Ayaka Sugawara, Hadas Kotek, Martin Hackl, & Kenneth Wexler MIT Glow in Asia IX at Mie University 9/6/2012 Slides available at http://web.mit.edu/ayakasug/www/acd.pdf Do English speaking children correctly interpret the Antecedent Contained Deletion (ACD) sentences like these? The Mermaid baked himi the same food that Cookie Monster*i did. Dora gave himi the same color paint that Smurf’si father did. Hei jumped over every fence that Kermit*i tried to. (Kiguchi & Thornton 2004) Miss Piggy wanted to drive every car that Kermit did. (Syrett & Lidz 2011) Kids behave like adults for the at least first three sentences. 2 In this talk: I will show how the previous studies revealed that kids have adult-like knowledge to interpret those sentences. I will discuss the results from a new set of experiments to test whether kids can do long-distance QR. The results support and strengthen what Syrett & Lidz (2011) report. 3 1. 2. Introduction – What is ACD? Previous studies - Kiguchi & Thornton (2004) “Kids know Binding Principles in ACD sentences” - Syrett & Lidz (2009) “Kids distinguish ACD from coordinate structure” - Syrett & Lidz (2011) “Kids can interpret both short & long distance QR” 3. 4. Our experiment Results and discussion 4 1. What is ACD? ACD is a special case of VP-ellipsis. (Bouton 1970, Sag 1976) “John watched every movie that Mary did.” VP-ellipsis requires the structural identity at LF between the VPant and the VPelid. John will [VPant eat an apple], but Bill won’t <VPelid eat an apple>. *<eat an orange> *<peeled an apple> *<have eaten an apple> In ACD sentences, for the VPelid and the VPant to have the identical LF, Quantifier Raising (QR) is necessary. 6 John watched every movie that Mary did. Without QR TP Antecedent VP 3VP John 3DP watched 3NP Elided VP every 3CP movie 3 Op 3TP that 3 Mary 3 did < VP > (1) 7 John watched every movie that Mary did. Without QR TP 3VP John 3DP Antecedent VP watched 3NP every 3CP Elided VP movie 3 Op 3TP that 3 Mary 3 did < VP > (1) J (did) [VP watch every movie that M did <watch every movie that M did <watch every movie that M did <watch … “infinite regress” problem 8 John watched every movie that Mary did. With QR qpDP 3VP 3NP John 3 every 3CP watch t movie 3 Op 3TP Elided VP that 3 Antecedent VP Mary 3 did <watch t > (1) 9 ACD necessarily involves the operation of QR. The target site of the QR must be above the antecedent VP in order for the elided VP to be identical to the antecedent VP by generating the trace. 10 Processing study on local and non-local ACD with adults (Hackl et al. 2012) “The lawyer was careful to remember … a. the/every fact that the young defense attorney presented b. the/every fact that the young defense attorney did c. the/every fact that the young defense attorney was … during the second cross examination period.” - Self-paced reading, followed by a question to ask the understanding of the sentence, with a rating task added. 180 sentences including fillers were presented word by word on a computer screen. Reaction Time was measured. The lawyer was careful to remember every fact that the attorney did. Local ACD TP wo L wo was careful wovP to qp 3 6 remember t every fact that A did <remember t> The lawyer was careful to remember the fact that the attorney was. NonNon-local ACD TP qp L qp PredP was qp qp 6 careful 3 every fact that A was to 3 <careful to remember t> remember t Hackl et al. (2012): Results (N=48) • Non-local QR requires more processing load than local QR. • Local QR induced by every facilitates ACD resolution. Non-ACD RC local non-local 2. Previous studies Sentences they tested: The Mermaid baked himi the same food that Cookie Monster*i did. Rejection of coreference: 85% Dora gave himi the same color paint that Smurf’si father did. Acceptance of coreference: 93.3% Hei jumped over every fence that Kermit*i tried to. Rejection of coreference: 94.2% 15 K&T in a nutshell Children (4- and 5-year olds) understand ACD sentences. QR is in child grammar. What kind of ACD sentences? ACD sentences in which Binding Principle B is relevant at LF. They know how Principle B works. ACD sentences in which Binding Principle C is relevant at LF. They know how Principle C works. Syrett & Lidz (2009) confirmed that children understand this type of ACD sentences correctly, instead of misapplying coordinate structure to them. The Mermaid baked him the same food, and Cookie Monster did so, too. 16 S & L (2011) in a nutshell Children (4-year-olds) can access the multiple possible interpretations in non-ACD VP-ellipsis & ACD sentences. Non-ACD Clifford asked Goofy to read the big books because Scooby did. (7) a. read the big books. b. asked Goofy to read the big books. ACD “Miss Piggy wanted to drive every car that Kermit did.” <drive t.