Some putative hallmarks of human language Creativity of linguistic meaning! 50k words in a language Open class words Concrete: dog, woman, house Abstract: belief, believe, jealousy Closed class words the, of, from, without, Infinitely many combinatory meanings, including impossible ones: I met a green cyclops in the garden today. John thought Jane knew that I met a green cyclops in the garden today. 1
Human linguistic creativity and semantic and pragmatic operations. Semantic combinations can be fickle The blue sky, bruise, eyes, ribbon, water of the Caribbean. And metaphorical The meeting was a zoo. And shaped by pragmatics Do you have the time? And can you pass the salt? It s cold in here. Human linguistic creativity and grammatical structure 2
Grammatical knowledge is represented independently of meaning. Colorless green ideas sleep furiously. Furiously sleep ideas green colorless. Human linguistic creativity and grammatical structure Grammar changes meaning The dog bit the man. The man bit the dog. The man was bitten by the dog. The dog was bitten by the man. 3
And grammar allows us to combine infinitely Bob thinks I told Jane to order Sarah to fire the new guy. S! NP VP NP! Det N S VP! V NP Recursion the man that the cops saw was a drug dealer 8 4
Influential idea in the cognitive science of language: human brains are computational systems for processing generative grammars. Classical view of parsing in NLP: Parser + Grammar The student forgot the answer. Parser Grammar:! S! NP VP! NP! D N! VP! V NP! VP! V S! VP! V!! VP! V NP NP!!! V! {forgot, ate}! N! {student, answer}! D The S NP VP NP N V D N student forgot the answer. 5
Levels of Representation and Processing Semantics: EXAMINED [DETECTIVE AGENT, EVIDENCE THEME ] S Agent and Theme are thematic roles (Who did what to whom?) Syntactic structure: NP VP NP D N V D N The detective examined the evidence. 11! Humans are limited-resource language processors Chomsky s distinction between competence : knowledge of the abstract rules of the grammar performance : the ability to apply those rules to linguistic input in producing or comprehending a sentence Limited by memory, attention, etc 12 6
just because it s grammatical doesn t mean we can understand it The dog that the rat that the cat ate bit died Center embedding. The dog that the rat that the cat ate bit died The dog died that the rat bit that the cat ate 14 7
Center embedding isn t always this difficult The patient that the doctor I know treated recovered from the illness. The airstrip that the plane the pilot flew landed at was tiny. The gazelle that the cheetah the tourists watched hunted escaped into the forest. 15 Ambiguity Lung Cancer in Women Mushrooms Teacher Strikes Idle Kids Stud Tires Out Shot Off Woman's Leg Helps Nicklaus to 66 Enraged Cow Injures Farmer with Axe Miners Refuse to Work after Death Squad Helps Dog Bite Victim Juvenile Court Tries Shooting Defendant Stolen Painting Found by Tree Killer Sentenced to Die for Second Time in 10 Years Drunk Drivers Paid $1,000 in 1984 Grandmother of Eight Makes Hole in One Two Convicts Evade Noose, Jury Hung 16! 8
Garden Path Errors and syntactic ambiguity The horse raced past the barn fell. The horse was raced past the barn (by somebody). The horse fell. CORRECT INTERPRETATION The horse raced past the barn. The barn fell. WRONG INTERPRETATION The cotton clothing is made of grows in Mississippi. The old man the boat. We painted the wall with cracks. 17 The modularity of syntactic processing? Language understanding is controlled by a syntactic processing system, which: Builds phrase structural representations (parses) Is fast and automatic Is functionally independent from outside influences (such as semantic or pragmatic processing). Operates with limited resources Precedes and guides semantic interpretation 18 9
Fodor (1983): many mental functions are accomplished by modules, which fulfill certain properties ( at least to some interesting extent ): 1. Domain specificity 2. Vulnerable to selective breakdown 3. Informational encapsulation 4. Obligatory firing 5. Fast, probably due to 3 and 4 above 6. Shallow outputs, the output of modules is very simple 7. Limited accessibility 8. Characteristic ontogeny, there is a regularity of development 9. Fixed neural architecture Syntax First models of syntactic ambiguity resolution (e.g., Frazier, 1987; 1989). The witness examined Stage 1: Syntactic commitment (strictly serial) Minimal processing of input word (e.g., D, N, V, etc.) Knowledge of phrase structure grammar Syntactic processing heuristics (e.g., avoid complexity) Stage 2: Semantic interpretation Assign thematic roles (e.g., Agent & Theme) Evaluate and revise if implausible. 20 10
The Garden-Path model: syntax first (Frazier, 1987; 1989). Initial commitments based on limited syntactic information: Coarse-grained lexical input: POS-only Temporary ambiguities resolved by syntactic heuristics Minimal Attachment Late Closure Second stage assigns thematic roles (Agent, Theme, etc) Stage 1 Commitments can be revised in second stage. Predicts frequent and highly uniform pattern of misanalysis. 11