Nazgul

Nazgul is a neural translation service built on top of Marian (formerly known as AmuNMT). Marcin Junczys-Dowmunt, Tomasz Dwojak, Hieu Hoang (2016). Is Neural Machine Translation Ready for Deployment? A Case Study on 30 Translation Directions (https://arxiv.org/abs/1610.01108). We use the following AmuNMT clone to produce attention info: https://github.com/barvins/amunmt

Nazgul can be used separately, but it pairs up nicely with Sauron: https://github.com/TartuNLP/sauron

Dependencies

NLTK

pip install nltk

translating needs following NLTK modules: perluniprops, punkt, nonbreaking_prefixes

$ python
>>> import nltk
>>> nltk.download()
Downloader> d perluniprops
Downloader> d punkt
Downloader> d nonbreaking_prefixes

How to run

Download and compile AmuNMT according to instructions from https://github.com/barvins/amunmt

To run the nazgul server:

python nazgul_json.py -c config.yml -e truecase.mdl -s 12345

Flags:

-c -- config file to be used for AmuNMT run
-e -- truecasing model file
-s -- socket on which the server will listen (default: 12345)

Configuration file

An example cofiguration (for more info see https://github.com/barvins/amunmt):

# Paths are relative to config file location
relative-paths: yes

# performance settings
beam-size: 5
normalize: yes
cpu-threads: 4
gpu-threads: 0

# scorer configuration
scorers:
  F0:
    path: model.npz
    type: Nematus

# scorer weights
weights:
  F0: 1.0

# vocabularies
source-vocab: vocab.source.bped.json
target-vocab: vocab.target.bped.json

# bpe, disable de-bpe
bpe: codes.bpe
no-debpe: yes

# To produce attention info
return-alignment: yes

Input and output while serving

Input

Desired input depends on with which flags and configuration file options the service is started. When all options are enabled, the input can be just regular text, because service sentence- and word-tokenizes, splits into bpe, truecases and after translation turns the output into proper text.

Output

Output is in json format. Following serves as a sample of output format that is sent to client.

msg = json.dumps({'raw_trans': trans,
                  'raw_input': raw_in,
                  'weights': weights,
                  'final_trans': detokenized.replace('@@ ', '').encode('utf-8').strip()
                  }, encoding='utf-8')
c.send(msg)

Testing

For testing a simple python testing script is in the same folder with server file (test_nazgul_json.py).

kspar/nazgul