Lucene 9.1.0 stempel API
Stempel - Algorithmic Stemmer for Polish Language
A method for conflation of different inflected word forms is an important component of many Information Retrieval systems. It helps to improve the system's recall and can significantly reduce the index size. This is especially true for highly-inflectional languages like those from the Slavic language family (Czech, Slovak, Polish, Russian, Bulgarian, etc).
This page describes a software package consisting of high-quality stemming tables for Polish, and a universal algorithmic stemmer, which operates using these tables. The stemmer code is taken virtually unchanged from the Egothor project.
The software distribution includes stemmer tables prepared using an extensive corpus of Polish language (see details below).
This work is available under Apache-style Open Source license - the stemmer code is covered by Egothor License, the tables and other additions are covered by Apache License 2.0. Both licenses allow to use the code in Open Source as well as commercial (closed source) projects.
A short explanation is in order about the terminology used in this text.
In the following sections I make a distinction between stem and lemma.
Lemma is a base grammatical form (dictionary form, headword) of a word. Lemma is an existing, grammatically correct word in some human language.
Stem on the other hand is just a unique token, not necessarily making any sense in any human language, but which can serve as a unique label instead of lemma for the same set of inflected forms. Quite often stem is referred to as a "root" of the word - which is incorrect and misleading (stems sometimes have very little to do with the linguistic root of a word, i.e. a pattern found in a word which is common to all inflected forms or within a family of languages).
For an IR system stems are usually sufficient, for a morphological analysis system obviously lemmas are a must. In practice, various stemmers produce a mix of stems and lemmas, as is the case with the stemmer described here. Additionally, for some languages, which use suffix-based inflection rules many stemmers based on suffix-stripping will produce a large percentage of stems equivalent to lemmas. This is however not the case for languages with complex, irregular inflection rules (such as Slavic languages) - here simplistic suffix-stripping stemmers produce very poor results.
Lemmatization is a process of finding the base, non-inflected form
of a word. The result of lemmatization is a correct existing word,
often in nominative case for nouns and infinitive form for verbs. A
given inflected form may correspond to several lemmas (e.g. "found"
-> find, found) - the correct choice depends on the context.
Stemming is concerned mostly with finding a unique "root" of a word, which not necessarily results in any existing word or lemma. The quality of stemming is measured by the rate of collisions (overstemming - which causes words with different lemmas to be incorrectly conflated into one "root"), and the rate of superfluous word "roots" (understemming - which assigns several "roots" to words with the same lemma).
Both stemmer and lemmatizer can be implemented in various ways. The two most common approaches are:
- dictionary-based: where the stemmer uses an extensive dictionary of morphological forms in order to find the corresponding stem or lemma
- algorithmic: where the stemmer uses an algorithm, based on
general morphological properties of a given language plus a set of
The implementation described here uses an algorithmic method. This method and particular algorithm implementation are described in detail in . The main advantage of algorithmic stemmers is their ability to process previously unseen word forms with high accuracy. This particular algorithm uses a set of transformation rules (patch commands), which describe how a word with a given pattern should be transformed to its stem. These rules are first learned from a training corpus. They don't cover all possible cases, so there is always some loss of precision/recall (which means that even the words from the training corpus are sometimes incorrectly stemmed).
Algorithm and implementationThe algorithm and its Java implementation is described in detail in the publications cited below. Here's just a short excerpt from :
The tacit input of our method is a sample set (a so-called dictionary) of words (as keys) and their stems. Each record can be equivalently stored as a key and the record of key's transformation to its respective stem. The transformation record is termed a patch command (P-command). It must be ensured that P-commands are universal, and that P-commands can transform any word to its stem. Our solution[6,8] is based on the Levenstein metric , which produces P-command as the minimum cost path in a directed graph.
One can imagine the P-command as an algorithm for an operator (editor) that rewrites a string to another string. The operator can use these instructions (PP-command's): removal - deletes a sequence of characters starting at the current cursor position and moves the cursor to the next character. The length of this sequence is the parameter; insertion - inserts a character ch, without moving the cursor. The character ch is a parameter; substitution - rewrites a character at the current cursor position to the character ch and moves the cursor to the next character. The character ch is a parameter; no operation (NOOP) - skip a sequence of characters starting at the current cursor position. The length of this sequence is the parameter.
The P-commands are applied from the end of a word (right to left). This assumption can reduce the set of P-command's, because the last NOOP, moving the cursor to the end of a string without any changes, need not be stored."
Data structure used to keep the dictionary (words and their P-commands) is a trie. Several optimization steps are applied in turn to reduce and optimize the initial trie, by eliminating useless information and shortening the paths in the trie.
Finally, in order to obtain a stem from the input word, the word is passed once through a matching path in the trie (applying at each node the P-commands stored there). The result is a word stem.
(to be completed...)
