Deep Neural Network Applications in Low-Resource Languages
Event details
| Date | 12.07.2019 |
| Hour | 10:00 › 12:00 |
| Speaker | Mohammadreza Banaei |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Pierre Dillenbourg
Thesis advisor: Prof. Karl Aberer
Co-examiner: Prof. Robert West
Abstract
In recent years, Natural Language Processing (NLP) has experienced a considerable leap in the performance of neural models in different tasks. However, most of these novel models rely on extremely huge unlabeled/labeled corpora, which are not available for many low-resource languages. Our research goal is to design architectures that are able to perform well in low-resource scenarios. As background, we present three papers that are related to our current research objective. The first paper describes a word segmentation algorithm for handling rare words. The second paper proposes an architecture for using phonological features for transfer learning, and the last one introduces an algorithm for generating enriched word embeddings for low-resource scenarios.
Background papers
Neural machine translation of rare words with subword units, by R. Sennrich et al.
Phonologically aware neural model for named entity recognition in low resource transfer settings, by A. Bharadwaj et al.
Adapting Word Embeddings to New Languageswith Morphological and Phonological Subword Representations, by A. Chaudhary et al.
Exam president: Prof. Pierre Dillenbourg
Thesis advisor: Prof. Karl Aberer
Co-examiner: Prof. Robert West
Abstract
In recent years, Natural Language Processing (NLP) has experienced a considerable leap in the performance of neural models in different tasks. However, most of these novel models rely on extremely huge unlabeled/labeled corpora, which are not available for many low-resource languages. Our research goal is to design architectures that are able to perform well in low-resource scenarios. As background, we present three papers that are related to our current research objective. The first paper describes a word segmentation algorithm for handling rare words. The second paper proposes an architecture for using phonological features for transfer learning, and the last one introduces an algorithm for generating enriched word embeddings for low-resource scenarios.
Background papers
Neural machine translation of rare words with subword units, by R. Sennrich et al.
Phonologically aware neural model for named entity recognition in low resource transfer settings, by A. Bharadwaj et al.
Adapting Word Embeddings to New Languageswith Morphological and Phonological Subword Representations, by A. Chaudhary et al.
Practical information
- General public
- Free