[Paper Review] Low-resource Languages: A Review of Past Work and Future Challenges

Alexandre Magueresse, Vincent Carles+1 more|arXiv (Cornell University)|Jun 12, 2020

Natural Language Processing Techniques56 references36 citations

TL;DR

A survey of NLP approaches for low-resource languages, focusing on projection/alignment, dataset creation, linguistic tasks, speech, embeddings, MT, and future directions.

ABSTRACT

A current problem in NLP is massaging and processing low-resource languages which lack useful training attributes such as supervised data, number of native speakers or experts, etc. This review paper concisely summarizes previous groundbreaking achievements made towards resolving this problem, and analyzes potential improvements in the context of the overall future research direction.

Motivation & Objective

• Summarize historical and current methods addressing NLP for low-resource languages (LRLs).
• Analyze resource collection and automatic alignment techniques across tasks.
• Outline linguistic tasks, speech, embeddings, MT, and classification in LRL contexts.
• Highlight future research directions and the need for diversified datasets and language closeness metrics.

Proposed method

• Describe the projection (alignment) technique at document, sentence, and word levels and its role in transferring HRL annotations to LRLs.
• Review dataset creation efforts (REFLEX-LCTL, LORELEI, etc.) and automatic alignment methods across lexicon induction, sentence, and document levels.
• Discuss linguistic tasks (POS tagging, dependency parsing, NER/typing/linking, morphology induction) and how projection/embeddings are applied.
• Summarize speech recognition approaches (MLP transfer, HMMs) and their transfer learning considerations.
• Summarize embeddings and MT approaches (data augmentation, multilingual embeddings, transfer learning, language graphs).
• Identify cross-cutting challenges and propose directions like diversified datasets and language closeness metrics.

Experimental results

Research questions

• RQ1 What past approaches have effectively addressed data scarcity in LRLs through projection/alignment?
• RQ2 What are the main data collection and alignment strategies used for LRLs, and what tasks do they support?
• RQ3 How have linguistic tasks, embeddings, and MT been adapted for LRLs, and what are the core limitations?
• RQ4 What future directions and open questions does the literature identify for improving LRL NLP?

Key findings

• Projection/alignment remains central to leveraging HRL annotations for LRLs across document, sentence, and word levels.
• Numerous dataset efforts (REFLEX-LCTL, LORELEI) and automatic alignment techniques underpin progress, with document/sentence-level applications emphasized.
• Linguistic tasks such as POS tagging, dependency parsing, NER/NET/NEL, and morphology induction benefit from cross-lingual projections and multilingual embeddings.
• Speech recognition for LRLs relies on transfer learning and architecture adaptations (MLP/HMM), with ongoing challenges in data heterogeneity and entity detection.
• Multilingual embeddings and MT approaches (transfer learning, zero-shot translation, language graphs) show promise but depend on language similarity and data availability.
• A critical need identified is diverse, multilingual datasets and a robust closeness metric to evaluate cross-language generalization.