Technical Documentation

Den här sidan är inte översatt till svenska. Innehållet visas därför på engelska.

This section describes the internal structure of the Sparv pipeline and its usage and can be used as a developer's guide.

Downloading and installing the Sparv pipeline

The latest distribution of the pipeline can be downloaded from the installation page.

Annotating corpora with the Sparv pipeline can be done in the same manner as on Språkbanken's own servers which is described here (in Swedish only).

The import format is described here.

The pipeline's folder structure

        bin/       binary files
        makefiles/ Makefiles
        models/    linguistic models
        python/    Python modules
        tests/     tests

The pipeline structure

The Sparv corpus pipeline consists of Python modules which are located in the python directory. Most of the modules contain code to solve a single task, e.g. part-of-speech tagging, tokenisation etc. Some of the modules call external tools, e.g. Hunpos or the MaltParser.

The following is a short summary of most of the existing modules and their basic functionalities. More information on every script can be found in the doc strings of the functions.

annotate.pyDifferent types of small annotations.
compound.pyCompound analysis using SALDO.
crf.pySentence segmentation with Conditional Random Fields.
cwb.pyCreates the different output formats (VRT, XML, Corpus Workbench files).
dateformat.pyDate formatting.
diapivot.pyCreates the diachronic pivot, linking older Swedish texts to SALDO.
fileid.pyCreates IDs for every file in the corpus.
freeling.pyFor annotation of other languages. Calls the Freeling software and processes its output. (Only available in the AGPL-version of Sparv)
geo.pyAnnotates geographical features.
hist.pyDifferent functions used for annotation historical Swedish texts.
hunpos_morphtable.pyCreates a morphtable file for Hunpos, based on SALDO.
hunpos.pyPart-of-speech tagging with Hunpos.
info.pyCreates an info file which is used by CWB.
install.pyInstalls corpora on other machines.
lemgram_index.pyCreates a lemgramindex which is used by Korp.
lexical_classes.pyCreates annotations for lexical classes with different resources (Blingbring, SweFN).
lmflexicon.pyParses LMF lexicons.
malt.pySyntactic parsing with the MaltParser.
number.pyDifferent types of numbering used e.g. in structural attributes.
parent.pyAdds annotations for parent links and/or children links (needed for pipeline internal purposes).
readability.pyCreates annotations for readability measures.
relations.pyCreates data used by the word picture.
saldo.pyDifferent types of SALDO-annotations.
segment.pyAll kinds of segmentation/tokenisation (paragraph, sentence, word).
sent_align.pySentence linking of parallel corpora.
sentiment.pyCreates sentiment annotations.
suc2hunpos.pyCreates a test material based on SUC3 for Hunpos.
swener.pyCreates named-entity annotations with hfst-SweNER.
timespan.pyHandles time spans.
trainnstcomp_model.pyTrains a compound part-of-speech model used by the compound analysis.
trainstatsmodel.pyTrains a statistical model used by the compound analysis.
treetagger.pyFor annotation of other languages. Calls the TreeTagger software and processes its output.
vw.pyTopic modelling with vowpal wabbit.
word_align.pyWord linking of parallel corpora with fast_align and cdec.
wsd.pyCrates annotations for word sense disambiguation.
xmlanalyser.pyAn analyzer for pseudo-XML documents.
xmlparser.pyParses XML and generates files in the pipeline's working format.

The first thing that happens when annotating a corpus with the Sparv pipeline is that the XML input data is parsed by the module This module generates the following files for every input file (usually in the annotations directory):

  • a .@TEXT file containing the source text with anchors
  • one file for each structural attribute captured, where each line consists of a reference to two anchors indicating the span of the attribute, and the value of the attribute

The output of most modules are files in the above described format, with references to anchors and data for the indicated span. The part-of-speech tagger module for instance creates a file containing an anchor span per word and the part-of-speech as value.


Below is a list with some of the most common annotations and which module they are built with:

  • word
    The string containing a token (word or punctuation). This annotation is created by
  • msd
    A morphosyntactic descriptor. Created by with word as input data.
  • pos
    A simplified MSD, preserving only its initial part. Created by with msd as input data.
  • baseform, lemgram, sense
    Citation form, lemgram, and SALDO-ID. Created by, with word, msd and ref as input files. Uses a SALDO model which is created with
  • ref
    En numrering av varje token inom varje mening. Skapas av
  • dephead, deprel
    Dependency head and dependency relation. Created by, with word, msd and pos as input data. Uses the MALT parser with a model for Swedish.

Developing a new module

This part describes how to create a new Python module within the Sparv pipeline.

Usually one does not have to define the format for the input and output data since there are methods defined for reading and writing these.

Below is a minimal example script which could be integrated in the Sparv pipeline:

    # -*- coding: utf-8 -*-
    import sb.util as util

    def changecase(word, out, case):
        # Read the annotation "word"
        words = util.read_annotation(word)

        # Transform every word to upper or lower case
        for w in words:
            if case == "upper":
                words[w] = words[w].upper()
            elif case == "lower":
                words[w] = words[w].lower()

        # Write the result (out) to the indicated annotation (words)
        util.write_annotation(out, words)

    if __name__ == '__main__':

Most modules start with importing util which contains functions for reading and writing annotations in the data format used within the pipeline.

Then you can define functions that may be used by Sparv. These functions usually take paths to existing annotation files as arguments as well as a path to the annotation file that should be created by the function.

The function util.read_annotation() is used to read existing annotations. It returns a dictionary object with anchors as keys and annotation values as values.

A module function can be concluded with calling util.write_annotation() which takes a path to an annotation file and a dictionary as arguments and saves the dictionary contents to the specified file path.

A module should also contain some code that makes its functions available from the command line. This is what the last two lines in the above example are there for. As arguments to the function you can list all the names of the functions you want to be able to use from the command line in the following format: alias=name_of_function. The chosen alias can then be specified as a flag in the command line. The first function may be set as default by omitting the alias.


When running the Sparv pipeline its different modules are coordinated with makefiles. There are two global makefiles containing configurations and rules for the creation of annotations (Makefile.config, Makefile.rules). Furthermore, each corpus has its own makefile (Makefile) that contains information about the data format and about which annotations should be used. The two global makefiles are imported by the corpus specific makefile.

Documentation for the different corpus specific makefiles can be found in Makefile.template. This file is distributed together with the other makefiles in the Sparv pipeline.

Makefile.rules is the file containing all rules that create annotations. Makefile.config contains variables that can be configured.

A rule in Makefile.rules can for example look like this:

    %.token.uppercase: %.token.word
        $(python) -m --changecase --out $@ --word $(1) --case upper

The above rule creates the annotation uppercase from the Python example above. It takes the annotation word as input and calls the function uppercase in the script with out, word and case as arguments.

Adding a new annotation can in most cases be done by adding a rule like in the above example to Makefile.rules. The name of the annotation must also be added to the list of annotations i the corpus specific makefile.