Stemming Algorithms

Xapian uses the Snowball Stemming Algorithms. At present, these support Danish, Dutch, English, Finnish, French, German, Hungarian, Italian, Norwegian, Portuguese, Romanian, Russian, Spanish, Swedish, and Turkish. There are also implementations of Lovins' English stemmer, Porter's original English stemmer, the Kraaij-Pohlmann Dutch stemmer, and a variation of the German stemmer which normalises umlauts.

We'd like to add stemmers for more languages - see the Snowball site for information on how to contribute.

What is a stemming algorithm?

A stemming algorithm is a process of linguistic normalisation, in which the variant forms of a word are reduced to a common form, for example,

    connection
    connections
    connective          --->   connect
    connected
    connecting

It is important to appreciate that we use stemming with the intention of improving the performance of IR systems. It is not an exercise in etymology or grammar. In fact from an etymological or grammatical viewpoint, a stemming algorithm is liable to make many mistakes. In addition, stemming algorithms - at least the ones presented here - are applicable to the written, not the spoken, form of the language.

For some of the world's languages, Chinese for example, the concept of stemming is not applicable, but it is certainly meaningful for the many languages of the Indo-European group. In these languages words tend to be constant at the front, and to vary at the end:

                   -ion
                   -ions
            connect-ive
                   -ed
                   -ing

The variable part is the `ending', or `suffix'. Taking these endings off is called `suffix stripping' or `stemming', and the residual part is called the stem.

Endings

Another way of looking at endings and suffixes is to think of the suffix as being made up of a number of endings. For example, the French word

            confirmatives

can be thought of as `confirm' with a chain of endings,

            -atif (adjectival ending - morphological)
    plus    -e    (feminine ending - grammatical)
    plus    -s    (plural ending - grammatical)

-atif can also be thought of as -ate plus -if. Note that the addition of endings can cause respellings, so -e changes preceding `f' to `v'.

Endings fall into two classes, grammatical and morphological. The addition of -s in English to make a plural is an example of a grammatical ending. The word remains of the same type. There is usually only one dictionary entry for a word with all its various grammatical endings. Morphological endings create new types of word. In English -ise or -ize makes verbs from nouns (`demon', `demonise'), -ly makes adverbs from adjectives (`foolish', `foolishly'), and so on. Usually there are separate dictionary endings for these creations.

Language knowledge

It is much easier to write a stemming algorithm for a language when you are familiar with it. If you are not, you will probably need to work with someone who is, and who can also explain details of grammar to you. Best is a professional teacher or translator. You certainly don't need to have a world authority on the grammar of the language. In fact too much expertise can get in the way when it comes to the very practical matter of writing the stemming algorithm.

Vocabularies

Each stemmer is issued with a vocabulary in data/voc.txt, and its stemmed form in data/voc.st. You can use these for testing and evaluation purposes.

Raw materials

A conventional grammar of a language will list all the grammatical endings, and will often summarise most of the morphological endings. A grammar, plus a dictionary, are therefore basic references in the development of a stemming algorithm, although you can dispense with them if you have an excellent knowledge of the language. What you cannot dispense with is a vocabulary to try the algorithm out on as it is being developed. Assemble about 2 megabytes of text. A mix of sources is best, and literary prose (conventional novels) usually gives an ideal mix of tenses, cases, persons, genders etc. Obviously the texts should be in some sense 'contemporary', but it is an error to exclude anything slightly old. The algorithm itself may well get applied to older texts once it has been written. For English, the works of Shakespeare in the customary modern spelling make a good test vocabulary.

From the source text derive a control vocabulary of words in sorted order. Sample vocabularies in this style are part of our Open Source release. If you make a small change to the stemming algorithm you should have a procedure that presents the change as a three column table: column one is the control vocabulary, column 2 the stemmed equivalent, and column 3 the stemmed equivalent after the change has been made to the algorithm. The effects of the change can be evaluated by looking at the differences between columns two and three.

The first job is to come up with a list of endings. This can be done by referring to the grammar, the dictionary, and also by browsing through the control vocabulary.

