Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Computing Character-level Word Embeddings from a Neural Language Model capturing Lexical Relationships

: Natious, Livin

Fulltext urn:nbn:de:0011-n-5347399 (1.9 MByte PDF)
MD5 Fingerprint: 5b25913f3994e71b6375396bfbef99bf
Created on: 26.2.2019

Bonn, 2019, VII, 60 pp.
Bonn, Univ., Master Thesis, 2019
Master Thesis, Electronic Publication
Fraunhofer IAIS ()

We describe a novel approach to generate high-quality lexical word embeddings from an Enhanced Neural Language Model (NLM) with countertting technique. The generated lexical word embeddings perform better on Natural Language Processing (NLP) tasks that demand synonym/antonym distinction. The quality of generated lexical word embeddings is measured using a downstream NLP task (here, Paraphrase Detection task). We also propose a new deep neural network for Paraphrase Detection task developed as a part of our word embedding evaluation.