Build and test with
$> make
from one language to another: cf.
$> python3 induce.py -h
induce.py- propagate monolingual word embeddings to other languages
- using dictionaries or translation tables (weighted average)
- uses FastText embeddings as a basis (hard-wired)
- given source language S with embeddings E_S and target language T without embeddings
- for every t in T: e_s(t) = \sum_{s\in S} P(t|s) e(s)
- optional: given the initial embeddings E_S, a corpus (
-c FILE1[..n]) and an integer n (currently, 50)- for every s in S for which s not in E_S: e_s(s) = avg_c avg_{x in c} e_s(x) with c contexts, x words in context and x in E_S
- truly multilingual embeddings can be achieved by:
- for a set of languages L_1,...,L_n, embedding spaces E_L1, ..., E_Ln
- for any word l in any of these languages, concatenate: e(l) = e_L1(l) x e_L2(l) x ... x e_Ln(l)
- perform dimensionality reduction, keep the weight matrix in order to be able to project future embeddings
- notes
- in order to be balanced, multilingual embeddings that are concatenated must all have the same dimensionality
- avg may be a poor choice for embedding aggregation (drift towards center), maybe better harmonic mean
- as we want to stay interpretable and deterministic, we can actually do statistical instead of neural dimensionality reduction, e.g., using TruncatedSVD
bootstrap contextualized (sense) embeddings
- in translation studies, it is sometimes assumed that one translation corresponds to one sense
- over parallel text, we can concat observed embeddings instead of induced embeddings (and likewise, aggregate over induced embeddings from multiple source languages), so that we arrive at contextualized embeddings (each representing an observed translation set ~ sense)
- using this parallel text as a basis, we can then train Bert (etc.) embeddings and learn their mapping to contextualized sense embeddings => interpretable contextualized embeddings
cf. FastText