Human Sentence ProcessingLexicalized ParsingSlide 3Human Performance: Self-Paced Reading ExperimentsSlide 5Big Human-Parsing Debate of the 1990’s: How Soon Does Semantics Come Into Play?Eye TrackingSlide 8Head-Mounted Eye TrackerVideotapeThe Visual World ParadigmSlide 12Slide 13Slide 14Slide 15PP Attachment AmbiguityOne referent contextSlide 18Two-referent contextOne-Referent ContextProcessing at Syllable LevelProcessing Stress Information600.465 - Intro to NLP - J. Eisner 1Human Sentence Processing600.465 - Intro to NLP - J. Eisner 2Lexicalized Parsingpeel the apple on the towelambiguousput the apple on the towelput loves on (is the other reading even possible?)put the apple on the towel in the boxVP[head=put] V[head=put] NP PPVP[head=put] V[head=put] NP PP[head=on]study the apple on the towelstudy dislikes on (how can the PCFG express this?)VP[head=study] VP[head=study] PP[head=on]study it on the towelit dislikes on even more – PP can’t attach to pronoun600.465 - Intro to NLP - J. Eisner 3Lexicalized Parsingthe plan that Natasha would swallow ambiguous between content of plan and relative clausethe plan that Natasha would snooze snooze dislikes a direct object (plan)the plan that Natasha would makemake likes a direct object (plan)the pill that Natasha would swallowpill can’t express a content-clause the way plan doespill is a probable direct object for swallowHow to express these distinctions in a CFG or PCFG?600.465 - Intro to NLP - J. Eisner 4Human Performance: Self-Paced Reading ExperimentsPretend you’re pressing a key to get each word:The shop sold to the bank was quite old. We included this sentence just to distract you.The necklace sold to the bank was quite old.The lawyer examined by the judge was silly.The evidence examined by the judge was silly.600.465 - Intro to NLP - J. Eisner 5Human Performance: Self-Paced Reading ExperimentsWhat happens to a self-paced reader when she goes down the garden path?Eventually she has to backtrack.That’s when she pauses: “point of disambiguation.”The lawyer examined by the judge ... Why isn’t the second sentence a garden path:The shop sold to the bank was quite old. The necklace sold to the bank was quite old.People are sensitive to frequency!600.465 - Intro to NLP - J. Eisner 6Big Human-Parsing Debate of the 1990’s: How Soon Does Semantics Come Into Play?How fast is semantics? Use it constantly, or only as a last resort?Hypothesis 1: Rely mainly on syntactic heuristicsGet a parse this way, th en interpret it semanticallyBacktrack & fix if we can’t finish the parse or it makes little semantic senseSample heuristic: When you build a PP, attach it to the most recently built thing you canSample heuristic: When an NP starts a sentence, it’s the subjectIf true, people should backtrack on “the necklace sold to the bank was …”Hypothesis 2: Rely mainly on syntactic probabilities using head wordsSmarter version of hypothesis 1: “necklace” isn’t a common subject for “sell”Explains why no backtracking on “the necklace sold to the bank was …”Hypothesis 3: Consider full semantics of a constituent as soon as it’s builtFull interpretation as soon as we build a constituentSo semantic analysis and backtracking are never delayedGarden paths result from genuinely ambiguous prefix, not slow semanticsHypothesis 4: Consider full semantics of a constituent even before it’s builtStart interpreting a constituent before hearing it allSemantics before syntax! (opposite of hypothesis 1)600.465 - Intro to NLP - J. Eisner 7Eye TrackingSelf-paced reading data too flaky to answer these questions.Brain imaging is too slow and coarse.Track people’s eye movements as they read.They don’t backtrack on “the necklace sold to the bank was …”Cleaner data; eliminates hypothesis 1. But reading is an artificial task – people didn’t evolve to be good readers.600.465 - Intro to NLP - J. Eisner 8Eye TrackingThe eyes are the window to the soul.You are constantly scanning your environment.Why? Evolution!highly accurate at jumping to objects3-4 of these “saccades” per second on averagefast motion (90 degrees of arc in 100 millisec)low latency – short wires connect eyes to brain600.465 - Intro to NLP - J. Eisner 9Head-Mounted Eye TrackerLike looking into someone’s thoughtsAs they happen, in a real environment!600.465 - Intro to NLP - J. Eisner 10VideotapeFrom Mike Tanenhaus’s lab– University of RochesterEye cameraScene cameraslide courtesy of M. Tanenhaus (modified)600.465 - Intro to NLP - J. Eisner 11The Visual World Paradigmlook at the five of heartslook at the other five of heartstotal time: 4.15 seconds(as shown on video)now put the five of hearts that is below the eight of clubs above the three of diamonds600.465 - Intro to NLP - J. Eisner 12The Visual World Paradigmnow put the five of hearts that is below the eight of clubs above the three of diamondsSubject looks at 5 shortly after point of disambiguation (underlined) – only one 5 below an 8Where would point of disambiguation beif only one of the 5 was below something?What if both 5 were below an 8? (8, 8)00.10.20.30.40.50.60.70.85 hearts that is below the eight of clubs. . .Probability of eye movementEarly disambiguation condition888 88MiscellaneousRelatum-targetAlternativeTarget8slide courtesy of M. Tanenhaus00.10.20.30.40.50.60.70.8Mid disambiguation condition5 hearts that is below the eight of clubs. . .Probability of eye movement8101010108888MiscellaneousRelatum-alternative10Relatum-target8TargetAlternativeslide courtesy of M. Tanenhaus(the video)00.10.20.30.40.50.60.70.8Late disambiguation condition5 hearts that is below the eight of clubs. . .Probability of eye movementMiscellaneousRelatum-targetRelatum-alternativeTargetAlternativeslide courtesy of M. Tanenhaus600.465 - Intro to NLP - J. Eisner 16PP Attachment AmbiguityPut the apple on the towel in the box.Only one apple Garden PathTwo apples use PP to clarify which apple, no garden pathslide courtesy of M. Tanenhaus (modified)Put the apple on the towel in the box.One referent contextA,A'BCD,B'"Put the apple on the towel in the box."A B C DA'B'"Put the apple that's on the towel in the box."ms 0 500 1000 1500
View Full Document