We release pre-trained word embeddings:
- 200-dimensional GloVe vectors (.text);
- 300-dimensional CBOW (.text) and SkipGram (.text) vectors;
- 200-dimensional fastText vectors (.text, .bin), trained using SkipGram architecture, with char n-grams up to length 3.
The training data for these models was collected from various sources:
a. Wikipedia;
b. fiction texts taken from the open part of the EANC corpus;
c. HC Corpora containing blogs and news articles collected by Hans Christensen from public sources in 2011;
d. digitized and reviewed part of Armenian soviet encyclopedia (as of February 2018) taken from Wikisource;
e. texts from news websites on the following topics: economics, events, art, sports, law, politics, blogs and interviews.
The texts were preprocessed by lowercasing all tokens and removing punctuation, digits. The final dataset contained 90.5 million tokens.
In addition, we publish an adaptation of the word analogy task (Mikolov et al., 2013a) for the Armenian language to serve as benchmark for intrinsic evaluation of vectors. The task contains 5 semantic and 8 syntactic sections, with 15646 analogy questions in total.
For extrinsic evalution of vectors in a classification task, we release a dataset of over 12000 news articles from iLur.am, categorized into 7 classes: sport, politics, weather, economy, accidents, art, society. The articles are split into train (2242k tokens) and test sets (425k tokens).
For more details, refer to the paper.