revised CUNY abstract
any last minute comments appreciated. greg? are you awake?
Sunday, December 9, 2007
Boston, Chen, & Hale 4
Thanks to Paul and everyone for your great comments! The latest version of the paper is up, and submitted.
Boston Chen and Hale Version 4
Boston Chen and Hale Version 4
Saturday, December 8, 2007
Patil, Bartek & Lewis
Thanks in advance for your comments.
The Interplay of Locality and Surprisal
Umesh Patil, Brian Bartek and Richard Lewis
Dependency Locality Theory (DLT) (Gibson 1998, 2000) predicts increasing processing difficulty at the verb as argument-verb distance is increased. However, recently Levy (2007), using surprisal (Hale 2001), has argued that increasing argument-head distance can sharpen the expectation for the verb, resulting in a facilitation at the verb -- an antilocality effect. We show that, although in general DLT and surprisal appear to make conflicting predictions, for certain constructions this is not necessarily the case.
An example is Grodner and Gibson's (2005) experiment (see 1.a-c). They varied argument-head distance in object-extracted RCs with an unmodified subject (1-a), a PP modified subject (1-b), and an RC-modified subject (1-c). The DLT predicts monotonically increasing processing difficulty at the first verb (supervised) in (1-a) to (1-c) due to the increase in predicted integration cost. By contrast, surprisal predicts that the increase in the number of preverbal constituents facilitates the expectation of upcoming verb. Using a self-paced reading study Grodner and Gibson provide evidence that supports DLT's predictions.
We demonstrate that surprisal's antilocality prediction can be seen in these materials after DLT's effect on reading times is partialled out. Bartek et al (2006) replicated Grodner and Gibson's experiment using eyetracking, and also carried out a second experiment with the same design but using short (4-6 words), high-frequency words (>50-per-million occurrences).
For these two experiments, we fitted a linear mixed-effects model with fixation durations at the verb 'supervised' as dependent measure and DLT's predictions as a fixed effect (sentences and participants were the random effects), and examined the residuals, which constitute the variance unexplained by DLT. Interestingly, we found that, in the Grodner and Gibson replication the mean residual values showed a pattern consistent with surprisal's predictions (monotonically decreasing reading times in 1-a to 1-c at the verb); this was principally seen in first-pass measures such as single fixation duration, first fixation duration, and first pass reading time. In the second experiment, which had the same design but short, high-frequency words, the antilocality effect was evident in the residuals only for single fixation duration.
In sum, we show that surprisal's predictions can be seen even in data that apparently show locality effects as predicted by DLT. Moreover, surprisal's predictions are seen largely in early eyetracking measures; this is consistent with the fact that the surprisal value depends on the satisfaction of the expectation for the upcoming verb, which occurs immediately after lexical access.
(1)
a. The administrator who the nurse supervised scolded the medic while ...
b. The administrator who the nurse from the clinic supervised scolded the medic while ...
c. The administrator who the nurse who was from the clinic supervised scolded the medic while . . .
References:
1. Bartek, B., Smith, M., Lewis, R., and Vasishth, S (1998). Linking working memory processes with self-paced reading and eyetracking measures: How lexical processing load and SPR exaggerate locality effects. In Proceedings of the CUNY sentence processing conference, La Jolla, CA.
2. Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1-76.
3. Gibson, E. (2000). Dependency locality theory: A distance-based theory of linguistic complexity. In A. Marantz, Y. Miyashita, & W. O’Neil (Eds.). Image, language, brain: Papers from the first mind articulation project symposium. Cambridge, MA: MIT Press.
4. Grodner, D. and Gibson, E. (2005). Some consequences of the serial nature of linguistic
input. Cognitive Science, 29(2):261–290.
5. Hale, J. (2001). A probabilistic Earley parser as a psycholinguistic model. In Proceedings of
NAACL, volume 2, pages 159–166.
6. Levy, R. (2007). Expectation-based syntactic comprehension. In press, Cognition.
The Interplay of Locality and Surprisal
Umesh Patil, Brian Bartek and Richard Lewis
Dependency Locality Theory (DLT) (Gibson 1998, 2000) predicts increasing processing difficulty at the verb as argument-verb distance is increased. However, recently Levy (2007), using surprisal (Hale 2001), has argued that increasing argument-head distance can sharpen the expectation for the verb, resulting in a facilitation at the verb -- an antilocality effect. We show that, although in general DLT and surprisal appear to make conflicting predictions, for certain constructions this is not necessarily the case.
An example is Grodner and Gibson's (2005) experiment (see 1.a-c). They varied argument-head distance in object-extracted RCs with an unmodified subject (1-a), a PP modified subject (1-b), and an RC-modified subject (1-c). The DLT predicts monotonically increasing processing difficulty at the first verb (supervised) in (1-a) to (1-c) due to the increase in predicted integration cost. By contrast, surprisal predicts that the increase in the number of preverbal constituents facilitates the expectation of upcoming verb. Using a self-paced reading study Grodner and Gibson provide evidence that supports DLT's predictions.
We demonstrate that surprisal's antilocality prediction can be seen in these materials after DLT's effect on reading times is partialled out. Bartek et al (2006) replicated Grodner and Gibson's experiment using eyetracking, and also carried out a second experiment with the same design but using short (4-6 words), high-frequency words (>50-per-million occurrences).
For these two experiments, we fitted a linear mixed-effects model with fixation durations at the verb 'supervised' as dependent measure and DLT's predictions as a fixed effect (sentences and participants were the random effects), and examined the residuals, which constitute the variance unexplained by DLT. Interestingly, we found that, in the Grodner and Gibson replication the mean residual values showed a pattern consistent with surprisal's predictions (monotonically decreasing reading times in 1-a to 1-c at the verb); this was principally seen in first-pass measures such as single fixation duration, first fixation duration, and first pass reading time. In the second experiment, which had the same design but short, high-frequency words, the antilocality effect was evident in the residuals only for single fixation duration.
In sum, we show that surprisal's predictions can be seen even in data that apparently show locality effects as predicted by DLT. Moreover, surprisal's predictions are seen largely in early eyetracking measures; this is consistent with the fact that the surprisal value depends on the satisfaction of the expectation for the upcoming verb, which occurs immediately after lexical access.
(1)
a. The administrator who the nurse supervised scolded the medic while ...
b. The administrator who the nurse from the clinic supervised scolded the medic while ...
c. The administrator who the nurse who was from the clinic supervised scolded the medic while . . .
References:
1. Bartek, B., Smith, M., Lewis, R., and Vasishth, S (1998). Linking working memory processes with self-paced reading and eyetracking measures: How lexical processing load and SPR exaggerate locality effects. In Proceedings of the CUNY sentence processing conference, La Jolla, CA.
2. Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1-76.
3. Gibson, E. (2000). Dependency locality theory: A distance-based theory of linguistic complexity. In A. Marantz, Y. Miyashita, & W. O’Neil (Eds.). Image, language, brain: Papers from the first mind articulation project symposium. Cambridge, MA: MIT Press.
4. Grodner, D. and Gibson, E. (2005). Some consequences of the serial nature of linguistic
input. Cognitive Science, 29(2):261–290.
5. Hale, J. (2001). A probabilistic Earley parser as a psycholinguistic model. In Proceedings of
NAACL, volume 2, pages 159–166.
