Presentation on theme: "Methods: A Brief Survey Eva M. Fernández ABRALIN24FEB05 S NPVP."— Presentation transcript:
Methods: A Brief Survey Eva M. Fernández ABRALIN24FEB05 S NPVP
S NPVP …Tree-Building is Really Happening? Is syntactic structure psychologically real? even though it’s abstract… and not at all present in the signal? A simple test: RSVP task (see slides from 22FEB05)
RSVP Paradigm Center-screen, word-by-word display Timing: N ms per word (here: N = 500) Sentence-recall task
Thebeautifulblackcatchasedthecolorfulball. The beautiful black cat chased the colorful ball.
Blackcolorfultheballchasedcatbeautifulthe. Black colorful the ball chased cat beautiful the.
S NPVP What Information is Used? In building (recovering) syntactic structure, the information used includes: LEXICAL INFORMATION MORPHO-SYNTACTIC INFORMATION PHONOLOGICAL INFORMATION (including prosody) SEMANTIC INFORMATION GRAMMAR
S NPVP Three Operations of the Syntax Building simple structure… “Mary often speaks about Joe.” * “Mary often speak about Joe.” Combining simple sentences into complex ones… “Mary told Joe something.” + “Joe had his shoelace untied.” “Mary told Joe he had his shoelace untied.” Moving elements of sentences around… “Mary told Joe what” “What did Mary tell Joe __ ?”
S NPVP It follows that… If the grammar is consulted during sentence processing, we should expect that the system will dislike: UNGRAMMATICAL sentences COMPLEX sentences sentences involving lots of MOVEMENT dislike = take longer to process? be unable to process? just how could one measure the parser’s tastes? could one just ask the listener/reader?
S NPVP A Potential Problem You might reject a grammatical sentence because it’s prescriptively bad: Where’s the library at? Who did Joe meet at the party? Mary’s love life doesn’t concern you and I. Me and you shouldn’t be talking about Mary’s affairs. This is historic times. (GWB) etc.!
S NPVP Another (More Important) Problem You might reject a grammatical sentence because it’s hard (or impossible!) to process: Mary put the candy on the table in her mouth. When Madonna sings the song is always a hit. The son of Paraoh’s daughter looked at himself in the mirror. The cat the dog the boy walked bit meowed. The machine covered with paper plates handles with chrome. etc.!
S NPVP Processing Difficulty: Why? Why does the parser dislike some sentences more than others? Because of the listener’s inadequate knowledge of language? Because of sentence processing routines that have gone wrong, that have applied incorrectly?
S NPVP Processing Routines Defined Mechanisms that operate during production and perception: in constant contact with the grammar based on working-memory limitations Minimal Attachment: build the simplest structure Late Closure: attach locally Minimal Chains: posit the fewest filler-gap dependencies
S NPVP Psychological Reality, Again... How can you tell that the processing routines are being followed? Observe their work in action… Physics: which will fall faster, 1kg of feathers or 1kg of lead? Psycholinguistics: which linguistic stimulus will be understood more easily and faster, one with a Minimal Attachment violation or one without?
S NPVP Forster & Chambers, 1973 (described in Forster, 1979): stimuli: letter sequences were either words or orthographically legal non-words, e.g., thamon tasks: (i)naming (“pronounce ASAP”) (ii)lexical decision results: lexical decision times (608 ms) >naming times (508 ms)
S NPVP Forster, 1974 (also described in Forster, 1979): stimuli: grammatical, ungrammatical, non-word sentences 1.The scouts annoyed the lady 2.The bicycle the calculated cognac 3.The plane gleashed the passengers tasks: (i)are all items in input familiar words? (ii)is sentence meaningful? results:“no significant difference between decision times for the two tasks, and indeed for some types of sentences, the sentence task was slightly faster than the lexical task” (p. 30)
S NPVP Organization of Language Processor and GPS (after Forster, 1979) LEXICON MESSAGE PROCESSOR SYNTACTIC PROCESSOR LEXICAL PROCESSOR Gral PROBLEM SOLVER Gral CONCEPTUAL KNOWLEDGE INPUT FEATURES DECISION OUTPUT
S NPVP Ambiguity At almost any point during structure-building, the parser has multiple options (LOCAL AMBIGUITY!) S NPVP V NP We knew Ann’s date… S NPVP V S We knew Ann’s date… We knew Ann’s date, Joe. We knew Ann’s date would embarrass her at the party.
S NPVP Multiple Local Ambiguities… Have the soldiers marched into the barracks… … ? …, please! … had their supper? … take their boots off, would you?
S NPVP Ambiguity… Not to be confused with vagueness: “I guess the movie was interesting.” Occurs when a lexical string has two possible structures, GLOBALLY: “The bird is ready to eat.” “Visiting relatives can be a real nuisance.” “Joe said Mary called him yesterday.” “Joe saw Mary with the telescope.” “Joe saw a dog next to a kitten with an orange sweater.” or LOCALLY: “We knew Ann’s date “Have the soldiers marched …, Joe. … would embarrass her at the party. … into the barracks? … into the barracks, please! … into the barracks eaten?
