Previously Triggered Dependency Search Overrides Local Coherence Effects

Size: px

Start display at page:

Download "Previously Triggered Dependency Search Overrides Local Coherence Effects"

Joseph Jacobs
5 years ago
Views:

1 Previously Triggered Dependency Search Overrides Local Coherence Effects 20 TH AMLAP SEPTEMBER LAUREN ACKERMAN NINA KAZANINA MASAYA YOSHIDA

2 The Question Does online long-distance dependency formation resolve local ambiguity?

3 Road Map 1. Introduction Local Attachment Active Search Competition for the parser s attention 2. Eye-tracking Study Early measures Late measures A possible conflict? 3. Discussion

4 Local Attachment As the parser encounters a string, attempts to form a coherent analysis: [VP hums ]][IP[NP Annie s song] ] [CP Whenever [IP he hums [VP hums]][ip [NP Annie s song] * one possible structure-building method

5 Local Attachment As the parser encounters a string, it attempts to form a coherent analysis: [VP hums ]][IP[NP Annie s song] ] [CP Whenever [IP he hums [VP hums]][ip [NP Annie s song] * one possible structure-building method

6 Local Attachment As the parser encounters a string, it attempts to form a coherent analysis: [VP hums ]][IP[NP Annie s song] ] [CP Whenever [IP he hums [VP hums]][ip [NP Annie s song] * one possible structure-building method

7 Local Attachment As the parser encounters a string, it attempts to form a coherent analysis: [CP Whenever [IP he hums Intransitive or Transitive? [VP hums ]][IP[NP Annie s song] ] [VP hums]][ip [NP Annie s song] * one possible structure-building method

8 Local Attachment As the parser encounters a string, it attempts to form a coherent analysis: [VP hums ]][IP[NP Annie s song] ] [CP Whenever [IP he hums Annie's song [VP hums]][ip [NP Annie s song] * one possible structure-building method

9 Local Attachment As the parser encounters a string, it attempts to form a coherent analysis: [VP hums [NP Annie s song] ] [CP Whenever [IP he hums Annie's song [VP hums]][ip [NP Annie s song] * one possible structure-building method

10 Local Attachment As the parser encounters a string, it attempts to form a coherent analysis: Local Attachment [VP hums [NP Annie s song] ] [CP Whenever [IP he hums Annie's song [VP hums]][ip [NP Annie s song] e.g., Tabor & Hutchins (2004); Tabor, Galantucci & Richardson (2004) * one possible structure-building method

11 Active Search The parser also seems to attempt to complete dependencies as early as possible (e.g., van Gompel & Liversedge 2003, Kazanina et al. 2007) [CP Whenever he hums, [IP Annie s song

12 Active Search The parser also seems to attempt to complete dependencies as early as possible (e.g., van Gompel & Liversedge 2003, Kazanina et al. 2007) Male NP [CP Whenever he hums, [IP Annie s song

13 Active Search The parser also seems to attempt to complete dependencies as early as possible (e.g., van Gompel & Liversedge 2003, Kazanina et al. 2007) Male NP [CP Whenever he hums, [IP Annie s song

14 Active Search The parser also seems to attempt to complete dependencies as early as possible (e.g., van Gompel & Liversedge 2003, Kazanina et al. 2007) Male NP [CP Whenever he hums, [IP Annie s song

15 Active Search The parser also seems to attempt to complete dependencies as early as possible (e.g., van Gompel & Liversedge 2003, Kazanina et al. 2007) Male NP? [CP Whenever he hums, [IP Annie s song

16 Active Search The parser also seems to attempt to complete dependencies as early as possible (e.g., van Gompel & Liversedge 2003, Kazanina et al. 2007) No!! [CP Whenever he hums, [IP Annie s song

17 Active Search The parser also seems to attempt to complete dependencies as early as possible (e.g., van Gompel & Liversedge 2003, Kazanina et al. 2007) GMME [CP Whenever he hums, [IP Annie s song