6. Levy, R. (2007). Expectation-based syntactic comprehension. In press, Cognition.
Boston Chen & Hale 3
Thanks Kyle and Titus for your comments. I re-worked the second and third paragraphs, and used Titus' suggestions for the phrasing. What do you all think?
Boston, Chen, & Hale version 3
Comments are very much appreciated!
Boston, Chen, & Hale version 3
Comments are very much appreciated!
Friday, December 7, 2007
Gerth, beim Graben & Vasishth (Version 4)
Thanks for the great comments!
edit: typo corrected
edit: changed Haegeman reference to Chomsky(1981)
Tensor product models for language-related brain potentials
Sabrina Gerth (1), Peter beim Graben (2), & Shravan Vasishth (1)
(1) Institute for Linguistics, University of Potsdam, Germany
(2) School of Psychology and Clinical Language Sciences, University of Reading, UK
A central question in computational psycholinguistics is how symbolic processing capabilities such as language are realized by neurodynamics in the human brain. An important online measure of language processing is event-related brain potentials (ERPs). However, ERPs are currently interpreted purely phenomenologically, and few computational accounts exist (see Hagoort 2005 for a recent proposal). We use tensor product representations of symbolic structures (Smolensky 1990; Smolensky and Legendre 2006; Mizraji 1989, 1992) to model syntactic parsing as nonlinear dynamics in neural activation space, and argue that ERP differences are due to trajectories which explore different regions in that space.
In an ERP experiment on the processing of German subject-object ambiguities, the sentence (1b) evoked a P600 ERP
compared to (1a) reflecting an initial garden path interpretation. Starting (for simplicity) with a Government and Binding formulation (Chomsky 1981) of the sentences (1a,b), we construct a locally ambiguous context-free grammar from the phrase structure trees. This grammar is decomposed into its unambiguous parts representing the two alternative processing strategies, namely subject preference against object preference in order to construct two appropriate deterministic pushdown recognizers (Aho & Ullman 1972).
Using the tensor product representation, the syntactic categories of the disambiguated grammars are mapped onto linearly independent filler vectors, while positions in a labeled binary tree are given as a basis of three-dimensional space. In order to build a parallel processor (Lewis 1998) the two parses for the subject-object sentence (regular vs. garden path) and the other two for the object-subject sentence respectively were linearly superimposed in activation space. Then, model ERPs are obtained as the first principal component.
Our model is able to describe, at least qualitatively, the obtained ERP results by trajectories that explore functionally and causally different regions in activation space while pursuing different language processing strategies. During its transient evolution, the trajectories of the model diverged exactly when the garden path was encountered which shows remarkable resemblance with the P600 effect in the ERP.
(1a) Die Rednerin hat den Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][SUB] at the congress sought
'The speaker has sought the advisor at the congress'
(1b) Die Rednerin hat der Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][OBJ] at the congress sought
'The speaker has been sought by the advisor'
Chomsky N. (1981) Lectures on Government and Binding: The Pisa Lectures. Mouton de Gruyter.
Hagoort P. (2005) On Broca, brain, and binding: a new framework. Trends in Cognitive Science 9(9):416 - 423.
Lewis RL (1998) Reanalysis and limited repair parsing: Leaping off the garden path. In: Fodor and Ferreira (1998), pp 147 - 285.
Mizraji E. (1989) Context-dependent association in linear distributed memories. Bulletin of Mathematical
Biology 51(2):195 - 205.
Smolensky P. (1990) Tensor product variable binding and the representation of symbolic structures in connectionist systems. Artificial Intelligence 46:159 - 216.
edit: typo corrected
edit: changed Haegeman reference to Chomsky(1981)
Tensor product models for language-related brain potentials
Sabrina Gerth (1), Peter beim Graben (2), & Shravan Vasishth (1)
(1) Institute for Linguistics, University of Potsdam, Germany
(2) School of Psychology and Clinical Language Sciences, University of Reading, UK
A central question in computational psycholinguistics is how symbolic processing capabilities such as language are realized by neurodynamics in the human brain. An important online measure of language processing is event-related brain potentials (ERPs). However, ERPs are currently interpreted purely phenomenologically, and few computational accounts exist (see Hagoort 2005 for a recent proposal). We use tensor product representations of symbolic structures (Smolensky 1990; Smolensky and Legendre 2006; Mizraji 1989, 1992) to model syntactic parsing as nonlinear dynamics in neural activation space, and argue that ERP differences are due to trajectories which explore different regions in that space.
In an ERP experiment on the processing of German subject-object ambiguities, the sentence (1b) evoked a P600 ERP
compared to (1a) reflecting an initial garden path interpretation. Starting (for simplicity) with a Government and Binding formulation (Chomsky 1981) of the sentences (1a,b), we construct a locally ambiguous context-free grammar from the phrase structure trees. This grammar is decomposed into its unambiguous parts representing the two alternative processing strategies, namely subject preference against object preference in order to construct two appropriate deterministic pushdown recognizers (Aho & Ullman 1972).
Using the tensor product representation, the syntactic categories of the disambiguated grammars are mapped onto linearly independent filler vectors, while positions in a labeled binary tree are given as a basis of three-dimensional space. In order to build a parallel processor (Lewis 1998) the two parses for the subject-object sentence (regular vs. garden path) and the other two for the object-subject sentence respectively were linearly superimposed in activation space. Then, model ERPs are obtained as the first principal component.
Our model is able to describe, at least qualitatively, the obtained ERP results by trajectories that explore functionally and causally different regions in activation space while pursuing different language processing strategies. During its transient evolution, the trajectories of the model diverged exactly when the garden path was encountered which shows remarkable resemblance with the P600 effect in the ERP.
(1a) Die Rednerin hat den Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][SUB] at the congress sought
'The speaker has sought the advisor at the congress'
(1b) Die Rednerin hat der Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][OBJ] at the congress sought
'The speaker has been sought by the advisor'
Chomsky N. (1981) Lectures on Government and Binding: The Pisa Lectures. Mouton de Gruyter.
Hagoort P. (2005) On Broca, brain, and binding: a new framework. Trends in Cognitive Science 9(9):416 - 423.
Lewis RL (1998) Reanalysis and limited repair parsing: Leaping off the garden path. In: Fodor and Ferreira (1998), pp 147 - 285.
Mizraji E. (1989) Context-dependent association in linear distributed memories. Bulletin of Mathematical
Biology 51(2):195 - 205.
Smolensky P. (1990) Tensor product variable binding and the representation of symbolic structures in connectionist systems. Artificial Intelligence 46:159 - 216.
beim Graben, Gerth & Vasishth (Version 2)
Modeling language-related brain potentials by dynamical recognizers
Peter beim Graben (1), Sabrina Gerth (2) & Shravan Vasishth (2)
(1) School of Psychology and Clinical Language Sciences,
University of Reading, United Kingdom
(2) Institute for Linguistics,
University of Potsdam, Germany
Event-related brain potentials (ERP) are important neural correlates of cognitive processes. In the domain of
language processing, the N400 reflects lexical-semantic integration processes. Every incoming new word is
known to evoke a "bump" in the ERP wave whose amplitude co-varies with the ease of integration (Coles and
Rugg 1995).
Starting from experimental findings of Osterhout, Holcomb and Swinney (1994) on this issue, we propose a
computational model that is able to account for the N400 word integration effect. Our models combine two
different approaches from dynamical system theory, 1. Nonlinear Dynamical Automata (NDA: Moore 1990, beim
Graben et al. 2004), implementing Turing machines by autonomous nonlinear dynamics in phase space, and 2.