S NPVP When there’s an ambiguity… How does the parser go about choosing among the alternatives? Does the parser notice all of the alternatives? If so, are all of the alternatives kept active? A debate in the literature, indeed, but the consensus: THE PARSER NOTICES THE STRUCTURAL ALTERNATIVE THAT’S EASIEST TO BUILD, following strategies like: Minimal Attachment: build the simplest structure Late Closure: attach locally Minimal Chains: posit the fewest filler-gap chains
S NPVP AN EXPERIMENT The legionnaires marched into the desert and searched for the nearest oasis. The legionnaires marched into the desert surprised the Persian forces. MY HYPOTHESIS… 2 nd sentence is harder: violates Minimal Attachment let’s see how!
S VP marched into the desert NP the legion- naires The legionnaires marched into the desert and searched for the nearest oasis. conj and VP searched for the nearest oasis
marched into the desert S NP the legion- naires The legionnaires marched into the desert surprised the Persian forces. NP marched into the desert S VP surprised the Persian forces Uh - oh!
S NPVP THE EXPERIMENTAL METHOD Hypothesis testing you have an idea about how structures are built by the sentence processor you test it by designing an experiment you make predictions about how the experiment will come out you think about what it will mean if the experiment comes out the “wrong way” this will be easier if you use BINARY COMPARISONS
S NPVP PARTICIPANTS never too many? “clones” of each other selection criteria: language history, education history, sex, handedness, age, reading ability, etc. controlling: use background questionnaire use behavioral criteria
S NPVP MATERIALS never too many? instantiating the contrast you’re interested in normed in whatever relevant way interspersed among distractor items never too many? usual filler-target ratio, 2-1 or 3-1 pseudo-randomized lists featuring binary comparisons (unless absolutely necessary)
S NPVP PROCEDURE “speeded classification tasks”: “the subject of the experiment is presented with some item of linguistic input, which must be classified according to some experimenter-defined criterion” try several types! (ingenuity is essential) don’t be fooled by technology: just because it costs more doesn’t mean it’s more effective just because everyone says it’s online doesn’t mean it taps the phase of processing you want (e.g., how do you know that information hasn’t flowed through a “later” processor before the GPS makes a decision?)
S NPVP QUESTIONNAIRES a grammaticality / acceptability judgment task On a scale of 1-7, indicate how acceptable you think each of these sentences are. 1 = perfectly acceptable 7 = perfectly awful measure: most frequent response type
The legionnaires marched into the desert and searched for the nearest oasis The legionnaires marched into the desert surprised the Persian forces
S NPVP ASSETS & LIABILITIES of questionnaire procedures? flexible low-tech lab test 100 people at a time! off-line: metalinguistic awareness, prescriptive rules subjects might engage in undesirable behavior: strategies, looking back, etc.
S NPVP RSVP rapid serial visual presentation words presented center-screen 500 msec for each word presentation rate could be varied task is to remember entire sentence and repeat or write down measure: accuracy of recall
S NPVP SELF-PACED READING subject controls pace of presentation constraints on speed determined by experimenter: time-outs, instructions materials presented in a series of “chunks” word-by-word, phrase-by-phrase, clause-by-clause different types of displays centered, incremental, moving window measure: reading time
S NPVP KEEPING THE PARTICIPANTS HONEST asking questions after every trial after every N trials never? excluding error-prone subjects providing instant feedback the speed-accuracy trade-off: the faster you respond, the more likely you’ll make errors
Thestudenttoldtheprofessorthateveryonehatedalie. Who told a lie? the student the professor Did the professor lie? Did the student lie? correct 2359
S NPVP SAME-DIFFERENT SENTENCE-MATCHING
The legionnaires marched into the desert and searched for the nearest oasis.
The alligator with the sharp teeth inspected the rifle. The alligator with the sharp teeth inspected his rifle.
The legionnaires marched into the desert surprised the Persian forces.
Everyone at the party knew Ann’s date had made a fool of himself. Everyone at the party knew Ann’s mate had made a fool of himself.
S NPVP EYE-TRACKING: fine-grained visual span: 9 letters, + periphery focus & move: eyetracker records focus measurements: first fixations regressions total reading See how the eye moves across the page in normal reading: animation by Steve Mansfield & Tim Klitz
S NPVP Eyetracking: challenges a more “on-line” method than SPR? the problem of time during saccades the problem of as much time as necessary for a given item the problem of peripheral vision more limitations length of items limited by size of display costly equipment bite-bar laborious analyses
S NPVP EYETRACKING: coarse-grained “Real World” paradigm Head-mounted eyetrackers permit examining how the immediate visual context can affect sentence processing. Easy to use, with adults... and children Evidence of interactive system for language processing? (See Garrett, 2000, for discussion.)
S NPVP BRAIN IMAGING “distinct brains systems associated with several different structural domains…: phrase structure, inflection, movement and binding, and lexical semantics” (Garrett, 2000, p. 43) semantic versus syntactic task priorities: different electrode locations different timing early responses to syntax: inflexible, localized in classic language processing areas later responses to syntax: more sensitivity to interpretative factors more broadly distributed