Collaborative Natural Language Inference
summary
Proposes building models that natively support human–model collaboration and counterfactual feedback, combining probabilistic control with pretrained autoregressive nets. Implements PyTorch-Struct to compute marginals, argmax trees, entropy and samples by writing short dynamic programs and using second‑order gradients. Tests include basketball box‑score summarization (baseline outputs ~80% correct but needs high precision) and ListOps where LSTM/Transformer fail but a learned discrete tree with structured RL produces correct parses and generalizes, visualized as red/white correctness dots.
AI-written summary · about
chapters (6)
- 0:00Introduction to NLP Research
- 1:22Interpretable Machine Learning
- 2:54Collaborative Writing Case Study
- 4:28PyTorch Struct Framework
- 6:12Technical Algorithms and Gradients
- 8:20ListOps and Model Evaluation
related talks
17mNatural Language Processing in Automation
01h 30mToward Understanding In-context Learning
01h 10mModels and Programs: Better Together
01h 16mGeneralization in natural language processing
42mReasonable AI
46mWhy data-driven methods will shape the future (...)
01h 31mNatural Language Understanding: Foundations and State-of-the-Art
45mReasoning in Natural Language