Dynamical Recognizer (DR: Pollack 1991; see also Tabor 1998) where symbols act non-autonomously upon phase
space, represented by iterated function systems.
In a first step, we describe the sentence material of Osterhout, Holcomb and Swinney (1994) by an appropriate
context-free grammar, and construct a deterministic top-down recognizer for processing this grammar in a
second step. The parser is canonically represented by a generalized shift and subsequently mapped onto a
nonlinear dynamics on the unit square by means of a Gödel encoding (Moore 1990) in steps three and four.
Instead of incorporating the complete input tape of the automaton into the model, we do this only for the
fist two symbols of the input in order to simulate a "working memory" (Frazier and Fodor 1978). After each
attachment, this memory contains only one symbol in the most significant position such that a new word is
scanned from the environment into the second-most significant position (Wegner 1998). In our combined model,
the action of the scanned word upon the phase space of the NDA is represented by one particular function from
an DR's iterated function system.
Symbolically meaningful states of the combined NDA/DR are rectangular macrostates moving through the NDA's
phase space. As an model ERP measure we propose the area of these rectangles. As each scan operation is
reflected by a vertical squeezing of a macrostate, the measured ERP signal drops after each scanning, thus
emulating the word integration N400.
Our model is consistent with the dynamical system interpretation of ERPs (Basar 1980) and accounts for the
causal efficacy of dynamical systems models of symbolic computation (Fodor and Pylyshyn 1988) since different
regions in phase space are functionally different with respect to language processing.
References
Basar E (1980) EEG-Brain Dynamics. Relations between EEG and Brain Evoked Potentials. Elsevier/North Holland
Biomedical Press, Amsterdam.
Coles MGH, Rugg MD (1995) Event-related brain potentials: an introduction. In: Coles MGH, Rugg MD (eds)
Electrophysiology of Mind: Event- Related Brain Potentials and Cognition, Oxford University Press, Oxford,
chap 1.
Fodor J, Pylyshyn ZW (1988) Connectionism and cognitive architecture: A critical analysis. Cognition 28:3 –
71.
Frazier L, Fodor JD (1978) The sausage machine: A new two-stage parsing model. Cognition 6:291 – 326.
beim Graben P, Jurish B, Saddy D, Frisch S (2004) Language processing by dynamical systems. International
Journal of Bifurcation and Chaos 14(2):599 – 621.
Moore C (1990) Unpredictability and undecidability in dynamical systems. Physical Review Letters 64(20):2354
– 2357.
Osterhout L, Holcomb PJ, Swinney DA (1994) Brain potentials elicited by garden-path sentences: Evidence of
the application of verb information during parsing. Journal of Experimental Psychology: Learning, Memory, and
Cognition 20(4):786 – 803.
Pollack JB (1991) The induction of dynamical recognizers. Machine Learning 7:227 – 252, also published in
Port and van Gelder (1995), pp. 283 – 312.
Tabor W (1998) Dynamical automata. Technical Report TR98-1694, Cornell Computer Science Department,
Department of Psychology, Uris Hall, Cornell University, Ithaca, NY 14853.
Wegner P (1998) Interactive foundations of computing. Theoretical Computer Science 192:315 – 351.
Peter beim Graben (1), Sabrina Gerth (2) & Shravan Vasishth (2)
(1) School of Psychology and Clinical Language Sciences,
University of Reading, United Kingdom
(2) Institute for Linguistics,
University of Potsdam, Germany
Event-related brain potentials (ERP) are important neural correlates of cognitive processes. In the domain of
language processing, the N400 reflects lexical-semantic integration processes. Every incoming new word is
known to evoke a "bump" in the ERP wave whose amplitude co-varies with the ease of integration (Coles and
Rugg 1995).
Starting from experimental findings of Osterhout, Holcomb and Swinney (1994) on this issue, we propose a
computational model that is able to account for the N400 word integration effect. Our models combine two
different approaches from dynamical system theory, 1. Nonlinear Dynamical Automata (NDA: Moore 1990, beim
Graben et al. 2004), implementing Turing machines by autonomous nonlinear dynamics in phase space, and 2.
Dynamical Recognizer (DR: Pollack 1991; see also Tabor 1998) where symbols act non-autonomously upon phase
space, represented by iterated function systems.
In a first step, we describe the sentence material of Osterhout, Holcomb and Swinney (1994) by an appropriate
context-free grammar, and construct a deterministic top-down recognizer for processing this grammar in a
second step. The parser is canonically represented by a generalized shift and subsequently mapped onto a
nonlinear dynamics on the unit square by means of a Gödel encoding (Moore 1990) in steps three and four.
Instead of incorporating the complete input tape of the automaton into the model, we do this only for the
fist two symbols of the input in order to simulate a "working memory" (Frazier and Fodor 1978). After each
attachment, this memory contains only one symbol in the most significant position such that a new word is
scanned from the environment into the second-most significant position (Wegner 1998). In our combined model,
the action of the scanned word upon the phase space of the NDA is represented by one particular function from
an DR's iterated function system.
Symbolically meaningful states of the combined NDA/DR are rectangular macrostates moving through the NDA's
phase space. As an model ERP measure we propose the area of these rectangles. As each scan operation is
reflected by a vertical squeezing of a macrostate, the measured ERP signal drops after each scanning, thus
emulating the word integration N400.
Our model is consistent with the dynamical system interpretation of ERPs (Basar 1980) and accounts for the
causal efficacy of dynamical systems models of symbolic computation (Fodor and Pylyshyn 1988) since different
regions in phase space are functionally different with respect to language processing.
References
Basar E (1980) EEG-Brain Dynamics. Relations between EEG and Brain Evoked Potentials. Elsevier/North Holland
Biomedical Press, Amsterdam.
Coles MGH, Rugg MD (1995) Event-related brain potentials: an introduction. In: Coles MGH, Rugg MD (eds)
Electrophysiology of Mind: Event- Related Brain Potentials and Cognition, Oxford University Press, Oxford,
chap 1.
Fodor J, Pylyshyn ZW (1988) Connectionism and cognitive architecture: A critical analysis. Cognition 28:3 –
71.
Frazier L, Fodor JD (1978) The sausage machine: A new two-stage parsing model. Cognition 6:291 – 326.
beim Graben P, Jurish B, Saddy D, Frisch S (2004) Language processing by dynamical systems. International
Journal of Bifurcation and Chaos 14(2):599 – 621.
Moore C (1990) Unpredictability and undecidability in dynamical systems. Physical Review Letters 64(20):2354
– 2357.
Osterhout L, Holcomb PJ, Swinney DA (1994) Brain potentials elicited by garden-path sentences: Evidence of
the application of verb information during parsing. Journal of Experimental Psychology: Learning, Memory, and
Cognition 20(4):786 – 803.
Pollack JB (1991) The induction of dynamical recognizers. Machine Learning 7:227 – 252, also published in
Port and van Gelder (1995), pp. 283 – 312.
Tabor W (1998) Dynamical automata. Technical Report TR98-1694, Cornell Computer Science Department,
Department of Psychology, Uris Hall, Cornell University, Ithaca, NY 14853.
Wegner P (1998) Interactive foundations of computing. Theoretical Computer Science 192:315 – 351.
