Cross–lingual mappings of contextual word embedding ELMo

Ljupche Milosheski (2019) Cross–lingual mappings of contextual word embedding ELMo. EngD thesis.

Preview

PDF Download (663Kb)

Abstract

To work with textual data, machine learning algorithms, in particular, neural networks, require word embeddings – vector representations of words in high-dimensional space. There are languages with a small amount of available resources. Exploiting the knowledge from the well-resourced languages for under-resourced languages is possible with cross-lingual embeddings by aligning the embeddings of one language with the vector space of another language. Existing methods for aligning embeddings are intended for context-independent embeddings, where every word has one representation. We propose a method, based on a dictionary and a parallel corpus aligns contextual embeddings, which capture more information about the context in which words appear. The proposed method requires a small amount of bilingual data, which is available for many language pairs. We empirically show that the proposed method outperforms the baseline obtained by alignment of context-independent embeddings.

Actions (login required)