18 Active Search The parser also seems to attempt to complete dependencies as early as possible (e.g., van Gompel & Liversedge 2003, Kazanina et al. 2007) But only when the grammar allows. [CP Whenever he [VP hums [NP Annie s song

19 Active Search The parser also seems to attempt to complete dependencies as early as possible (e.g., van Gompel & Liversedge 2003, Kazanina et al. 2007) But only when the grammar allows. IP [CP Whenever he [VP hums [NP Annie s song he VP hums Annie's song The NP is c-commanded by the pronoun (thus coreference violates Condition C)(Chomsky 1981)

20 Active Search As the parser encounters a string, it must attempt to form a coherent analysis: [VP hums [NP Annie s song] ] [CP Whenever [IP he hums Annie's song [VP hums]][ip [NP Annie s song] * one possible structure-building method

21 Active Search As the parser encounters a string, it must attempt to form a coherent analysis: [VP hums [NP Annie s song] ] [CP Whenever [IP he hums Annie's song Coreference Dependency [VP hums]][ip [NP Annie s song] e.g., Kazanina et al. (2007) * one possible structure-building method

22 What is at stake? Whenever he hums Annie s song Local Attachment biases the parser to analyze Annie s song as the embedded object

23 What is at stake? Whenever he hums Annie s song Local Attachment biases the parser to analyze Annie s song as the embedded object Active Search biases the parser to complete the dependency as early as possible, with Annie s song analyzed as a matrix subject

24 What is at stake? Whenever he hums Annie s song Local Attachment biases the parser to analyze Annie s song as the embedded object Active Search biases the parser to complete the dependency as early as possible, with Annie s song analyzed as a matrix subject These two biases pull the parser in different directions

25 What is at stake? Whenever he hums Annie s song Local Attachment biases the parser to analyze Annie s song as the embedded object Active Search biases the parser to complete the dependency as early as possible, with Annie s song analyzed as a matrix subject These two biases pull the parser in different directions Does the parser prioritize one?

26 Previous studies Kwon & Sturt (2014) Before PRO running, Martin had breakfast. Before PRO washing the kids

27 Road Map 1. Introduction Local Attachment Active Search Competing pressures 2. Eye-tracking Study Early measures Late measures A possible conflict? 3. Discussion

28 Predictions Hypothesis #1: Parser prioritizes Hypothesis #2: Parser prioritizes

29 Hypothesis #1 Whenever he hums Annie s song The parser prioritizes forming a local attachment

30 Hypothesis #1 The parser prioritizes forming a local attachment Whenever he hums Annie s song ambiguous

31 Hypothesis #1 The parser prioritizes forming a local attachment Whenever he hums Annie s song ambiguous CP IP CP IP whenever he hums NP VP Annie s song V Matrix subject he VP hums NP Annie s song Embedded Object

32 Hypothesis #1 The parser prioritizes forming a local attachment Whenever he hums Annie s song ambiguous CP IP CP IP whenever he hums NP VP Annie s song V Matrix subject Requires non-local resolution he VP hums NP Annie s song Embedded Object Can have local resolution

33 Hypothesis #1 The parser prioritizes forming a local attachment Whenever he hums Annie s song ambiguous CP IP CP IP whenever he hums NP VP Annie s song V Matrix subject Requires non-local resolution he VP hums NP Annie s song Embedded Object Can have local resolution

34 Hypothesis #1 The parser prioritizes forming a local attachment Whenever he hums Annie s song is. CP ambiguous matrix subject predicted IP CP IP whenever he hums NP VP Annie s song V Matrix subject Requires non-local resolution he VP hums NP Annie s song Embedded Object Can have local resolution

35 Predictions Hypothesis #1: Parser prioritizes Hypothesis #2: Parser prioritizes Embedded object is c- commanded by cataphor Thus cannot corefer No slow-down predicted if gender mismatches Parser is surprised by the matrix verb Slow-down predicted