Thursday, December 6, 2007
trans-rochester team abstract
Hi everyone here is my first draft of what I'm calling the "trans-Rochester team" abstract, even though Greg never really was in Rochester.
http://comp.lin.msu.edu/cuny08/ellipsisCUNY08.pdf
have at it!
http://comp.lin.msu.edu/cuny08/ellipsisCUNY08.pdf
have at it!
Boston Chen & Hale
Here's a link to our abstract, "A cross-linguistic analysis of garden-pathing in a statistical dependency parser":
BostonChenHale.pdf
And here's the .tex file if anyone's interested:
BostonChenHale.tex
We appreciate all comments and input!
Marisa
BostonChenHale.pdf
And here's the .tex file if anyone's interested:
BostonChenHale.tex
We appreciate all comments and input!
Marisa
transatlantic team abstract
Here is my draft abstract for the transatlantic team involving Shravan, Marisa, Umesh, Reinhold and myself.
http://comp.lin.msu.edu/cuny08/hale.pdf
It takes advantage of Peter Gordon's openness to diagrams.
http://comp.lin.msu.edu/cuny08/hale.pdf
It takes advantage of Peter Gordon's openness to diagrams.
diagrams are ok as long as you don't go crazy
Hi everyone, I got this response back from Peter Gordon re: the language disallowing diagrams in CUNY submissions:
On Dec 5, 2007, at 9:41 PM, Peter C. Gordon wrote:
John
I have reviewed the issues with last year's organizers (who used the same instructions, which were inherited from before, which were handed down from ...). The problem is not in the reviewing but in compiling a program that has at least some level of typographical coherence. PDFs make that hard. But, I think that the scientific issues win out and that you should submit PDFs with diagrams if you believe that is the best way to convey the work. Reviewers will be encouraged to use the PDFs rather than the text (reviewers frequently ignore instructions).
Please try to keep the PDFs as uncrowded as possible (no tiny fonts, no extremely clever organization where every space is adjusted to the exact half point distance that makes it work) because your PDF will be converted into text and then reformatted to a common style for printing if it is included in the program. We will try to make the abstracts appear as they should but we will not kill ourselves over it. In putting the program together we will follow the practice stated by one of last year's organizers "I simply made executive/editorial decisions about how much I was willing to deal with any violations of the rules."
Thanks for checking, we look forward to getting your abstract(s).
Peter
On Dec 5, 2007, at 9:41 PM, Peter C. Gordon wrote:
John
I have reviewed the issues with last year's organizers (who used the same instructions, which were inherited from before, which were handed down from ...). The problem is not in the reviewing but in compiling a program that has at least some level of typographical coherence. PDFs make that hard. But, I think that the scientific issues win out and that you should submit PDFs with diagrams if you believe that is the best way to convey the work. Reviewers will be encouraged to use the PDFs rather than the text (reviewers frequently ignore instructions).
Please try to keep the PDFs as uncrowded as possible (no tiny fonts, no extremely clever organization where every space is adjusted to the exact half point distance that makes it work) because your PDF will be converted into text and then reformatted to a common style for printing if it is included in the program. We will try to make the abstracts appear as they should but we will not kill ourselves over it. In putting the program together we will follow the practice stated by one of last year's organizers "I simply made executive/editorial decisions about how much I was willing to deal with any violations of the rules."
Thanks for checking, we look forward to getting your abstract(s).
Peter
The upshot sounds so me like, we can do diagrams just keep in mind some of the organizers and reviewers don't know how to handle them.
von der Malsburg and Vasishth Version 2
## shortened, new title
A New Method for Analyzing Eye Movements in Reading that Is Sensitive to Fixation Patterns and Fixation Durations
Titus von der Malsburg and Shravan Vasishth
In the investigation of reading, word-based measures such as gaze duration and regression probability, are commonly used for analyzing empirical data. However, these measures are not always adequate when the effect of an experimental manipulation is expected to appear in patterns of fixations rather than in local measures. One example is the study of reanalysis strategies in temporarily ambiguous sentences. We address this shortcoming by proposing alternative methods for analyzing eye movements that are based on the notion of similarity of scanpaths. For this purpose we introduce a measure that quantifies the similarity of two scanpaths with respect to their fixation patterns and their temporal dynamics.
Our metric is a member of the class of edit-distances of which the Levenshtein metric is an instance (Levenshtein 1968). Edit-distances have already been used for analysing eye tracking data in a variety of studies outside reading research, for example in experiments studying scene perception or the usability of graphical user interfaces [citation]. While edit-distances capture the topological similarities of scanpaths, they operate on a representation of scanpaths that neglects two crucial properties: The first is fixation duration, which is one of the most important and best-studied sources of evidence in online reading research. The second is the spatial distance of regions of interest. The observation of parafoveal preview effects, however, suggest that the spatial distance does play an important role and should be accounted for [cite something that demonstrates preview effects].
We addressed these deficits in standard edit-distance measures by replacing the cost function of the Levenshtein metric with a function that compares fixations with respect to their duration and distance. The way we incorporate distance is motivated by findings about acuity in human visual perception (Rovamo, Virsu, Näsänen, 1978). The resulting measure operates on a parsimonious representation of eye movement data and can efficiently be computed (O(m*n), where n, m are the numbers of fixations of the two compared scanpaths).
Possible approaches to analysing data based on this measure are: 1.) Cluster analysis of sets of scanpaths in order to identify reading strategies, 2.) analysis of variance of the pairwise similarities in sets of scanpaths for identifying different populations, 3.) application of string interpolation methods and optimization techniques for computing prototypical scanpaths for sets of scanpaths.
Areas of investigation where these methods might also be useful comprise the evaluation of models of oculo-motor control, classification of reading disorders, analysis of cognitive strategies in experiments using the visual world paradigm.
As opposed to approaches using word-based measures, the proposed similarity-based methods enable the analysis of the full available data and do not necessitate averaging of variables thereby preserving possibly crucial information. The main benefit of our similarity measure is that it is sensitive to fixation patterns and the temporal dynamics in a scanpaths at the same time.
A New Method for Analyzing Eye Movements in Reading that Is Sensitive to Fixation Patterns and Fixation Durations
Titus von der Malsburg and Shravan Vasishth
In the investigation of reading, word-based measures such as gaze duration and regression probability, are commonly used for analyzing empirical data. However, these measures are not always adequate when the effect of an experimental manipulation is expected to appear in patterns of fixations rather than in local measures. One example is the study of reanalysis strategies in temporarily ambiguous sentences. We address this shortcoming by proposing alternative methods for analyzing eye movements that are based on the notion of similarity of scanpaths. For this purpose we introduce a measure that quantifies the similarity of two scanpaths with respect to their fixation patterns and their temporal dynamics.
Our metric is a member of the class of edit-distances of which the Levenshtein metric is an instance (Levenshtein 1968). Edit-distances have already been used for analysing eye tracking data in a variety of studies outside reading research, for example in experiments studying scene perception or the usability of graphical user interfaces [citation]. While edit-distances capture the topological similarities of scanpaths, they operate on a representation of scanpaths that neglects two crucial properties: The first is fixation duration, which is one of the most important and best-studied sources of evidence in online reading research. The second is the spatial distance of regions of interest. The observation of parafoveal preview effects, however, suggest that the spatial distance does play an important role and should be accounted for [cite something that demonstrates preview effects].
