microNER: A microservice for German Named Entity Recognition

We publish several pre-trained models for Named Entity Recognition wrapped into a micro-service based on Docker to allow for easy integration of German NER into other applications via a JSON API. With F-Scores above 82% for the GermEval’14 dataset and above 85% for the CoNLL’03 dataset, the micro-service achieves (near) state-of-the-art performance for this task.

The service relies on bidirectional recurrent neural networks and CRF. For details, see this paper.

Installation

Using microNER requires Docker to be installed and running.

Further, it uses pre-trained fastText embeddings which require quite some disk space. 15 GB should be available on your hard disk where Docker images are stored.

Download the image form Docker hub:

docker pull uhhlt/microner:v0.1

And run the image with a port exposed on your host, e.g. 5001:

docker run -p 5001:5001 uhhlt/microner:v0.1

Usage

Once the docker container is running, it accepts JSON requests in the following format:

{
  "meta" : {
    "model" : "MODELNAME"
  },
  "data" : {
    "sentences" : [
        "This is sentence 1.",
        "This is sentence 2."
    ],
    "tokens" : [
		["Token", "sequence", "as", "array"], 
		["Token", "seqeunce", "as", "array"]
	]
  }
}

Four models are available which can be selected via the meta parameter in the JSON request:

1. germeval.h5
2. germeval-inner.h5
3. germeval-conll.h5
4. conll.h5

Text for named entity recognition needs to be given in the data object either pre-tokenized (which is advised since you probably want to keep control on that) passed as an array of an array of strings in the tokens property, or in a array of separate sentences in the sentences property.

The micro-service returns sentence-wise arrays of tokens attached with their respective entity label.

Test

From the command line, you can test a running instance of microNER with curl:

curl --request POST http://localhost:5001/api --header "Content-Type: application/vnd.api+json" --data '
{
  "meta" : {
    "model" : "germeval-conll.h5"
  },
  "data" : {
    "sentences" : [
        "Ich bin Sabine Müllers erster CDU-Generalsekretär.",
        "Tom F. Manteufel wohnt auf der Insel Borkum."
    ],
    "tokens" : [
		["Ich", "bin", "Franz", ",", "du", "bist", "Micha", "."], 
		["Micha", "wohnt", "in", "Berlin", "."]
	]
  }
}
'

Build your own Docker image

Checkout this git repository and run within the project directory

docker build -t my_ner_image .

Citation

If you use the micro-service, consider citing this paper:

• Wiedemann, Gregor; Jindal, Raghav; Biemann, Chris (2018): microNER: A Micro-Service for German Named Entity Recognition based on BiLSTM-CRF. In: Proceedings of the 14th Conference on Natural Language Processing (KONVENS), S. 165–171.

@inproceedings{Wiedemann.2018h,
 author    = {Wiedemann, Gregor and Jindal, Raghav and Biemann, Chris},
 title     = {microNER: A Micro-Service for German Named Entity Recognition 
             based on BiLSTM-CRF},
 pages     = {165--171},
 booktitle = {Proceedings of the 14th Conference on Natural Language 
             Processing (KONVENS)},
 year      = {2018}
}