36 Hypothesis #2 The parser prioritizes forming a coreference dependency Whenever he hums Annie s song ambiguous

37 Hypothesis #2 The parser prioritizes forming a coreference dependency Whenever he hums Annie s song ambiguous CP IP CP IP whenever he hums NP VP Annie s song V Matrix subject he VP hums NP Annie s song Embedded Object

38 Hypothesis #2 The parser prioritizes forming a coreference dependency Whenever he hums Annie s song ambiguous CP IP CP IP whenever he hums NP VP Annie s song V Matrix subject No c-command, Coreference possible he VP hums NP Annie s song Embedded Object C-command, Coreference impossible

39 Hypothesis #2 The parser prioritizes forming a coreference dependency Whenever he hums Annie s song ambiguous CP IP CP IP whenever he hums NP VP Annie s song V Matrix subject No c-command, Coreference possible he VP hums NP Annie s song Embedded Object C-command, Coreference impossible

40 Hypothesis #2 The parser prioritizes forming a coreference dependency Whenever he hums Annie s song is. CP ambiguous IP matrix V predicted CP IP whenever he hums NP VP Annie s song V Matrix subject No c-command, Coreference possible he VP hums NP Annie s song Embedded Object C-command, Coreference impossible

41 Predictions Hypothesis #1: Parser prioritizes Hypothesis #2: Parser prioritizes Embedded object is c- commanded by cataphor Thus cannot corefer No slow-down predicted if gender mismatches Parser is surprised by the matrix verb Slow-down predicted Matrix subject is not c- commanded by cataphor Thus can corefer Slow-down at matrix subject if gender mismatches Parser expects the matrix verb No slow-down predicted

Matched Mismatched No Comma Whenever she was about to hum Annie s favorite song was Whenever he was about to hum Annie s favorite song was Comma Whenever she was about to hum, Annie s favorite song

42 Matched Mismatched No Comma Whenever she was about to hum Annie s favorite song was Whenever he was about to hum Annie s favorite song was Comma Whenever she was about to hum, Annie s favorite song was Whenever he was about to hum, Annie s favorite song was Eye-tracking methodology: 36 Northwestern undergrads 2 x 2 design, Gender Congruency vs Comma Presented in full sentences, one sentence at a time Comprehension questions after every sentence All results analyzed by Linear Mixed Effects Regressions (LMER)

43 No Comma Comma Matched Whenever she was about to hum Annie s favorite song was Whenever she was about to hum, Annie s favorite song was Mismatched Whenever he was about to hum Annie s favorite song was Whenever he was about to hum, Annie s favorite song was If local attachment: Annie s song is misanalysed as an embedded object à RT slowdown at matrix verb in No Comma conditions

44 Matched Mismatched No Comma Whenever she was about to hum Annie s favorite song was Whenever he was about to hum Annie s favorite song was Comma Whenever she was about to hum, Annie s favorite song was Whenever he was about to hum, Annie s favorite song was If dependency formation: Annie s favorite song is analysed as a Matrix subject à RT slowdown in Gender-mismatch conditions

First Pass Reading Time Annie s Matrix subject region: Gender Mismatch Effect Supports Hypothesis #2 Parser prioritizes forming a dependency early

45 First Pass Reading Time Annie s Matrix subject region: Gender Mismatch Effect Supports Hypothesis #2 Parser prioritizes forming a dependency early Parser attempts to form a cataphoric dependency at matrix subject Main effect of Gender: p<.03 Main effect of Comma: n.s. Interaction (Gender vs Comma) : n.s.

Regression Path Time was spillover region Matrix verb spillover region: Reading time slow-down Supports Hypothesis #1 Parser prioritizes forming a local

46 Regression Path Time was spillover region Matrix verb spillover region: Reading time slow-down Supports Hypothesis #1 Parser prioritizes forming a local attachment analysis Parser is surprised by the intransitive verb Main effect of Gender: n.s. Main effect of Comma: p<.001 Interaction (Gender vs Comma) : n.s.

