oo}{k}{l}{e}' '{e}{p}{a}' '{x}{a}{n}{a}{p}{a}' '{e}{p}{e}' '{p}{e}{r}{y}{p}{a}'
           '{s}{u}{n}{a}{th}{r}{o}' '{d}{a}{n}{e}' '{k}{l}{e}' '{ch}{a}{r}{t}{o}{p}{a}' '{e}{x}{a}{r}{ch}{a}' '{m}{e}{t}{e}{p}{e}' '{a}{p}{o}{k}{l}{e}'

It looks like this bug has been there since this stemmer was first added.

[ojwb]

A related issue I noticed is that ισα produces an empty stem. Producing an empty stem for a non-empty input is undesirable in general for a stemmer; in this case ισα seems to mean "same" and so probably stemming it to ισ would be good.

The essential cause of this is that the has_min_length test is only applied to the input, but removed suffixes can reduce the stem length below that, even removing the whole word as a suffix (or combination of suffixes). If I test tweaking the code to prevent the -ισα suffix removal then -α is removed instead giving the desired result. Probably this stemmer would benefit from a "minimum length of the remaining stem" check before suffix removal, like most of the other stemmers have. That wouldn't address the error reported here though.