Rules for removing endings

If a word has an ending, E, when should E be removed? Various criteria come into play here. One is the knowledge we have about the word from other endings that might have been removed. If a word ends with a grammatical verb ending, and that has been removed, then we have a verb form, and the only further endings to consider are morphological endings that create verbs from other word types. At this level the system of endings gives rise to a small state table, which can be followed in devising the algorithm. In Latin derived languages, there is a state table of morphological endings that roughly looks like this:

   -IC (adj) -+->  -ATION (noun)
              +->  -ITY (noun)
              +->  -MENT (adv)
              \->  -AT (verb)     -+->   -IV (adj)    -+->   -ITY (noun)
                                   |                   \->   -MENT (adv)
                                   \->   -OR (noun)

 -ABLE (adj) -+->  -ITY (noun)
              \->  -MENT (adv)

  -OUS (adj) --->  -MENT (adv)

The ending forms take different values in different languages. In French, -OR becomes `-eur' (m.) or `-rice' (f.), -AT disappears into the infinitive form of a verb. In English, -MENT becomes `-ly', and then one can recognise,

   -IC-ATION   fortification
   -IC-ITY     electricity
   -IC-MENT    fantastically
   -AT-IV      contemplative
   -AT-OR      conspirator
   -IV-ITY     relativity
   -IV-MENT    instinctively
   -ABLE-ITY   incapability
   -ABLE-MENT  charitably
   -OUS-MENT   famously

Trios, -IC-AT-IV etc., also occur, but sequences of length four, -IC-AT-IV-ITY and -IC-AT-IV-MENT, are absent (or occur very rarely).

Sometimes the validity of an ending depends on the immediately preceding group of letters. In Italian, for example, certain pronouns and pronoun groups attach to present participle and infinitive forms of verbs, for example,

scrivendole = scrivendo (writing) + le (to her)
mandarglielo = mandare (to give) + glielo (it to him)

If E is the ending, the possible forms are -andoE, -endoE, -arE, -erE, -irE, so E is removed in the context -Xndo or Yr, where X is a or e, and Y is a or e or i. See the attached_pronoun procedure in the Italian stemmer.

The most useful criterion for removing an ending, however, is to base the decision on the syllable length of the stem that will remain. This idea was first used in the English stemming algorithm, and has been found to be applicable in the other stemming algorithms too. If C stands for a sequence of consonants, and V for a sequence of vowels, any word can be analysed as,

        [C] V C ... V C [V]

where [..] indicates arbitrary presence, and V C ... V C means V C repeated zero or more times. We can find successive positions 0, 1, 2 ... in a word corresponding to each vowel-consonant stretch V C,

        t h u n d e r s t r i c k e n
           0     1         2     3   4

The closer E is to the beginning of the word, the more unwilling we should be remove it. So we might have a rule to remove E if at is after position 2, and so on.

Developing the algorithm

Build the algorithm up bit by bit, trying out a small number of ending removals at a time. For each new ending plus rule added, decide whether, on average, the stemming process is improved or degraded. If it is degraded the rule is unhelpful and can be discarded.

This sounds like common sense, but it is actually very easy to fall into the trap of endlessly elaborating the rules without looking at their true effect. What you find eventually is that you can be improving performance in one area of the vocabulary, while causing a similar degradation of performance in another area. When this happens consistently it is time to call a halt to development and to regard the stemming algorithm as finished.

It is important to realise that the stemming process cannot be made perfect. For example, in French, the simple verb endings -ons and -ent of the present tense occur repeatedly in other contexts. -ons is the plural form of all nouns ending -on, and -ent is also a common noun ending. On balance it is best not to remove these endings. In practice this affects -ent verb endings more than -ons verb endings, since -ent endings are commoner. The result is that verbs stem not to a single form, but to a much smaller number of forms (three), among which the form given by the true stem of the verb is by far the commonest.

If we define errors A and B by,

error A: removing an ending when it is not an ending
error B: not removing an ending when it is an ending

Then removing -ent leads to error A; not removing -ent leads to error B. We must adopt the rule that minimises the number of errors - not the rule that appears to be the most elegant.