The following Polish corpora have been used:
- Polish dictionary from ispell distribution
- Wzbogacony korpus sÅ‚ownika frekwencyjnego
- The Bible (so called "TysiÄ…clecia") - unauthorized electronic version
- Analizator morfologiczny SAM v. 3.4 - this was used to recover lemmas missing from other texts
This step was the most time-consuming - and it would probably be
even more tedious and difficult if not for the
Python. The source texts had to be
brought to a common encoding (UTF-8) - some of them used quite ancient
encodings like Mazovia or DHN - and then scripts were written to
collect all lemmas and
inflected forms from the source texts. In cases when the source text
I used the SAM analyzer to produce lemmas. In cases of ambiguous
lemmatization I decided to put references to inflected forms from all
All grammatical categories were allowed to appear in the corpus,
i.e. nouns, verbs, adjectives, numerals, and pronouns. The resulting
corpus consisted of roughly 87,000+ inflection sets, i.e. each set
consisted of one base form (lemma) and many inflected forms. However,
because of the nature of the training method I restricted these sets to
include only those where there were at least 4 inflected forms. Sets
with 3 or less inflected forms were removed, so that the final corpus
consisted of ~69,000 unique sets, which in turn contained ~1.5 mln
I tested the stemmer tables produced using the implementation described above. The following sections give some details about the testing setup.
The testing procedure was as follows:
- the whole corpus of ~69,000 unique sets was shuffled, so that the input sets were in random order.
- the corpus was split into two parts - one with 30,000 sets (Part 1), the other with ~39,000 sets (Part 2).
- Training samples were drawn in sequential order from the Part 1. Since the sets were already randomized, the training samples were also randomized, but this procedure ensured that each larger training sample contained all smaller samples.
- Part 2 was used for testing. Note: this means that the testing run used only words previously unseen during the training phase. This is the worst scenario, because it means that stemmer must extrapolate the learned rules to unknown cases. This also means that in a real-life case (where the input is a mix between known and unknown words) the F-measure of the stemmer will be even higher than in the table below.
The following table summarizes test results for varying sizes of training samples. The meaning of the table columns is described below:
- training sets: the number of training sets. One set consists of one lemma and at least 4 and up to ~80 inflected forms (including pre- and suffixed forms).
- testing forms: the number of testing forms. Only inflected forms were used in testing.
- stem OK: the number of cases when produced output was a correct (unique) stem. Note: quite often correct stems were also correct lemmas.
- lemma OK: the number of cases when produced output was a correct lemma.
- missing: the number of cases when stemmer was unable to provide any output.
- stem bad: the number of cases when produced output was a stem, but already in use identifying a different set.
- lemma bad: the number of cases when produced output was an incorrect lemma. Note: quite often in such case the output was a correct stem.
- table size: the size in bytes of the stemmer table.
|Training sets||Testing forms||Stem OK||Lemma OK||Missing||Stem Bad||Lemma Bad||Table size [B]|
I also measured the time to produce a stem (which involves
traversing a trie,
retrieving a patch command and applying the patch command to the input
On a machine running Windows XP (Pentium 4, 1.7 GHz, JDK 1.4.2_03
for tables ranging in size from 1,000 to 20,000 cells, the time to
single stem varies between 5-10 microseconds.
This means that the stemmer can process up to 200,000 words per second, an
outstanding result when compared to other stemmers (Morfeusz - ~2,000
w/s, FormAN (MS Word analyzer) - ~1,000 w/s).
The package contains a class
which you can use to produce reports
like the one below:
--------- Stemmer benchmark report: -----------
Stemmer table: /res/tables/stemmer_2000.out
Input file: ../test3.txt
Number of runs: 3
RUN NUMBER: 1 2 3
Total input words 1378176 1378176 1378176
Missed output words 112 112 112
Time elapsed [ms] 6989 6940 6640
Hit rate percent 99.99% 99.99% 99.99%
Miss rate percent 00.01% 00.01% 00.01%
Words per second 197192 198584 207557
Time per word [us] 5.07 5.04 4.82
The results of these tests are very encouraging. It seems that using the training corpus and the stemming algorithm described above results in a high-quality stemmer useful for most applications. Moreover, it can also be used as a better than average lemmatizer.
Both the author of the implementation (Leo Galambos, <leo.galambos AT egothor DOT org>) and the author of this compilation (Andrzej Bialecki <ab AT getopt DOT org>) would appreciate any feedback and suggestions for further improvements.
- Galambos, L.: Multilingual Stemmer in Web Environment, PhD Thesis, Faculty of Mathematics and Physics, Charles University in Prague, in press.
- Galambos, L.: Semi-automatic Stemmer Evaluation. International Intelligent Information Processing and Web Mining Conference, 2004, Zakopane, Poland.
- Galambos, L.: Lemmatizer for Document Information Retrieval
Systems in JAVA.<http://www.informatik.uni-trier.de/%7Eley/db/conf/sofsem/sofsem2001.html#Galambos01>
SOFSEM 2001, Piestany, Slovakia.