vukbatanovic / scstemmers Goto Github PK

(I would like to wrap these stemmers into an Elasticsearch analysis plugin for use on Wikipedia, Wiktionary, and other Serbian, Croatian, Serbo-Croatian, and possibly Bosnian wiki projects. However there are some issues that prevent that from happening.)

I believe that these lines and similar in other stemmers are causing a big problem for non-Serbian text.

For example, तथा is made up of त + थ + ा; the first two characters are letters according to Java, while the last character is not a letter. So, तथा gets stemmed as " ा". Words that are all non-Serbian characters that Java recognizes as letters, like ऊपर, get stemmed as an empty string. This happens to lots of other writing systems besides Devanagari (Hindi), including Greek, Armenian, Hebrew, Arabic, Japanese, Chinese, Korean, and many others (all of which are common in Wikipedia and Wiktionary).

Similarly, words with non-Serbian Latin or Cyrillic letters get stemmed very oddly. ğaʿala, ḍaẓaġ, and ṯaẖaḏ all get stemmed as "aa", mostly because several letters are dropped. Эмин is stemmed as "min" because the initial Э is ignored.

The limits in stemDual1Word for stemmers 1 and 2 are not being applied when the stemmer is run from the command line.

Words like baca, bena, beno, cima, demo, deni, dete, dnem, erem, este, gama, goste, hove, iste, ivan, ivana and others that match suffixes in the rules are being stemmed to nothing by stemmer 1. Stemmer 2 is a little less aggressive.

Similarly, all of these words are stemmed to just "p" by stemmer 1 (stemmer 2 is a bit less aggressive): p, paba, paca, paja, paka, past, pega, pekao, pela, pele, pena, peni, pete, peti, peć, peći, pijem, piju, pila, pile, piliće, pima, pisala, pisali, pisalo, pisan, pisana, pisane, pisani, pisao, pisati, pismo, pite, piti, piće, piše, pišući, poste, pula, puli, pune, puste, puta, pute, puti, puto, puše, п, пава, пади, пака, паст, пата, пајa, паја, пева, пега, пегле, пекао, пела, пеле, пели, пени, пено, пете, пети, пећ, пећи, пила, пиле, пили, пилиће, пило, пирели, писала, писале, писали, писало, писан, писана, писане, писани, писано, писао, писати, писах, писаше, писаће, писаћу, писмо, писте, пите, пити, пише, пишем, пишеш, пишући, пијем, пијемо, пијеш, пију, пиће, пићу, плац, плаца, плима, плиће, пому, поци, поче, птит, пула, пуле, пули, пуне, пусте, пута, путе, пути, путо, пучем.

The suffix transformation in the Ljubešić-Pandžić stemmer is not restricted to only modifying the suffix.

I couldn't find a real world Serbian or Croatian word that had the problem, but the name "Kovačevac" does show up in URLs, for example, without diacritics as "kovacevac", and both instances of "vac" get changed to "vca", giving the final output as "kovcaevc", rather than "kovacevc".

I didn't test it, but looking at the original Python source, it seems to limit substitutions to just suffixes.