Vectors/home

From Nordic Language Processing Laboratory
Jump to: navigation, search

Background

The purpose of the NLPL repository of word vectors (or embeddings, which can comprise both ‘classic’, count-based and ‘modern’, dense models, including the contextualized ones) is to make available a large and carefully curated collection of large-scale distributional semantic models for many languages. For general background, please see Fares et al. (2017).

For interactive exploration and download of the repository, there is an on-line explorer. The underlying data is stored in the NLPL community directory below /cluster/shared/nlpl/data/vectors/ (on Saga) and /projappl/nlpl/data/vectors/ (on Puhti). The repository is versioned, in the sense of assigning release numbers to different stages of repository construction. Each repository entry, thus, is assigned a unique and persistent identifier; once published, a repository entry will never change (to aid replicability). The initial release (providing some two dozen models) was published in May 2017 as version 1.0. In March 2018, version 1.1 supersedes this initial release, adding a large number of models and languages (including those from the 2017 UD parsing shared task) and re-packaging the models from the original release in a more standardized format (see below). In December 2019, version 2.0 was released, which added BERT and ELMo models (including large monolingual models for Norwegian), made metadata more consistent, and ensured that binary format models are always provided (to increase the loading speed, in comparison to the models stored as plain text).

Repository Contents

The on-line browser dynamically presents parts of the information encoded for programmatic access in the repository catalogue, which is represented as a JSON file in the top-level repository directory, with catalogue names corresponding to each repository version, e.g. /cluster/shared/nlpl/data/vectors/20.json (on Saga) for the current repository release.

The catalogue contains three top-level sections, one each for corpora (data sources), algorithms (model creation tools), and models (resulting sets of word vectors). NLPL users with access to Saga and Puhti can read the catalogue file directly from the NLPL community directory, for example when executing a series of experiments that make use of different pre-trained sets of word vectors. Further documentation of the catalogue metadata is available as a separate page.

Each repository entry (i.e. set of word vectors, or ‘model’) is packaged in the form of a .zip archive, with uniform conventions for file naming inside the file, using the model.txt and model.bin entries for the actual vectors. Each archive includes the relevant excerpts from the catalogue as a file meta.json to help identify the specific contents; a README file included with each model entry provides a life-time unique identifier, e.g. http://vectors.nlpl.eu/repository/20/3.zip for model #3 in the 2.0 release of the repository.

Using NLPL Models In-Situ

To avoid data duplication, it is recommended to load models from the NLPL repository directly from the NLPL community directory, when working on Saga or Puhti. Repository entries are uniformly packaged as .zip compressed archives, but the uniform naming scheming makes it possible to directly read one or more of the model files from the archive.

In Python, for example, something along the following lines should work to iterate over all of the entries in the model

import zipfile
import gensim
repository = "/cluster/shared/nlpl/data/vectors/20"
with zipfile.ZipFile(repository + "/30.zip", "r") as archive:
  stream = archive.open("model.txt")
  for line in stream:
    ...

Alternatively, if working in a framework like gensim

  model = gensim.models.KeyedVectors.load_word2vec_format(stream, binary=False, unicode_errors='replace')

Binary fastText models (stored as parameters.bin files) should be first extracted from the .zip archive, and then loaded with

  model = gensim.models.fasttext.load_facebook_vectors("parameters.bin")

Future Work

  1. Prepare Version 2.0, now with /cluster/shared/nlpl/data/vectors/ as the master copy (DONE).
  2. The life-time handle for each model should be included in the JSON catalogue (in addition to being listed in the README file) (DONE).
  3. For classic models, redundantly add binary binary model.bin for faster loading (DONE).
  4. Each corpus should be listed as a separate entry; corpus combinations go into the array-valued corpus property on models (DONE).
  5. Where applicable, there should be an array-valued documentation field (of string, typically URLs) on corpora and models (DONE).
  6. The maintainers property may be over-promising, as often third-party models are in practice unmaintained; maybe rename to creator (DONE).
  7. Document (and possibly re-design) the metadata scheme; maybe invent a fourth category for a process applied prior model training.