Irregular forms

Linguistic irregularities slip through the net of a stemming algorithm. The English stemmer stems `cows' to `cow', but does not stem `oxen' to `ox'. In reality this matters much less than one might suppose. In English, the irregular plurals tend to be of things that were common in Anglo-Saxon England: oxen, sheep, mice, dice - and lice. Men, women and children are of course common today, but the very commonness of these words makes them of less importance in IR systems. Similar remarks may be said about irregular verbs in English, the total number of which is around 150. Here, the fact that verbs are used perhaps rather less than nouns and adjectives in IR queries helps account for the unimportance of verb irregularities in IR performance. There are in English more significant irregularities in morphological changes such as `receive' to `reception', `decide' to `decision' etc., which correspond, ultimately, to irregularities in the Latin verbs from which these words derive. But again working IR systems are rarely upset by lack of resolution of these forms.

An irregularity of English which does cause a problem is the word `news'. It is now a singular noun, and is never regarded as the plural of `new'. This, and a few more howlers, are placed in a table, irregular_forms, in the English stemming algorithm. Similar tables are provided in the other stemming algorithms, with some provisional entries. The non-English stemming algorithms have not been used enough for a significant list of irregular forms to emerge, but as they do, they can be placed in the irregular_forms table.

Using stemming in IR

In earlier implementations of IR systems, the words of a text were usually stemmed as part of the indexing process, and the stemmed forms only held in the main IR index. The words of each incoming query would then be stemmed similarly. When the index terms were seen by the user, for example during query expansion, they would be seen in their stemmed form. It was important therefore that the stemmed form of a word should not be too unfamiliar in appearance. A user will be comfortable with seeing `apprehend', which stands for 'apprehending', `apprehended' as well as `apprehend'. More problematical is `apprehens', standing for `apprehension', `apprehensive' etc., but even so, a trained user would not have a problem with this. In fact all the Xapian stemming algorithms are built on the assumption that it leave stemmed forms which it would not be embarrassing to show to real users, and we suggest that new stemming algorithms are designed with this criterion in mind.

A superior approach is to keep each word, W, and its stemmed form, s(W), as a two-way relation in the IR system. W is held in the index with its own posting list. s(W) could have its separate posting list, but this would be derivable from the class of words that stem to s(W). The important thing is to have the W <-> s(W) relation. From W we can derive s(W), the stemmed form. From a stemmed form s(W) we can derive W plus the other words in the IR system which stem to s(W). Any word can then be searched on either stemmed or unstemmed. If the stemmed form of a word needs to be shown to the user, it can be represented by the commonest among the words which stem to that form.

Stopwords

It has been traditional in setting up IR systems to discard the very commonest words of a language - the stopwords - during indexing. A more modern approach is to index everything, which greatly assists searching for phrases for example. Stopwords can then still be eliminated from the query as an optional style of retrieval. In either case, a list of stopwords for a language is useful.

Getting a list of stopwords can be done by sorting a vocabulary of a text corpus for a language by frequency, and going down the list picking off words to be discarded.

The stopword list connects in various ways with the stemming algorithm:

The stemming algorithm can itself be used to detect and remove stopwords. One would add into the irregular_forms table something like this,

   "", /* null string */

   "am/is/are/be/being/been/"    /* BE */
   "have/has/having/had/"        /* HAD */
   "do/does/doing/did/"          /* DID */
   ...                           /* multi-line string */

so that the words `am', `is' etc. map to the null string (or some other easily recognised value).

Alternatively, stopwords could be removed before the stemming algorithm is applied, or after the stemming algorithm is applied. In this latter case, the words to be removed must themselves have gone through the stemmer, and the number of distinct forms will be greatly reduced as a result. In Italian for example, the four forms

    questa     queste    questi    questo

(meaning `that') all stem to

    quest

In the xapian-data directory in the SVN repository, each language represented in the stemming section has, in addition to a large test vocabulary, a useful stopword list in both source and stemmed form. The source form, in the file stopsource, is carefully annotated, and the derived file, stopwords, contains an equivalent list of sorted, stemmed, stopwords.