Style Transfer for Data Augmentation in Sequence Labelling Tasks - September 2022
Center for Language & Speech Processing(CLSP), JHU via YouTube
Overview
Explore data augmentation techniques using style transfer for improving domain adaptation in sequence labeling tasks, specifically named entity recognition (NER) on social media data. Learn about transforming out-of-domain labeled data to stylistically match target data, enhancing NER prediction performance through combined training on generated and in-domain data. Discover recent empirical results and gain insights into research projects addressing the challenge of scarce labeled data in text understanding and language analysis. Delve into topics such as noise in reconstruction, denoising techniques, data evaluation, comparison methods, alignment strategies, supervision approaches, and paraphrase generation. Examine model overviews, result comparisons, and cherry-picked examples to understand the effectiveness of these techniques in improving NER performance on social media platforms.
Syllabus
Introduction
Motivation
Noise in Reconstruction
Denoising Reconstruction
Reconstruction Details
Data
Evaluation
Comparison
Alignment
Supervision
Linearization
Paraphrase Generation
Model Overview
Results
Comparing Results
Cherry Pick Examples
Conclusion
Questions
Linearisation
Taught by
Center for Language & Speech Processing(CLSP), JHU