Most languages of the world are inflected, meaning that words can change their form to express differences in the following:
-
Number: fox, foxes
-
Tense: pay, paid, paying
-
Gender: waiter, waitress
-
Person: hear, hears
-
Case: I, me, my
-
Aspect: ate, eaten
-
Mood: so be it, were it so
While inflection aids expressivity, it interferes with retrievability, as a single root word sense (or meaning) may be represented by many different sequences of letters. English is a weakly inflected language (you could ignore inflections and still get reasonable search results), but some other languages are highly inflected and need extra work in order to achieve high-quality search results.
Stemming attempts to remove the differences between inflected forms of a
word, in order to reduce each word to its root form. For instance foxes may
be reduced to the root fox, to remove the difference between singular and
plural in the same way that we removed the difference between lowercase and
uppercase.
The root form of a word may not even be a real word. The words jumping and
jumpiness may both be stemmed to jumpi. It doesn’t matter—as long as
the same terms are produced at index time and at search time, search will just
work.
If stemming were easy, there would be only one implementation. Unfortunately, stemming is an inexact science that suffers from two issues: understemming and overstemming.
Understemming is the failure to reduce words with the same meaning to the same
root. For example, jumped and jumps may be reduced to jump, while
jumping may be reduced to jumpi. Understemming reduces retrieval;
relevant documents are not returned.
Overstemming is the failure to keep two words with distinct meanings separate.
For instance, general and generate may both be stemmed to gener.
Overstemming reduces precision: irrelevant documents are returned when they
shouldn’t be.
A lemma is the canonical, or dictionary, form of a set of related words—the
lemma of paying, paid, and pays is pay. Usually the lemma resembles
the words it is related to but sometimes it doesn’t — the lemma of is,
was, am, and being is be.
Lemmatization, like stemming, tries to group related words, but it goes one step further than stemming in that it tries to group words by their word sense, or meaning. The same word may represent two meanings—for example,wake can mean to wake up or a funeral. While lemmatization would try to distinguish these two word senses, stemming would incorrectly conflate them.
Lemmatization is a much more complicated and expensive process that needs to understand the context in which words appear in order to make decisions about what they mean. In practice, stemming appears to be just as effective as lemmatization, but with a much lower cost.
First we will discuss the two classes of stemmers available in Elasticsearch—[algorithmic-stemmers] and [dictionary-stemmers]—and then look at how to choose the right stemmer for your needs in [choosing-a-stemmer]. Finally, we will discuss options for tailoring stemming in [controlling-stemming] and [stemming-in-situ].