> <want to drive t.> 17 Experiment 1 (non (nonnon-ACD VP Ellipsis) Embedded condition: There are two piles of books, big books and small books. Scooby “Goofy, read the big books.” (Scooby reads the small books.) Clifford “Goofy, read the small books, because Scooby read them and liked them.” Puppet “Clifford asked Goofy to read the small books because Scooby did.” Yes (A: 81%, C: 69%) 18 Experiment 1 (non (nonnon-ACD VP Ellipsis) Matrix condition: There are two piles of books, big books and small books. Scooby “Goofy, read the big books.” (Scooby reads the small books.) Clifford “Goofy, read the big books, because Scooby always has good ideas.” Puppet “Clifford asked Goofy to read the big books because Scooby did.” Yes (A: 94%, C: 71%) 19 Summary for Experiment 1 (nonnon-ACD VP Ellipsis) Children (4-year-olds) can access the multiple possible interpretations in non-ACD ambiguous VP-ellipsis sentences. They behave like adults. 20 Experiment 2 (ACD) Subjects: 24 four-year-olds, 30 undergrads Procedure: Truth Value Judgment Task followed by justification of the answer. Embedded/matrix condition is a between subject factor. Target sentences: (8) (9) Miss Piggy wanted PROsubj to drive every car that Kermit did. Clifford asked Goofy PROobj to read every book that Scooby did. 21 Experiment 2: Embedded condition Kermit has red cars, which are old, and black cars, which are new. K drives K wants to drive He has driven the red ones and is tired of them. He wants to drive the black ones but is not allowed to do so. Miss Piggy asks him to drive some, and he reluctantly drives the red ones. After the drive, Miss Piggy gets very excited and wants to drive some. 22 Experiment 2: Embedded condition: Miss Piggy wants to drive the red ones. K drives K wants to drive Puppet “Miss Piggy wanted to drive every car that Kermit did.” Yes (Adults: 68%, 68% Children: 46%) 46% 23 Experiment 2: Matrix condition Kermit has red cars, which are old, and black cars, which are new. K drives K wants to drive He has driven the red ones and is tired of them. He wants to drive the black ones but is not allowed to do so. Miss Piggy asks him to drive some, and he reluctantly drives the red ones. After the drive, Miss Piggy gets very excited and wants to drive some. 24 Experiment 2: Matrix condition: Miss Piggy agrees that the red ones are not exciting enough, and wants to drive the black ones. K drives K wants to drive Puppet “Miss Piggy wanted to drive every car that Kermit did.” Yes (Adults: 50%, 50% Children: 38%) 38% 25 Miss Piggy wanted to drive every car that Kermit did. Embedded wo P wo want wo to qp 3 6 drive t every car that K did Matrix <drive t> qp P qp qp 6 want 3 every car that K did <want to drive t> to 3 drive t Discussion Although children do poorly (Emb: 46%, Mat: 38%), the justifications for rejection suggest that both adults and kids are accessing the other reading. e.g. Miss Piggy wanted to drive the new cars. Kermit drived the old cars. (in Matrix condition) In this experiment, embedded and matrix condition was a between subject factor, so each person is assigned only one condition. 7 kids and 12 adults provided justifications for both readings within an experiment session. Drawing on those justifications, they conclude that kids can access both interpretations. 