Natural Language Understanding: Foundations and State-of-the-Art
summary
Survey of three concrete semantics approaches: distributional (Word2Vec / LSA) for broad coverage, frame semantics to expose argument structure, and model‑theoretic executable semantic parsing that maps sentences to programs. Covers neural seq2seq parsers with data‑recombination, the sparse‑reward problem of learning from question→answer only, and a WikiTableQuestions demo (compute steps over tables). Describes SHRDLERN: 100 Mechanical Turk users taught a block world; pragmatic modeling speeds learning for cooperative users. Notes reading‑comprehension models rose from ~51% to ~82% accuracy.
AI-written summary · about
chapters (18)
- 0:00Introduction to speaker
- 0:28Importance of language
- 1:28Alan Turing and AI
- 3:43History of NL systems
- 5:39The statistical turn
- 7:29ML and NLP interaction
- 10:02Properties of language
- 21:03Uncertainty in language
- 25:00Distributional semantics
- 34:12Frame semantics
- 38:20Semantic role labeling
- 47:34Executable semantic parsing
- 58:20Learning with supervision
- 1:07:55Interactive language learning
- 1:12:34Experimental results
- 1:17:47Pragmatics and modeling
- 1:24:19Conclusion and future
- 1:30:43Questions and answers
related talks
49mDeep Learning Applied to Natural Language Processing (NLP)
45mReasoning in Natural Language
01h 00mHuman and machine inductive biases for compositional linguistic generalization
17mNatural Language Processing in Automation
01h 16mGeneralization in natural language processing
30mSemantic Search and Neural Information Retrieval
36mCompositional distributional semantics for modelling natural language
55mDemystifying Natural Language Processing across several industry verticals