We addressed these deficits in standard edit-distance measures by replacing the cost function of the Levenshtein metric with a function that compares fixations with respect to their duration and distance. The way we incorporate distance is motivated by findings about acuity in human visual perception (Rovamo, Virsu, Näsänen, 1978). The resulting measure operates on a parsimonious representation of eye movement data and can efficiently be computed (O(m*n), where n, m are the numbers of fixations of the two compared scanpaths).
Possible approaches to analysing data based on this measure are: 1.) Cluster analysis of sets of scanpaths in order to identify reading strategies, 2.) analysis of variance of the pairwise similarities in sets of scanpaths for identifying different populations, 3.) application of string interpolation methods and optimization techniques for computing prototypical scanpaths for sets of scanpaths.
Areas of investigation where these methods might also be useful comprise the evaluation of models of oculo-motor control, classification of reading disorders, analysis of cognitive strategies in experiments using the visual world paradigm.
As opposed to approaches using word-based measures, the proposed similarity-based methods enable the analysis of the full available data and do not necessitate averaging of variables thereby preserving possibly crucial information. The main benefit of our similarity measure is that it is sensitive to fixation patterns and the temporal dynamics in a scanpaths at the same time.
Engelmann, Kuo and Vasishth
The Chinese subject-object relative asymmetry:
A computational, experimental and corpus investigation
Felix Engelmann, Kueilan Kuo and Shravan Vasishth
Cross-linguistically, subject relative clauses tend to be easier to
process than object relatives. Recently, Hsiao and Gibson (2003)
presented a surprising result that was subsequently replicated by Lin and
Garnsey (2007): in Chinese, subject relatives like (1a) are harder to
process than object relatives (1b).
This result is striking because it dramatically shrinks the space of
theoretical explanations for relative-clause processing. Theories that
rely on word-order regularities (Christiansen and MacDonald 2002,
Bever 1970) can predict the Hsiao and Gibson finding correctly: word
order in object relatives is SVO (the canonical order for Chinese) and
in object relatives it is VOS. By contrast, other theories fail to
explain the Chinese facts. For example, (i) Keenan and Comrie (1977)
rely on the greater accessibility of subjects compared to objects,
(ii) MacWhinney (1977) invokes the need to shift perspective from
subject to object in object relatives, which is harder in object
relatives than subject, (iii) Mitchell (1995) argues that structural
frequency of syntactic frames determines parsing difficulty, and since
subject relatives are more frequent than object, they should be easier
to process.
We make two contributions to this interesting finding by Hsiao and
Gibson.
First, we examine the simple recurrent network (SRN) model, which has
been cited in the context of the Chinese results (Hsiao and Gibson
2003: 15) as an instantiation of the word order theory. The
predictions of an SRN such as Christiansen and MacDonald's have never
been compared to reading times seen in Chinese relative clauses -- it
is merely assumed that the SRN makes the correct predictions for the
results Hsiao and Gibson found. We present the predictions of an SRN
for Chinese relative clause processing, and show that the predictions
are rather mixed and do not point to clear support for an SRN-based
explanation. In our simulations, we ensured that the SRN displays
nearly-identical behavior to published connectionist simulations
involving Dutch and German center embeddings and ungrammatical English
center embeddings -- the model architecture is thus comparable to
previous published models. The SRN was trained on a simplified
grammar of Chinese based on structural frequencies in the Sinica
Corpus 3.0 (5 million words). On the region preceding the relativizer
de the network predictions showed a tendency towards a subject
relative preference which is contrary to Hsiao and Gibson's
results. The other regions did not show any significant difference.
The difficulties faced by the connectionist model in predicting the
Chinese reading times cast doubt on the connectionist instantiation of
the word order regularity explanation.
Second, we repeated the Hsiao and Gibson study using their stimuli in
a self-paced reading study. We found, contrary to their results, that
Chinese in fact shows a subject preference; independent research on
Chinese relatives supports our findings (Lin and Bever 2006), (Su et
al 2007).
We suggest that the best explanation for the subject-object relative
clause asymmetry lies in the relative frequencies of syntactic frames:
subject relatives in Chinese are, as in many other languages, more
frequent than object relatives, and this frequency pattern predicts
the subject-relative preference correctly. Using the Sinica Corpus,
we present evidence supporting this view. In addition, we argue that
the frequency patterns of noun-phrases animacy (animates tend to occur
before inanimates more frequently) may exert an independent influence
on processing difficulty. Independent evidence consistent with this
idea is presented by Lin and Garnsey (2007).
In sum, based on a series of simulations and a self-paced reading
experiment, we argue that (a) the word-order explanation for relative
clauses as instantiated by connectionist models may be problematic,
and (b) theories that predict the subject-preference
cross-linguistically (structural-frequency, accessibility, and
perspective) may be more compelling.
1a. Subject relative
[GAP yaoqing fuhao de] guanyuan xinhuaibugui
invite tycoon gen official have bad intentions
'The official who invited the tycoon has bad intentions.'
1b. Object relative
[fuhao yaoqing GAP de] guanyuan xinhuaibugui
tycoon invite gen official have bad intentions
'The official who the tycoon invited has bad intentions.'
MacDonald,M. C., & Christiansen, M. H. (2002). Reassessing working
memory: A reply to Just
and Carpenter andWaters and
Caplan. Psychological Review, 109(1), 35-54.
Hsiao, F. P.-F., & Gibson, E. (2003). Processing relative clauses in
Chinese. Cognition, 90, 3-27.
Lin, Y. B., & Garnsey, S.M. (2007). Plausibility and the resolution of
temporary ambiguity in relativeclause comprehension in mandarin. In
Proceedings of the CUNY Sentence Processing Conference.
A computational, experimental and corpus investigation
Felix Engelmann, Kueilan Kuo and Shravan Vasishth
Cross-linguistically, subject relative clauses tend to be easier to
process than object relatives. Recently, Hsiao and Gibson (2003)
presented a surprising result that was subsequently replicated by Lin and
Garnsey (2007): in Chinese, subject relatives like (1a) are harder to
process than object relatives (1b).
This result is striking because it dramatically shrinks the space of
theoretical explanations for relative-clause processing. Theories that
rely on word-order regularities (Christiansen and MacDonald 2002,
Bever 1970) can predict the Hsiao and Gibson finding correctly: word
order in object relatives is SVO (the canonical order for Chinese) and
in object relatives it is VOS. By contrast, other theories fail to
explain the Chinese facts. For example, (i) Keenan and Comrie (1977)
rely on the greater accessibility of subjects compared to objects,
(ii) MacWhinney (1977) invokes the need to shift perspective from
subject to object in object relatives, which is harder in object
relatives than subject, (iii) Mitchell (1995) argues that structural
frequency of syntactic frames determines parsing difficulty, and since
subject relatives are more frequent than object, they should be easier
to process.
We make two contributions to this interesting finding by Hsiao and
Gibson.
First, we examine the simple recurrent network (SRN) model, which has
been cited in the context of the Chinese results (Hsiao and Gibson
2003: 15) as an instantiation of the word order theory. The
predictions of an SRN such as Christiansen and MacDonald's have never
been compared to reading times seen in Chinese relative clauses -- it
is merely assumed that the SRN makes the correct predictions for the
results Hsiao and Gibson found. We present the predictions of an SRN
for Chinese relative clause processing, and show that the predictions
are rather mixed and do not point to clear support for an SRN-based
explanation. In our simulations, we ensured that the SRN displays
nearly-identical behavior to published connectionist simulations
involving Dutch and German center embeddings and ungrammatical English
center embeddings -- the model architecture is thus comparable to
previous published models. The SRN was trained on a simplified
grammar of Chinese based on structural frequencies in the Sinica
Corpus 3.0 (5 million words). On the region preceding the relativizer
de the network predictions showed a tendency towards a subject
relative preference which is contrary to Hsiao and Gibson's
results. The other regions did not show any significant difference.
