Word-Embeddings-Dict

This is an akka-based microservice to act as dictionary for multilanguage word embeddings.

The service run in two modes:

embeddings conversion
dictionary

Embeddings conversion

The first mode is activated by setting the environment variable DO_CONVERSION=true. It will convert embeddings from csv to parquet files. The service will then terminate. All the embeddings must reside in /data/ft/ (a mounted volume) and have filename of the form

wiki.<language>_en.vec

e.g.

wiki.fr_en.vec

except English which will be

wiki.en.vec

The reason for this names is that all vectors for languages other than English are aligned with the English ones.

See scripts/download.sh and scripts/rename.sh to download the raw vectors from Amazon AWS.

Dictionary

This is the main mode of execution. The service will provide

vector dictionary for all supported languages
synonyms (k-nn)
analogies (king - man + woman ~= queen; also k-nn)

Check com.haufe.umantis.ds.embdict.messages for possible message queries.

Build

sbt docker:publishLocal 
docker-compose build 
docker-compose up -d

Deployment

This dictionary service needs access to word vectors in /data/ft through a mounted volume. Ports 5150 and 5151 need to be exposed.

Environment Variables (from docker-compose.yml)

OPENBLAS_NUM_THREADS: 1
LANGUAGES: ${LANGUAGES:-ar,bg,ca,cs,da,de,el,en,es,et,fi,fr,he,hr,hu,id,it,mk,nl,no,pl,pt,ro,ru,sk,sl,sv,tr,uk,vi}
DICTIONARY_SIZE: ${DICTIONARY_SIZE:-all}
DO_CONVERSION: ${DO_CONVERSION:-false}
RESCALE_VECTORS: ${RESCALE_VECTORS:-true}
DICT_TYPES: ${DICT_TYPES:-parallel}

Adding new messages

This services uses akka-kryo-serialization to serialize messages instead of the classic Java serializer for performance reasons. If new messages need to be added, they have to be explicitly listed in the configuration file. Check src/main/resources/kryo_serializer.conf for further information.

Contacts

This software is released under the terms of the GNU GPL 3 License. It was developed by Nicola Bova at Haufe-Umantis, Barcelona, Spain.

Email: nicola dot bova at gmail dot com

Haufe-Lexware / word-embeddings-dict