Tagger Handler

The Tagger request handler, AKA the "SolrTextTagger", is a "text tagger".

Given a dictionary (a Solr index) with a name-like field, you can post text to this request handler and it will return every occurrence of one of those names with offsets and other document metadata desired. It’s used for named entity recognition (NER).

The tagger doesn’t do any natural language processing (NLP) (outside of Lucene text analysis) so it’s considered a "naive tagger", but it’s definitely useful as-is and a more complete NER or ERD (entity recognition and disambiguation) system can be built with this as a key component. The SolrTextTagger might be used on queries for query-understanding or large documents as well.

To get a sense of how to use it, jump to the tutorial below.

The Tagger request handler does not yet support a sharded index. It can be used in a cluster running in SolrCloud mode, but the collection that stores the tag dictionary must be a single-sharded collection. Despite this limitation, tens to hundreds of millions of names (documents) can be supported; the maximum is mostly limited only by memory.

Tagger Configuration

To configure the tagger, your Solr schema needs 2 fields:

  • A unique key field (see Unique Key for how to define a unique key in your schema). Recommended field settings: set docValues=true.

  • A tag field, which must be a TextField, with ConcatenateGraphFilterFactory at the end of the index chain (not the query chain): Set preservePositionIncrements=false on that filter. Recommended field settings: omitNorms=true, omitTermFreqAndPositions=true and maybe specify the postings format — see performance tips.

The text field’s index analysis chain, aside from needing ConcatenateGraphFilterFactory at the end, can otherwise have whatever tokenizer and filters suit your matching preferences. It can have multi-word synonyms and use WordDelimiterGraphFilterFactory for example. However, do not use FlattenGraphFilterFactory as it will interfere with ConcatenateGraphFilterFactory. Position gaps (e.g., stop words) get ignored; it’s not (yet) supported for the gap to be significant.

The text field’s query analysis chain, on the other hand, is more limited. There should not be tokens at the same position, thus no synonym expansion — do that at index time instead. Stop words (or any other filter introducing a position gap) are supported. At runtime the tagger can be configured to either treat it as a tag break or to ignore it.

Your solrconfig.xml needs the solr.TagRequestHandler defined, which supports defaults, invariants, and appends sections just like the search handler.

For configuration examples, jump to the tutorial below.

Tagger Parameters

The tagger’s execution is completely configurable with request parameters. Only field is required.



Default: none

The tag field that serves as the dictionary. This is required; you’ll probably specify it in the request handler.



Default: none

You can specify some number of filter queries to limit the dictionary used for tagging. This parameter is the same one used by the solr.SearchHandler.



Default: 10000

The maximum number of documents to return. This parameter is the same as is used by the solr.SearchHandler.



Default: none

Solr’s standard parameter for listing the fields to return. This parameter is the same one used by the solr.SearchHandler.



Default: NO_SUB

Choose the algorithm to determine which tags in an overlapping set should be retained versus being pruned away. Options are:

  • ALL: Emit all tags.

  • NO_SUB: Don’t emit a tag that is completely within another tag (i.e., no subtag).

  • LONGEST_DOMINANT_RIGHT: Given a cluster of overlapping tags, emit the longest one (by character length). If there is a tie, pick the right-most. Remove any tags overlapping with this tag then repeat the algorithm to potentially find other tags that can be emitted in the cluster.



Default: false

If true, return the matched text in the tag response. This will trigger the tagger to fully buffer the input before tagging.



Default: 1000

The maximum number of tags to return in the response. Tagging effectively stops after this point.



Default: false

If true, suppress errors that can occur if, for example, you enable synonym expansion at query time in the analyzer, which you normally shouldn’t do. Let this default to false unless you know that such tokens can’t be avoided.



Default: false

A boolean flag that causes stopwords (or any condition causing positions to skip like >255 char words) to be ignored as if they aren’t there. Otherwise, the behavior is to treat them as breaks in tagging on the presumption your indexed text-analysis configuration doesn’t have a StopWordFilter defined. By default the indexed analysis chain is checked for the presence of a StopWordFilter and if found then ignoreStopWords is true if unspecified. You probably shouldn’t have a StopWordFilter configured and probably won’t need to set this parameter either.



Default: false

If true, indicates the input is XML and the offsets of returned tags should be adjusted as necessary to allow for the client to insert an opening and closing element at the tag offset pair. If it isn’t possible to do so then the tag will be omitted. You are expected to configure HTMLStripCharFilterFactory in the schema when using this option. This will trigger the tagger to fully buffer the input before tagging.

