Multi-Lingual Support in Nutch

Jérôme Charron

17 June 2005



The goal of this proposal is to provide a solution for multi-lingual support in Nutch. Multi-lingual support means to be able to use a language specific Analyzer during searching and analysing.


The configuration of this behaviour is done using the standard plugin.includes and plugin.excludes properties of the nutch configuration file. For instance, to add French, English and German analysis support for both document analysis and query parsing, just add in the Nutch configuration file:


If no language specific analyzer plugin is specified in the configuration then the default NutchAnalyzer implementation (ie NutchDocumentAnalyzer will be used for document analysis (similar to the actual implementation), and no language dependent analysis will be performed on the Query.

Nutch Analysis Plugin

NutchAnalyzer Extension Point

The NutchAnalyzer interface defines a new Nutch extension point. This interface is simply an abstract class that extends the Lucene Analyzer class, so that Lucene analyzers could be easily integrated as NutchAnalyzer plugins. NutchAnalyzer extensions should define the attribute "lang" in order to be associated to a particular language code.


The AnalyzerFactory class is responsible of instanciating the NutchAnalyzer implementation to use depending on the Nutch analysis plugins configuration and a specified language code.

The AnalyzerFactory policy for finding the NutchAnalyzer extension to use is to return the first one that match a specified language code. If none is found, then the default NutchDocumentAnalyzer is returned.

Document Analysis

Language Code Source

The language specific document analysis is based on the result of the LanguageIdentifierPlugin. It just call the AnalyzerFactory with the lang attribute provided by the LanguageIdentifierPlugin.

Code modifications

Impacts on the actual code is very light to add multi-lingual capabilities for document analysis. First, on the part of the code of the IndexSegment class that add a document to the index, the line


should be replaced by

indexWriter.addDocument(doc, AnalyzerFactory.get(doc.get("lang")));

so that, the IndexWriter is called with the good Analyzer implementation.

Second, the NutchDocumentAnalyzer class must implement the NutchAnalyzer class.

Query Analysis

Language Code Source

The language specific query analysis is based on a lang attribute like for the Document Analysis. But the lang attribute in this case must be retrieved from the front-end using the following policy:

  1. Use an optional lang attribute provided by the search interface.

  2. If no such attribute is provided by the search interface, then uses the Browser language.
  3. (try to identify the query language using the LanguageIdentifierPlugin)(Note by KurosakaTeruhiko: This probably won't work well because queries are usually too short to tell a language. "chat" can be English or French, for example. What language is "Euro"?)

Code modifications

The query analysis requires more code modifications than the document analysis. The first impact on the searcher code is to be able to get the lang attribute from the front-end. This is done by adding the following method in the Query class:

public static Query parse(String queryString, String lang) throws IOException;

This method then uses the AnalyzerFactory to retrieve the analyzer to use for parsing the query terms. The NutchAnalysis class that is only used as a query parser, will be renamed to QueryParser (like in Lucene), and will be very similar to the Lucene QueryParser by providing a parse method with an Analyzer as parameter:

public static Query parse(String query, Analyzer analyzer);

--- (Note by AlessandroGasparini) In Nutch 0.8.1 this is already implemented in the Query class:

public static Query parse(String queryString, String queryLang, Configuration conf)
throws IOException {
    return fixup(NutchAnalysis.parseQuery(
            queryString, AnalyzerFactory.get(conf).get(queryLang), conf), conf);

More Notes

(Note by AlessandroGasparini) The specified analizers performs language Stemming so if you would you can search for "retrieval" and you can get all the documents in which the stem "retriev" appears. On my configuration i used the LanguageIdentifier, and the analysis-(language) plugins for both the crawler and the web application. I've attached my nutch-site.xml to show the complete plugin configuration nutch-site.xml

The use of stemming needs to be carefully evaluated for a given corpus and target audience. There are a few aspects to be considered:


MultiLingualSupport (last edited 2009-09-20 23:09:42 by localhost)