27 The reliable justifications (clearly indicating either reading) by children were given to them 54% of the time (52/96). We are not sure if the justifications are good enough to make arguments. So we’d like to clarify. Where is the difference between chance-level and good understanding? Accuracy rate by children was “Emb: 46%, Mat: 38%.” Is there a possibility that they just guessed, and only to defend themselves they randomly picked one of the possible antecedents when asked for justifications? Adults’ accuracy rate was low (Emb: 68%, Mat: 50%). This does not show that the correct interpretations are what children should know. 28 3. Our experiment 29 Motivation Can children really target multiple landing sites of QR in ACD sentences? Do they make judgments according to the reading they get? Target sentences (10) (11) (12) (13) Donald Duck wanted to be the same color that Mickey Mouse was. Elmo wanted to be the same animal that Lisa was. Donald Duck wanted to be the same color that Mickey Mouse did. Elmo wanted to be the same animal that Lisa did. 30 (10) (11) (12) (13) Donald Duck wanted to be the same color that Mickey Mouse was. I know what happened. Elmo wanted to be the same animal that Lisa was. Donald Duck wanted to be the same color that Mickey Mouse did. I know what happened. Elmo wanted to be the same animal that Lisa did. Contributions The sentences (12, 13) test whether kids can do the longlong-distance QR (and give the judgment sticking to the interpretation). “No” conditions allow us to see whether children can correctly reject the unavailable reading. 31 TVJT: stories on a computer screen 61 children (4;2-7;9, mean 5;4) at daycare centers in Boston area and at Boston Children’s Museum. 4 target trials (all designated to be false) and 3 filler trials (2 to be true, 1 to be false) Yes-sayers (N=12), no-sayers (N=4), those who did not complete the experiment (N=2), and those who did not answer two filler items correctly (N=11) were excluded Short vs. Long ACD was a between-subject factor. Short ACD (N=15), Long ACD (N=16) We only had target sentences that are designated to be false, to strengthen the argument. We have unambiguously matrix-targeting ACD sentences, to see whether children actually access the interpretation. Filler items are quite similar to target sentences, e.g. “Mickey Mouse wanted to be at the same party that Donald Duck went to.” Short ACD Cookie Monster wanted to be the same thing that Dora was. No (short QR) Total participants: 61 Included participants: 31 (excl. 12 yes-sayers, 4 nosayers, 2 not-completing, 11answering 0-1 fillers correctly) Short condition: 15 Long condition: 16 Accuracy rate w/ short: 62% (per item: 73%, 40%, 60%, 73%) Accuracy rate w/ long: 81% 75%) (per item: 81%, 94%, 75%, Significant difference between long QR and short QR. Long QR is significantly better than Short QR (81.2% vs. 61.5%). Accuracy rate Long QR 81.2% Short QR 61.5% Wilcoxon rank sum test with continuity correction, W=1544, p=0.016 Calculating p-value according to the standard deviations (SD) (assuming that the chance=.5) and their z-scores also results in significant results (p=0.017) Both above chance. (Assumption: chance = 0.5) Long QR SD z-score p-value w/.5 Short QR 0.063 0.064 4.96 1.86 2*pnorm(-abs(long.z)) 7.0e-07 2*pnorm(-abs(short.z)) 0.06 Given that we only had false target sentences and that children have yes-bias, the “chance”-level would be smaller than 50%. We observe significant (long) and marginally significant (short) differences from “50% chance.” This suggests that children do understand ACD sentences, both short and long. As for the reason why children do better with long ACD than short ACD, we have ideas to be tested by running another set of experiments. We are currently running two versions of the experiment. With “every” Elmo wanted to be every animal that Lisa was/did. With “about to” Donald Duck was about to look into the same box that Mickey Mouse did/was. We anticipate that with “every,” the trend will flip, i.e., lower accuracy rate for long QR and higher accuracy rate for short QR. This anticipation stems from the processing data from adults (Hackl et al 2012). Bouton, L. F. (1970) “Antecedent contained pro-forms,” in Papers from the Sixth Regional Meeting of the Chicago Linguistics Society, 154-167. Hackl, Martin, Jorie Koster-Moeller & Jason Varvoutis (2012) “Quantification and ACD: Evidence from Real Time Sentence Processing,” Journal of Semantics 29 (2), 145-205. Kiguchi, Hirohisa & Rosalind Thornton (2004) “Binding Principles and ACD Constructions in Child Grammars,” Syntax 7, 234-271. Sag, Ivan (1976) Deletion and Logical Form, PhD dissertation, MIT. Syrett, Kristen & Jeffrey Lidz (2009) “QR in Child Grammar: Evidence from Antecedent-Contained Deletion,” Language Acquisition 16, 67-81. Syrett, Kristen & Jeffrey Lidz (2011) “Competence, Performance, and the Locality of Quantifier Raising: Evidence from 4-Year-Old Children,” Linguistic Inquiry 42, 305-337. 49