# Luke Zettlemoyer > ~1979– · Computer Scientist, NLP Researcher > > **Recorded contribution:** Semantic parsing; ELMo co-creator; AI2; University of Washington ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation Luke Zettlemoyer is a University of Washington professor and research director at Meta whose work spans semantic parsing, grounded language, pretrained representations, and large language models. Early work with Michael Collins learned mappings from sentences to logical forms; he later co-authored ELMo, which represented words through contextual states from a bidirectional language model. ## 2. The problem inherited Natural-language systems struggled to connect a sentence to executable meaning and treated each word as one fixed vector even though context changes meaning. ## 3. The central contribution Zettlemoyer advanced learned semantic parsing and co-developed ELMo-style contextual representations, helping move NLP from static features toward pretrained context-sensitive encoders. ## 4. Reconstruct the mechanism 1. Define a structured meaning representation or contextual prediction task. 2. Learn parameters that score candidate structures or token representations from sentences. 3. Use surrounding words to compute a different vector for the same token in each context. 4. Transfer the learned representation into downstream tasks and evaluate where semantics or grounding fails. ## 5. What changed downstream - Semantic parsers made language-to-structure learning a central NLP problem. - ELMo helped demonstrate the broad transfer value of contextual pretraining before BERT and later foundation models. ## 6. Attribution, limits, and uncertainty - ELMo was led by a multi-author Allen Institute/UW team, and contextual language modeling has many predecessors. - Contextual vectors can encode correlations without grounded meaning and inherit bias, privacy, and domain-shift problems from training data. ## 7. Reconstruction lab Write five ambiguous sentences containing one repeated word. Compare a static embedding baseline with context-derived features, then define a logical form for two commands and test compositional generalization. For the embedding task, measure whether the repeated word’s vectors separate by sense and whether neighboring words carry the signal. For semantic parsing, hold out a novel combination of familiar primitives rather than merely new wording. Compare sequence accuracy with exact denotation accuracy. These tests connect contextual representation and structured prediction: both seek representations that transfer, yet fluent contextual features do not by themselves guarantee compositional meaning or executable correctness. Use an adversarial paraphrase and a longer composition to distinguish genuine structural transfer from memorized lexical or template cues. ## 8. Evidence trail - [Luke Zettlemoyer](https://www.cs.washington.edu/people/faculty/luke-zettlemoyer/) — University of Washington - [Deep Contextualized Word Representations](https://aclanthology.org/N18-1202/) — NAACL - [Learning to Map Sentences to Logical Form](https://arxiv.org/abs/1207.1420) — arXiv --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*