Keynote Q&A – Barbara Plank
summary
Argues that combining deep multi-task learning with oddly available data sources yields practical gains: word embeddings pre-initialization on huge unlabeled corpora outperforms a linguist-heavy Slovene tagger; multiclass losses and hierarchical auxiliary outputs improve low-frequency labels; dynamic frameworks like Dynet (formerly CNN wrappers) ease task mixing; domain similarity via LDA and Jensen–Shannon helps select useful source instances; publishes a multilingual Myers‑Briggs dataset (7 languages, ~400–15,000 users per language) and reports working prototype experiments rather than vague promises.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:51Audience backgrounds
- 3:15Linguists in modern NLP
- 4:39Unsupervised learning in NLP
- 7:11Unlabeled data for robust models
- 8:38Evaluating word embeddings
- 11:28Multitask learning and sharing
- 13:43Rare labels in multitask setups
- 16:18DyNet frameworks
- 17:52Label sparsity in classification
- 19:17Domain adaptation metrics
- 22:52Combining domain metrics
- 24:42Data selection for adaptation
- 25:23Keystroke data improvement
- 26:27Personality from text
- 28:05Ethics in user modeling
- 29:51Further NLP study
related talks
39mTackling the Limits of Deep Learning
43mKeynote: Applied NLP in the age of Generative AI
40mKeynote: Open-source Multimodal AI in the Wild
54mKeynote: Can you trust your (large language) model?
05mKeynote: Teaching LLMs New Knowledge
51mEfficient Deep Learning with Humans in the Loop
44mKeynote: The Art of Language: Mastering Multilingual Challenges in LLMs- 05mKeynote: Teaching LLMs New Knowledge