Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

The great people of University of Turku have developed a new and better version of the parser,

New parser is also available as server and as Docker image.

Because of these reasons, I am happy to say that development of finnish-dep-parser-docker is ended and kazhar/finnish-dep-parser is now deprecated and obsolete.


Docker container and HTTP server for Finnish dependency parser.

Parser takes Finnish text as input and outputs CoNLL-U format.


Finnish dependency parser is originally used as a command line tool, but in this Docker container the parser is wrapped in HTTP server. The server code replaces some of the scripts/code with Java servlets and the result is that it is much faster to parse text.

Fortunately, Finnish dependency parser uses open source code, so I could include and modify some of the scripts and source code. Please see the server-directory to find out what code is used and modified. One such code is HFST optimized-lookup standalone library and command line tool by Helsinki Finite-State Technology.

Unfortunately, the code used in Finnish dependency parser appears not to handle concurrent requests well, so the server is currently forced to process one request at a time. Please see the code for details.

This Docker container includes also Finnish WordNet lexical database and some code to use it. WordNet code is extJWNL (Extended Java WordNet Library) and slightly modified so it works with Finnish WordNet.


Get container from Dockerhub:

  • docker pull kazhar/finnish-dep-parser

Or build docker image using:

  • docker build -t finnish-dep-parser .

Run docker container, expose port 9876 and exit using CTRL-C:

  • docker run -it --rm -p kazhar/finnish-dep-parser

Start docker container in detached mode and restart if container goes down:

  • docker run --restart always -d -p kazhar/finnish-dep-parser

Post file to parser using curl:

  • curl -H "Content-Type: text/plain" --data-binary "@test/text_1k.txt"
  • or
  • cat test/text_1k.txt | curl -H "Content-Type: text/plain" --data-binary @-

Use any programming language to HTTP POST Finnish text to this server and get CoNLL-U format back.

Use HTTP GET with text-param:

Use HTTP GET to get simple statistics of the parser:

Environment variables

Some environment variables can be used.

  • enable_cache, enable in-memory cache, 'true' or 'false'
    • Default is 'true'.
    • Cache size is 100MB.
  • server_feature, set features to start: DEP, LEMMA, FWN or ALL.
    • Separate features using comma to set features. For example: -e "server_feature=DEP,FWN"
    • Default is ALL to start all features.
    • DEP-feature starts dependency parser.
    • FWN-feature enables FinWordNet servlet. Context root is /finwordnet.
      • Requests are like /finwordnet?word=sana.
      • Response is un-compliant custom JSON.
    • LEMMA-feature starts a servlet that returns CoNLL-U format that includes only lemmas. Context root is /lemma.
  • slf4j_log_level, SLF4J logging level.
    • One of: 'trace', 'debug', 'info', 'warn', 'error' or 'off'.
    • Default is 'info'.
  • log_level, system out logging level.
    • 0=no logging of parsed text.
    • 1=log elapsed time and excerpt of parsed text.
    • Default is 0.
    • If log_level=0, you can see the latest parsed text by using GET-request to parser.


Directory parserfarm includes scripts to start finnish-dep-parser farm, 1 or more parser containers within single host accessible via proxy. Farm is implemented using scripts and plain containers without docker-compose or other similar stuff in order to have nothing but Docker runtime as a prereq.

Parsing of text is CPU intensive. If you use many parsers, they may consume all CPU resources. If you use few parsers, they may not take advantage of all resources. Good starting point is to have about as many parser containers as there are CPU cores in the server.

Note that each parser container requires over 2GB of memory. Java heap size is 2GB and in-memory cache size is 100MB. Good starting point is to assume that each parser needs about 2.5GB of memory (and remember that Docker host requires also memory).

Files in parserfarm-directory:

  •, this script starts parser farm. Usage:
    • Default version is 'latest'.
  •, this scripts starts a number of parsers without proxy. Usage:
    • Default version is 'latest'.
  •, this script stops all containers started with either of previous scripts.
  • haproxy_template.cfg, Config file template for haproxy. Modify to your needs. Copied to *haproxy.cfg" when starting the farm.
  • Dockerfile, Dockerfile to build local haproxy image.

If needed, modify the scripts to your requirements.


  • nproc, command to check number of cores in the Linux server.
    • assume 4 cores
  • 2 0.15, starts proxy and 2 parsers of version 0.15.

This kind of farming/scalability is not intended for production use. Use for example IBM Cloud Private for production. Community Edition is available to be used to check out IBM Cloud Private.


Everything in this repo, including all code is "AS IS". No support, no warranty, no fitness for any purpose, nothing is expressed or implied, not by me (nor my employer).


The Finnish dependency parsing pipeline is licensed under GPL-2.0. I guess this repo is also GPL-2.0 but I can not tell for sure, so I don't claim any licensing.

Other licenses apply to other code in this repo. See files in this repo for info about licenses.

FinnWordNet is from University of Helsinki and it is licensed under the Creative Commons Attribution (CC-BY) 3.0 license. As a derivative of the Princeton WordNet, FinnWordNet is also subject to the Princeton WordNet license.

If you want to know more and be sure, please seek legal advice.


Finnish dependency parser ( Docker container.






No packages published