The difficulties faced by the connectionist model in predicting the
Chinese reading times cast doubt on the connectionist instantiation of
the word order regularity explanation.
Second, we repeated the Hsiao and Gibson study using their stimuli in
a self-paced reading study. We found, contrary to their results, that
Chinese in fact shows a subject preference; independent research on
Chinese relatives supports our findings (Lin and Bever 2006), (Su et
al 2007).
We suggest that the best explanation for the subject-object relative
clause asymmetry lies in the relative frequencies of syntactic frames:
subject relatives in Chinese are, as in many other languages, more
frequent than object relatives, and this frequency pattern predicts
the subject-relative preference correctly. Using the Sinica Corpus,
we present evidence supporting this view. In addition, we argue that
the frequency patterns of noun-phrases animacy (animates tend to occur
before inanimates more frequently) may exert an independent influence
on processing difficulty. Independent evidence consistent with this
idea is presented by Lin and Garnsey (2007).
In sum, based on a series of simulations and a self-paced reading
experiment, we argue that (a) the word-order explanation for relative
clauses as instantiated by connectionist models may be problematic,
and (b) theories that predict the subject-preference
cross-linguistically (structural-frequency, accessibility, and
perspective) may be more compelling.
1a. Subject relative
[GAP yaoqing fuhao de] guanyuan xinhuaibugui
invite tycoon gen official have bad intentions
'The official who invited the tycoon has bad intentions.'
1b. Object relative
[fuhao yaoqing GAP de] guanyuan xinhuaibugui
tycoon invite gen official have bad intentions
'The official who the tycoon invited has bad intentions.'
MacDonald,M. C., & Christiansen, M. H. (2002). Reassessing working
memory: A reply to Just
and Carpenter andWaters and
Caplan. Psychological Review, 109(1), 35-54.
Hsiao, F. P.-F., & Gibson, E. (2003). Processing relative clauses in
Chinese. Cognition, 90, 3-27.
Lin, Y. B., & Garnsey, S.M. (2007). Plausibility and the resolution of
temporary ambiguity in relativeclause comprehension in mandarin. In
Proceedings of the CUNY Sentence Processing Conference.
Wednesday, December 5, 2007
Gerth, beim Graben & Vasishth (Version 3)
edit: title changed
Tensor product models for language-related brain potentials
Sabrina Gerth (1), Peter beim Graben (2), & Shravan Vasishth (1)
(1) Institute for Linguistics, University of Potsdam, Germany
(2) School of Psychology and Clinical Language Sciences, University of Reading, UK
A central question in computational psycholinguistics is how symbolic processing capabilities such
as language are realized by neurodynamics in the human brain. An important online measure of language
processing is event-related brain potentials (ERPs). However, ERPs are currently interpreted purely phenomenologically,
and few computational accounts exist (see Hagoort 2005 for a recent proposal). We use tensor product representations of symbolic structures (Smolensky 1990; Smolensky and Legendre 2006; Mizraji 1989, 1992) to model syntactic parsing as nonlinear dynamics in neural activation space, and argue that ERP differences are due to trajectories which explore different regions in that space.
In an ERP experiment on the processing of German subject-object ambiguities, the sentence (1b) evoked a P600
ERP compared to (1a) reflecting an initial garden path interpretation. Starting (for simplicity) with a Government and Binding
formulation (Haegeman 1994) of the sentences (1a,b), we construct a locally ambiguous context-free grammar
from the phrase structure trees. This grammar is decomposed into its unambiguous parts representing the two
alternative processing strategies, namely subject preference against object preference in order to construct
two appropriate deterministic pushdown recognizers (Aho & Ullman 1972).
Using the tensor product representation, the syntactic categories of the disambiguated grammars are mapped
onto linearly independent filler vectors, while positions in a labeled binary tree are given as a basis of
three-dimensional space. In order to built a parallel processor (Lewis 1998) the two parses for the
subject-object sentence (regular vs. garden path) and the other two for the object-subject sentence
respectively were linearly superimposed in activation space. Then, model ERPs are obtained as the first
principal component.
Our model is able to describe, at least qualitatively, the obtained ERP results by trajectories that explore
functionally and causally different regions in activation space while pursuing different language processing
strategies. During its transient evolution, the trajectories of the model diverged exactly when the garden
path was encountered which shows remarkable resemblance with the P600 effect in the ERP.
(1a) Die Rednerin hat den Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][SUB] at the congress sought
'The speaker has sought the advisor at the congress'
(1b) Die Rednerin hat der Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][OBJ] at the congress sought
'The speaker has been sought by the advisor'
Haegeman L. (1994) Introduction to Government & Binding Theory, Blackwell Textbooks in Linguistics, vol1, 2nd
edn. Blackwell Publishers, Oxford, 1st edition 1991.
Hagoort P. (2005) On Broca, brain, and binding: a new framework. Trends in Cognitive Science 9(9):416 - 423.
Lewis RL (1998) Reanalysis and limited repair parsing: Leaping off the garden path. In: Fodor and Ferreira (1998), pp 147 - 285.
Mizraji E. (1989) Context-dependent association in linear distributed memories. Bulletin of Mathematical
Biology 51(2):195 - 205.
Smolensky P. (1990) Tensor product variable binding and the representation of symbolic structures in
Tensor product models for language-related brain potentials
Sabrina Gerth (1), Peter beim Graben (2), & Shravan Vasishth (1)
(1) Institute for Linguistics, University of Potsdam, Germany
(2) School of Psychology and Clinical Language Sciences, University of Reading, UK
A central question in computational psycholinguistics is how symbolic processing capabilities such
as language are realized by neurodynamics in the human brain. An important online measure of language
processing is event-related brain potentials (ERPs). However, ERPs are currently interpreted purely phenomenologically,
and few computational accounts exist (see Hagoort 2005 for a recent proposal). We use tensor product representations of symbolic structures (Smolensky 1990; Smolensky and Legendre 2006; Mizraji 1989, 1992) to model syntactic parsing as nonlinear dynamics in neural activation space, and argue that ERP differences are due to trajectories which explore different regions in that space.
In an ERP experiment on the processing of German subject-object ambiguities, the sentence (1b) evoked a P600
ERP compared to (1a) reflecting an initial garden path interpretation. Starting (for simplicity) with a Government and Binding
formulation (Haegeman 1994) of the sentences (1a,b), we construct a locally ambiguous context-free grammar
from the phrase structure trees. This grammar is decomposed into its unambiguous parts representing the two
alternative processing strategies, namely subject preference against object preference in order to construct
two appropriate deterministic pushdown recognizers (Aho & Ullman 1972).
Using the tensor product representation, the syntactic categories of the disambiguated grammars are mapped
onto linearly independent filler vectors, while positions in a labeled binary tree are given as a basis of
three-dimensional space. In order to built a parallel processor (Lewis 1998) the two parses for the
subject-object sentence (regular vs. garden path) and the other two for the object-subject sentence
respectively were linearly superimposed in activation space. Then, model ERPs are obtained as the first
principal component.
