Making Natural Language Processing Robust to Sociolinguistic Variation
summary
Shows that embedding author positions in social networks and adding two bilinear compatibility terms to a neural/structured model yields personalized predictions without per-author labels. Uses linguistic homophily via author node embeddings derived from Twitter graphs, links strings to Wikipedia entities with structured prediction and SVM-style loss trained by SGD, and builds a K-basis ensemble for sentiment. Entity linking on Twitter outperforms S-MART; applying news-trained models to Twitter raises POS error from 3% to 15%, demonstrating practical gains.
AI-written summary · about
chapters (8)
- 0:00Start
- 5:48Finding tacit context in the social network
- 6:36Assortativity of entity references
- 14:08Language variation: a challenge for NLP
- 14:43Personalization by ensemble
- 15:58Homophily to the rescue?
- 17:43Network-driven personalization
- 19:20Variable sentiment words
related talks
31mAttacking Deep Learning-Based NLP Systems
43mA Framework for Transfer Learning on Multi-lingual Text Data for Sentiment Classification
45mDeep Learning for Natural Language Processing
46mWhy data-driven methods will shape the future (...)
11mCollaborative Natural Language Inference
23mTracking the Criminal of Fake News Based on a Unified Embedding
28mEmbeddings
31mKeynote Q&A – Barbara Plank