47 Total Time was Matrix verb region: Reading time slow-down Supports Hypothesis #1 Parser prioritizes forming a local attachment analysis Parser is surprised by the intransitive verb Main effect of Gender: n.s. Main effect of Comma: p<.02 Interaction (Gender vs Comma) : n.s.

48 Road Map 1. Introduction Local Attachment Active Search Competing pressures 2. Eye-tracking Study Early measures Late measures A possible conflict? 3. Discussion

49 Discussion Parser prioritizes cataphoric dependency formation (first-pass effects at the matrix subject) Online formation of a cataphoric dependency affects local ambiguity resolution But it shows effects of a local attachment analysis (regression path & total time at the matrix verb)

50 Discussion Multiple simultaneous representations (e.g., Slattery et al. 2013, Gibson & Pearlmutter 2000, Pearlmutter & Mendelsohn 1999) On each trial the parser maintains multiple representations

51 Discussion Multiple non-simultaneous representations (Kwon & Sturt 2014) On a given trial the parser is guided by either Active Search or Local Attachment Complementary subsets of trials showing a GMME at Annie's vs. garden path at 'was'

52 Thank you! SPECIAL THANKS TO: NSF GRANT BCS & BCS MICHAEL FRAZIER COREY NELSON THE NU SYNTAX/SEMANTICS LAB GROUP

53 References Christianson, Hollingworth, Halliwell & Ferreira. (2001). Thematic Roles Assigned along the Garden Path Linger. Cognitive Psychology 42, Ferreira, Baily & Ferraro. (2002). Good-Enough Representations in Language Comprehension. Current Directions in Psychological Science, Ferreira, Christianson & Hollingworth. (2001). Misinterpretations of Garden-Path Sentences: Implications for Models of Sentence Processing and Reanalysis. Journal of Psycholinguistic Research 30, Gahl, Jurafsky & Roland. (2004). Verb subcategorization frequencies: American English corpus data, methodological studies, and crosscorpus comparisons. Behavior Research Methods, Instruments, & Computers 36(3), Gibson & Pearlmutter. (2000). Distinguishing Serial and Parallel Parsing. Journal of Psycholinguistic Research 29(2), Kazanina, Lau, Lieberman, Yoshida & Phillips. (2007). The effect of syntactic constraints on the processing of backwards anaphora. Journal of Memory and Language 56(3), Konieczny, Weldle, Wolfer, Baumann, & Müller. (2010). Anaphors and Local Coherences. Talk presented at the 32nd Annual Conference of the Cognitive Science Society. August 2010, Portland, USA. Kwon & Sturt. (2014). The use of control information in dependency formation: An eye-tracking study. Journal of Memory and Language 73, Pearlmutter & Mendelsohn. (1999). Serial versus Parallel Sentence Comprehension. Unpublished manuscript. Slattery, Sturt, Christianson, Yoshida & Ferreira. (2013). Lingering misinterpretations of garden path sentences arise from competing syntactic representations. JML 69, Sturt, Sanford, Stewart & Dawydiak. (2004). Linguistic focus and good-enough representations: An application of the change-detection paradigm. Psychonomic Bulletin & Review 11(5), Tabor & Hutchins. (2004). Evidence for Self-Organized Sentence Processing: Digging-In Effects. Journal of Experimental Psychology 30(2), Tabor, Galantucci & Richardson. (2004). Effects of merely local syntactic coherence on sentence processing. JML 50, van Gompel & Liversedge. (2003). The influence of morphological information on cataphoric pronoun assignment. Journal of Experimental Psychology 29(1), 128.

54 Offline Forced-Coreference Acceptability

55 Effect of Punctuation on Transitivity Bias

56 Forced-Coreference Pilot for Eye-tracking

controlling for confounds in online measures of sentence complexity

controlling for confounds in online measures of sentence complexity Marten van Schijndel 1 July 28, 2015 1 Department of Linguistics, The Ohio State University van Schijndel Confounds in complexity July