Our model is able to describe, at least qualitatively, the obtained ERP results by trajectories that explore
functionally and causally different regions in activation space while pursuing different language processing
strategies. During its transient evolution, the trajectories of the model diverged exactly when the garden
path was encountered which shows remarkable resemblance with the P600 effect in the ERP.
(1a) Die Rednerin hat den Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][SUB] at the congress sought
'The speaker has sought the advisor at the congress'
(1b) Die Rednerin hat der Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][OBJ] at the congress sought
'The speaker has been sought by the advisor'
Haegeman L. (1994) Introduction to Government & Binding Theory, Blackwell Textbooks in Linguistics, vol1, 2nd
edn. Blackwell Publishers, Oxford, 1st edition 1991.
Hagoort P. (2005) On Broca, brain, and binding: a new framework. Trends in Cognitive Science 9(9):416 - 423.
Lewis RL (1998) Reanalysis and limited repair parsing: Leaping off the garden path. In: Fodor and Ferreira (1998), pp 147 - 285.
Mizraji E. (1989) Context-dependent association in linear distributed memories. Bulletin of Mathematical
Biology 51(2):195 - 205.
Smolensky P. (1990) Tensor product variable binding and the representation of symbolic structures in
beim Graben, Gerth and Vasishth, Version 1
Peter beim Graben (1), Sabrina Gerth (2) & Shravan Vasishth (2)
(1) School of Psychology and Clinical Language Sciences,
University of Reading, United Kingdom
(2) Institute for Linguistics,
University of Potsdam, Germany
Event-related brain potentials (ERP) are important neural correlates of cognitive processes. In the domain of
language processing, the N400 reflects lexical-semantic integration processes. Every incoming new word is
known to evoke a "bump" in the ERP wave whose amplitude co-varies with the ease of integration (Coles and
Rugg 1995).
Starting from experimental findings of Osterhout, Holcomb and Swinney (1994) on this issue, we propose a
computational model that is able to account for the N400 word integration effect. Our models combines two
different approaches from dynamical system theory, 1. Nonlinear Dynamical Automata (NDA: Moore 1990, beim
Graben et al. 2004), implementing Turing machines by autonomous nonlinear dynamics in a phase space, and 2.
Dynamical Recognizer (DR: Pollack 1991; see also Tabor 1998) where symbols act non-autonomously upon a phase
space, represented by iterated function systems.
In a first step, we describe the sentence material of Osterhout, Holcomb and Swinney (1994) by a appropriate
context-free grammar, and construct a deterministic top-down recognizer for processing this grammar in a
second step. The parser is canonically represented by a generalized shift and subsequently mapped onto a
nonlinear dynamics on the unit square by means of a Gödel encoding (Moore 1990) in steps three and four.
Instead of incorporating the complete input tape of the automaton into the model, we do this only for the
fist two symbols of the input in order to emulate a "working memory" (Frazier and Fodor1978). After each
attachment, this memory contains only one symbol in the most significant position such that a new word is
scanned from the environment into the second-most significant position (Wegner 1998). In our combined model,
the action of the scanned word upon the phase space of the NDA is represented by one particular function from
an DR's iterated function systems.
Symbolically meaningful states of the combined NDA/DR are rectangular macrostates moving through the NDA's
phase space. As an model ERP measure we propose the volume of these rectangles. As each scan operation is
reflected by a vertical squeezing of a macrostate, the measured ERP signal drops after each scanning, thus
emulating the word integration N400.
Our model is consistent with the dynamical system interpretation of ERPs (Basar 1980) and accounts for the
causal efficacy of dynamical systems models of symbolic computation (Fodor and Pylyshyn 1988) since different
regions in phase space are functionally different with respect to language processing.
References
Basar E (1980) EEG-Brain Dynamics. Relations between EEG and Brain Evoked Potentials. Elsevier/North Holland
Biomedical Press, Amsterdam.
Coles MGH, Rugg MD (1995) Event-related brain potentials: an introduction. In: Coles MGH, Rugg MD (eds)
Electrophysiology of Mind: Event- Related Brain Potentials and Cognition, Oxford University Press, Oxford,
chap 1.
Fodor J, Pylyshyn ZW (1988) Connectionism and cognitive architecture: A critical analysis. Cognition 28:3 –
71.
Frazier L, Fodor JD (1978) The sausage machine: A new two-stage parsing model. Cognition 6:291 – 326.
beim Graben P, Jurish B, Saddy D, Frisch S (2004) Language processing by dynamical systems. International
Journal of Bifurcation and Chaos 14(2):599 – 621.
Moore C (1990) Unpredictability and undecidability in dynamical systems. Physical Review Letters 64(20):2354
– 2357.
Osterhout L, Holcomb PJ, Swinney DA (1994) Brain potentials elicited by garden-path sentences: Evidence of
the application of verb information during parsing. Journal of Experimental Psychology: Learning, Memory, and
Cognition 20(4):786 – 803.
Pollack JB (1991) The induction of dynamical recognizers. Machine Learning 7:227 – 252, also published in
Port and van Gelder (1995), pp. 283 – 312.
Tabor W (1998) Dynamical automata. Technical Report TR98-1694, Cornell Computer Science Department,
Department of Psychology, Uris Hall, Cornell University, Ithaca, NY 14853.
Wegner P (1998) Interactive foundations of computing. Theoretical Computer Science 192:315 – 351.
(1) School of Psychology and Clinical Language Sciences,
University of Reading, United Kingdom
(2) Institute for Linguistics,
University of Potsdam, Germany
Event-related brain potentials (ERP) are important neural correlates of cognitive processes. In the domain of
language processing, the N400 reflects lexical-semantic integration processes. Every incoming new word is
known to evoke a "bump" in the ERP wave whose amplitude co-varies with the ease of integration (Coles and
Rugg 1995).
Starting from experimental findings of Osterhout, Holcomb and Swinney (1994) on this issue, we propose a
computational model that is able to account for the N400 word integration effect. Our models combines two
different approaches from dynamical system theory, 1. Nonlinear Dynamical Automata (NDA: Moore 1990, beim
Graben et al. 2004), implementing Turing machines by autonomous nonlinear dynamics in a phase space, and 2.
Dynamical Recognizer (DR: Pollack 1991; see also Tabor 1998) where symbols act non-autonomously upon a phase
space, represented by iterated function systems.
In a first step, we describe the sentence material of Osterhout, Holcomb and Swinney (1994) by a appropriate
context-free grammar, and construct a deterministic top-down recognizer for processing this grammar in a
second step. The parser is canonically represented by a generalized shift and subsequently mapped onto a
nonlinear dynamics on the unit square by means of a Gödel encoding (Moore 1990) in steps three and four.
Instead of incorporating the complete input tape of the automaton into the model, we do this only for the
fist two symbols of the input in order to emulate a "working memory" (Frazier and Fodor1978). After each
attachment, this memory contains only one symbol in the most significant position such that a new word is
scanned from the environment into the second-most significant position (Wegner 1998). In our combined model,
the action of the scanned word upon the phase space of the NDA is represented by one particular function from
an DR's iterated function systems.
Symbolically meaningful states of the combined NDA/DR are rectangular macrostates moving through the NDA's
phase space. As an model ERP measure we propose the volume of these rectangles. As each scan operation is
reflected by a vertical squeezing of a macrostate, the measured ERP signal drops after each scanning, thus
emulating the word integration N400.
