600.465 - Intro to NLP - J. Eisner 1 Human Sentence Processing 600.465 - Intro to NLP - J. Eisner 2 Lexicalized Parsing  peel the apple on the towel  ambiguous  put the apple on the towel  put loves on (is the other reading even possible?)  put the apple on the towel in the box  VP[head=put]  V[head=put] NP PP  VP[head=put]  V[head=put] NP PP[head=on]  study the apple on the towel  study dislikes on (how can the PCFG express this?)  VP[head=study]  VP[head=study] PP[head=on]  study it on the towel  it dislikes on even more – PP can’t attach to pronoun 600.465 - Intro to NLP - J. Eisner 3 Lexicalized Parsing  the plan that Natasha would swallow  ambiguous between content of plan and relative clause  the plan that Natasha would snooze  snooze dislikes a direct object (plan)  the plan that Natasha would make  make likes a direct object (plan)  the pill that Natasha would swallow  pill can’t express a content-clause the way plan does  pill is a probable direct object for swallow  How to express these distinctions in a CFG or PCFG? 600.465 - Intro to NLP - J. Eisner 4 Human Performance: Self-Paced Reading Experiments Pretend 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 5 Human Performance: Self-Paced Reading Experiments  What 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 6 Big 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 heuristics  Get a parse this way, then interpret it semantically  Backtrack & fix if we can’t finish the parse or it makes little semantic sense  Sample heuristic: When you build a PP, attach it to the most recently built thing you can  Sample heuristic: When an NP starts a sentence, it’s the subject  If true, people should backtrack on “the necklace sold to the bank was …”  Hypothesis 2: Rely mainly on syntactic probabilities using head words  Smarter 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 built  Full interpretation as soon as we build a constituent  So semantic analysis and backtracking are never delayed  Garden paths result from genuinely ambiguous prefix, not slow semantics  Hypothesis 4: Consider full semantics of a constituent even before it’s built  Start interpreting a constituent before hearing it all  Semantics before syntax! (opposite of hypothesis 1)600.465 - Intro to NLP - J. Eisner 7 Eye Tracking  Self-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 8 Eye Tracking  The eyes are the window to the soul.  You are constantly scanning your environment.  Why? Evolution!  highly accurate at jumping to objects  3-4 of these “saccades” per second on average  fast motion (90 degrees of arc in 100 millisec)  low latency – short wires connect eyes to brain 600.465 - Intro to NLP - J. Eisner 9 Head-Mounted Eye Tracker  Like looking into someone’s thoughts  As they happen, in a real environment! 600.465 - Intro to NLP - J. Eisner 10 Videotape  From Mike Tanenhaus’s lab – University of Rochester Eye cameraScene cameraslide courtesy of M. Tanenhaus (modified) 600.465 - Intro to NLP - J. Eisner 11 The Visual World Paradigm look at the five of hearts look at the other five of hearts total 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 diamonds  600.465 - Intro to NLP - J. Eisner 12 The Visual World Paradigm now put the five of hearts that is below the eight of clubs above the three of diamonds Subject looks at 5 shortly after point of disambiguation (underlined) – only one 5 below an 8 Where would point of disambiguation be if only one of the 5 was below something? What if both 5 were below an 8? (8, 8 )slide courtesy of M. Tanenhaus slide courtesy of M. Tanenhaus (the video) slide courtesy of M. Tanenhaus 600.465 - Intro to NLP - J. Eisner 16 PP Attachment Ambiguity Put the apple on the towel in the box.slide courtesy of M. Tanenhaus (modified) Put the apple on the towel in the box.One referent context slide courtesy of M. Tanenhaus garden path oops! backtrack One referent context slide courtesy of M. TanenhausTwo-referent context slide courtesy of M. Tanenhaus amazing lack of oops 600.465 - Intro to NLP - J. Eisner 20 One-Referent Context put the apple on the towel in the box put the candy bar on the saucer in the bowl put the salt shaker on the envelope in the bowl (same picture as before – as shown on video) 600.465 - Intro to NLP - J. Eisner 21 Processing at Syllable Level Don’t wait for constituent to finish Don’t even wait for word to finish! pick up the candy They’re already looking based on “can” – we know this because if there’s also a candle, they’re 50% likely to look at it first! 600.465 - Intro to NLP - J. Eisner 22 Processing Stress Information Touch the large red square. Touch the LARGE red