WARNING: The 2.x versions of Elasticsearch have passed their EOL dates. If you are running a 2.x version, we strongly advise you to upgrade.

This documentation is no longer maintained and may be removed. For the latest information, see the current Elasticsearch documentation.

Elasticsearch: The Definitive Guide [2.x] » Dealing with Human Language » Reducing Words to Their Root Form » Stemming in situ
« Controlling Stemming Stopwords: Performance Versus Precision »

Stemming in situedit

For the sake of completeness, we will finish this chapter by explaining how to index stemmed words into the same field as unstemmed words. As an example, analyzing the sentence The quick foxes jumped would produce the following terms:

Pos 1: (the)
Pos 2: (quick)
Pos 3: (foxes,fox) 
Pos 4: (jumped,jump)

The stemmed and unstemmed forms occupy the same position.

Read Is Stemming in situ a Good Idea before using this approach.

To achieve stemming in situ, we will use the keyword_repeat token filter, which, like the keyword_marker token filter (see Preventing Stemming), marks each term as a keyword to prevent the subsequent stemmer from touching it. However, it also repeats the term in the same position, and this repeated term is stemmed.

Using the keyword_repeat token filter alone would result in the following:

Pos 1: (the,the) 
Pos 2: (quick,quick) 
Pos 3: (foxes,fox)
Pos 4: (jumped,jump)

The stemmed and unstemmed forms are the same, and so are repeated needlessly.

To prevent the useless repetition of terms that are the same in their stemmed and unstemmed forms, we add the unique token filter into the mix:

PUT /my_index
{
  "settings": {
    "analysis": {
      "filter": {
        "unique_stem": {
          "type": "unique",
          "only_on_same_position": true 
        }
      },
      "analyzer": {
        "in_situ": {
          "tokenizer": "standard",
          "filter": [
            "lowercase",
            "keyword_repeat", 
            "porter_stem",
            "unique_stem" 
          ]
        }
      }
    }
  }
}

The unique token filter is set to remove duplicate tokens only when they occur in the same position.

The keyword_repeat token filter must appear before the stemmer.

The unique_stem filter removes duplicate terms after the stemmer has done its work.

Is Stemming in situ a Good Ideaedit

People like the idea of stemming in situ: “Why use an unstemmed field and a stemmed field if I can just use one combined field?” But is it a good idea? The answer is almost always no. There are two problems.

The first is the inability to separate exact matches from inexact matches. In this chapter, we have seen that words with different meanings are often conflated to the same stem word: organs and organization both stem to organ.

In Using Language Analyzers, we demonstrated how to combine a query on a stemmed field (to increase recall) with a query on an unstemmed field (to improve relevance). When the stemmed and unstemmed fields are separate, the contribution of each field can be tuned by boosting one field over another (see Prioritizing Clauses). If, instead, the stemmed and unstemmed forms appear in the same field, there is no way to tune your search results.

The second issue has to do with how the relevance score is calculated. In What Is Relevance?, we explained that part of the calculation depends on the inverse document frequency — how often a word appears in all the documents in our index. Using in situ stemming for a document that contains the text jump jumped jumps would result in these terms:

Pos 1: (jump)
Pos 2: (jumped,jump)
Pos 3: (jumps,jump)

While jumped and jumps appear once each and so would have the correct IDF, jump appears three times, greatly reducing its value as a search term in comparison with the unstemmed forms.

For these reasons, we recommend against using stemming in situ.

« Controlling Stemming Stopwords: Performance Versus Precision »