Home › Categories › Technology› Download

Download.pdf

towards-realistic-single-task-continuous-learning-research-for-ner.pdf

🔗 Source: assets.amazon.science

📊 Size: 303 KB

📄 Pages: 11 pages

⬇️ Downloads: 48

Summary

Researchers discuss the need for realistic single-task continuous learning benchmarks in natural language processing, particularly for named entity recognition (NER) tasks. They identify the lack of public datasets that mimic real-world scenarios, where data is dynamic and evolves over time. To address this, they create a continuous learning NER dataset from an existing publicly available dataset, StackOverflowNER, and release it along with the code. The dataset is split into time-based episodes, and an NER model is trained incrementally over five episodes. The results show limited forgetting and no regression. The researchers also propose a configurable distribution-based sampling of data and study the effectiveness of different single-task continuous learning strategies, including data rehearsal and a variation of the replay-based GDumb baseline.

Download File

Description

Researchers discuss the need for realistic single-task continuous learning benchmarks in natural language processing, particularly for named entity recognition...

Technical Information

File Format: PDF
File Size: 303 KB
Pages: 11
Language: EN
Total Downloads: 48
Last Updated: 1 month ago

Document Overview

This PDF document about Download provides comprehensive information and guidance. Whether you're a beginner or advanced user, this resource offers valuable insights into Download.