I'm writing to report a bug in the porter-stemmer algorithm supplied as part of the FTS3 implementation.
The stemmer has an inverted logic error that prevents it from properly stemming words of the following form: dry -> dri cry -> cri This means, for instance, that the following words don't stem the same: dried -> dri -doesn't match- dry cried -> cry -doesn't match- cry The bug seems to have been introduced as a simple logic error by whoever wrote the stemmer code. The original description of step 1c is here: http://snowball.tartarus.org/algorithms/english/stemmer.html Step 1c: replace suffix y or Y by i if preceded by a non-vowel which is not the first letter of the word (so cry -> cri, by -> by, say -> say) But the code in sqlite reads like this: /* Step 1c */ if( z[0]=='y' && hasVowel(z+1) ){ z[0] = 'i'; } In other words, sqlite turns the y into an i only if it is preceded by a vowel (say -> sai), while the algorithm intends this to be done if it is _not_ preceded by a vowel. But there are two other problems in that same line of code: (1) hasVowel checks whether a vowel exists anywhere in the string, not just in the next character, which is incorrect, and goes against the step 1c directions above. (amplify would not be properly stemmed to amplifi, for instance) (2) The check for the first letter is not performed (for words like "by", etc) I've fixed both of those errors in the patch below: /* Step 1c */ - if( z[0]=='y' && hasVowel(z+1) ){ + if( z[0]=='y' && isConsonant(z+1) && z[2] ){ z[0] = 'i'; } _______________________________________________ sqlite-users mailing list sqlite-users@sqlite.org http://sqlite.org:8080/cgi-bin/mailman/listinfo/sqlite-users