Solr’s parameters for controlling the response format are also supported, such as echoParams, wt, indent, etc.

Tutorial with Geonames

This is a tutorial that demonstrates how to configure and use the text tagger with the popular Geonames data set. It’s more than a tutorial; it’s a how-to with information that wasn’t described above.

Create and Configure a Solr Collection

Create a Solr collection named "geonames". For the tutorial, we’ll assume the default "data-driven" configuration. It’s good for experimentation and getting going fast but not for production or being optimal.

bin/solr create -c geonames

Configuring the Tagger

We need to configure the schema first. The "data driven" mode we’re using allows us to keep this step fairly minimal — we just need to declare a field type, 2 fields, and a copy-field.

The critical part up-front is to define the "tag" field type. There are many many ways to configure text analysis; and we’re not going to get into those choices here. But an important bit is the ConcatenateGraphFilterFactory at the end of the index analyzer chain. Another important bit for performance is postingsFormat=FST50 resulting in a compact FST based in-memory data structure that is especially beneficial for the text tagger.

Schema configuration:

curl -X POST -H 'Content-type:application/json'  http://localhost:8983/solr/geonames/schema -d '{
         "class":"solr.StandardTokenizerFactory" },
        {"class":"solr.ConcatenateGraphFilterFactory", "preservePositionIncrements":false }
         "class":"solr.StandardTokenizerFactory" },

  "add-field":{"name":"name", "type":"text_general"},

  "add-field":{"name":"name_tag", "type":"tag", "stored":false },

  "add-copy-field":{"source":"name", "dest":["name_tag"]}

Configure a custom Solr Request Handler:

curl -X POST -H 'Content-type:application/json' http://localhost:8983/solr/geonames/config -d '{
  "add-requesthandler" : {
    "name": "/tag",

Load Some Sample Data

We’ll go with some Geonames.org data in CSV format. Solr is quite flexible in loading data in a variety of formats. This cities1000.zip should be almost 7MB file expanding to a cities1000.txt file around 22.2MB containing 145k lines, each a city in the world of at least 1000 population.

Using bin/post:

bin/post -c geonames -type text/csv \
  -params 'optimize=true&maxSegments=1&separator=%09&encapsulator=%00&fieldnames=id,name,,alternative_names,latitude,longitude,,,countrycode,,,,,,population,elevation,,timezone,lastupdate' \

or using curl:

curl -X POST --data-binary @/path/to/cities1000.txt -H 'Content-type:application/csv' \

That might take around 35 seconds; it depends. It can be a lot faster if the schema were tuned to only have what we truly need (no text search if not needed).

In that command we said optimize=true&maxSegments=1 to put the index in a state that will make tagging faster. The encapsulator=%00 is a bit of a hack to disable the default double-quote.

Tag Time!

This is a trivial example tagging a small piece of text. For more options, see the earlier documentation.

curl -X POST \
  'http://localhost:8983/solr/geonames/tag?overlaps=NO_SUB&tagsLimit=5000&fl=id,name,countrycode&wt=json&indent=on' \
  -H 'Content-Type:text/plain' -d 'Hello New York City'

The response should be this (the QTime may vary):

        "name":["New York City"],

Tagger Performance Tips

  • Follow the recommended configuration field settings above. Additionally, for the best tagger performance, set postingsFormat=FST50. However, non-default postings formats have no backwards-compatibility guarantees, and so if you upgrade Solr then you may find a nasty exception on startup as it fails to read the older index. If the input text to be tagged is small (e.g., you are tagging queries or tweets) then the postings format choice isn’t as important.

  • "optimize" after loading your dictionary down to 1 Lucene segment, or at least to as few as possible.

  • For bulk tagging lots of documents, there are some strategies, not mutually exclusive:

    • Batch them. The tagger doesn’t directly support batching but as a hack you can send a bunch of documents concatenated with a nonsense word that is not in the dictionary like "ZZYYXXAABBCC" between them. You’ll need to keep track of the character offsets of these so you can subtract them from the results.

    • For reducing tagging latency even further, consider embedding Solr with EmbeddedSolrServer. See EmbeddedSolrNoSerializeTest.

    • Use more than one thread — perhaps as many as there are CPU cores available to Solr.