Our model is consistent with the dynamical system interpretation of ERPs (Basar 1980) and accounts for the
causal efficacy of dynamical systems models of symbolic computation (Fodor and Pylyshyn 1988) since different
regions in phase space are functionally different with respect to language processing.
References
Basar E (1980) EEG-Brain Dynamics. Relations between EEG and Brain Evoked Potentials. Elsevier/North Holland
Biomedical Press, Amsterdam.
Coles MGH, Rugg MD (1995) Event-related brain potentials: an introduction. In: Coles MGH, Rugg MD (eds)
Electrophysiology of Mind: Event- Related Brain Potentials and Cognition, Oxford University Press, Oxford,
chap 1.
Fodor J, Pylyshyn ZW (1988) Connectionism and cognitive architecture: A critical analysis. Cognition 28:3 –
71.
Frazier L, Fodor JD (1978) The sausage machine: A new two-stage parsing model. Cognition 6:291 – 326.
beim Graben P, Jurish B, Saddy D, Frisch S (2004) Language processing by dynamical systems. International
Journal of Bifurcation and Chaos 14(2):599 – 621.
Moore C (1990) Unpredictability and undecidability in dynamical systems. Physical Review Letters 64(20):2354
– 2357.
Osterhout L, Holcomb PJ, Swinney DA (1994) Brain potentials elicited by garden-path sentences: Evidence of
the application of verb information during parsing. Journal of Experimental Psychology: Learning, Memory, and
Cognition 20(4):786 – 803.
Pollack JB (1991) The induction of dynamical recognizers. Machine Learning 7:227 – 252, also published in
Port and van Gelder (1995), pp. 283 – 312.
Tabor W (1998) Dynamical automata. Technical Report TR98-1694, Cornell Computer Science Department,
Department of Psychology, Uris Hall, Cornell University, Ithaca, NY 14853.
Wegner P (1998) Interactive foundations of computing. Theoretical Computer Science 192:315 – 351.
Engelmann, Kuo and Vasishth (Version 2)
## Thanks Titus; those comments were very helpful.
The Chinese subject-object relative asymmetry
A computational and corpus investigation
Felix Engelmann, Kueilan Kuo and Shravan Vasishth
Cross-linguistically, subject relative clauses tend to be easier to
process than object relatives. However, Hsiao and Gibson (2003)
presented a surprising result that was recently replicated by Lin and
Garnsey (2007): in Chinese, subject relatives like (1a) are harder to
process than object relatives (1b).
This result is striking because it dramatically shrinks the space of
theoretical explanations for relative-clause processing. Theories that
rely on word-order regularities (Christiansen and MacDonald 2002,
Bever 1970) can predict the Hsiao and Gibson finding correctly: word
order in object relatives is SVO (the canonical order for Chinese) and
in object relatives it is VOS. By contrast, other theories fail to
explain the Chinese facts. For example, (i) Keenan and Comrie (1977)
rely on the greater accessibility of subjects compared to objects,
(ii) MacWhinney (1977) invokes the need to shift perspective from
subject to object in object relatives, which is harder in object
relatives than subject, (iii) Mitchell (1995) argues that structural
frequency of syntactic frames determines parsing difficulty, and since
subject relatives are more frequent than object, they should be easier
to process.
We make two contributions to this interesting finding by Hsiao and
Gibson.
First, we examine connectionist models, which have been cited (Hsiao
and Gibson 2003: 15) as an instantiation of the word order theory. The
predictions of a simple recurrent network model such as Christiansen
and MacDonald's have never been compared to reading times seen in
Chinese relative clauses -- it is merely assumed that the word-order
theories (like the connectionist model) make the correct predictions
for the results Hsiao and Gibson found. We present the predictions of
a simple recurrent network for Chinese relative clause processing, and
show that the predictions are rather mixed and do not point to clear
support for an SRN-based explanation. In our simulations, we ensured
that our SRN displays identical behavior to published connectionist
simulations involving Dutch and German center embeddings and
ungrammatical English center embeddings -- the model architecture is
thus identical to previous published models. The difficulties faced by
the connectionist model in predicting the Chinese reading times cast
doubt on this instantiation of the word order regularity explanation.
Second, we repeated the Hsiao and Gibson study using their stimuli in
a self-paced reading study. We found, contrary to their results, that
Chinese in fact shows a subject preference; independent research on
Chinese relatives supports our findings (Lin and Bever 2006), (Su et
al 2007).
We suggest that the best explanation for the subject-object relative
clause asymmetry lies in the relative frequencies of syntactic frames:
subject relatives in Chinese are, as in many other languages, more
frequent than object relatives, and this frequency-pattern predicts
the subject-relative preference correctly. Using Sinica Corpus (5
million words), we present evidence supporting this view. In addition,
we argue that the frequency patterns of noun-phrases animacy (animates
tend to occur before inanimates more frequently) may exert an
independent influence on processing difficulty. Independent evidence
consistent with this idea is presented by Lin and Garnsey (2007).
In sum, based on a series of simulations and a self-paced reading
experiment, we argue that (a) the word-order explanation for relative
clauses as instantiated by connectionist models may be problematic,
and (b) theories that predict the subject-preference
cross-linguistically (structural-frequency, accessibility, and
perspective) may be more compelling.
1a. Subject relative
[GAP yaoqing fuhao de] guanyuan xinhuaibugui
invite tycoon gen official have bad intentions
'The official who invited the tycoon has bad intentions.'
1b. Object relative
[fuhao yaoqing GAP de] guanyuan xinhuaibugui
tycoon invite gen official have bad intentions
'The official who the tycoon invited has bad intentions.'
MacDonald,M. C., & Christiansen, M. H. (2002). Reassessing working
memory: A reply to Just
and Carpenter andWaters and
Caplan. Psychological Review, 109(1), 35-54.
Hsiao, F. P.-F., & Gibson, E. (2003). Processing relative clauses in
Chinese. Cognition, 90, 3-27.
Lin, Y. B., & Garnsey, S.M. (2007). Plausibility and the resolution of
temporary ambiguity in relativeclause comprehension in mandarin. In
Proceedings of the CUNY Sentence Processing Conference.
The Chinese subject-object relative asymmetry
A computational and corpus investigation
Felix Engelmann, Kueilan Kuo and Shravan Vasishth
Cross-linguistically, subject relative clauses tend to be easier to
process than object relatives. However, Hsiao and Gibson (2003)
presented a surprising result that was recently replicated by Lin and
Garnsey (2007): in Chinese, subject relatives like (1a) are harder to
process than object relatives (1b).
This result is striking because it dramatically shrinks the space of
theoretical explanations for relative-clause processing. Theories that
rely on word-order regularities (Christiansen and MacDonald 2002,
Bever 1970) can predict the Hsiao and Gibson finding correctly: word
order in object relatives is SVO (the canonical order for Chinese) and
in object relatives it is VOS. By contrast, other theories fail to
explain the Chinese facts. For example, (i) Keenan and Comrie (1977)
rely on the greater accessibility of subjects compared to objects,
(ii) MacWhinney (1977) invokes the need to shift perspective from
subject to object in object relatives, which is harder in object
relatives than subject, (iii) Mitchell (1995) argues that structural
frequency of syntactic frames determines parsing difficulty, and since
subject relatives are more frequent than object, they should be easier
to process.
We make two contributions to this interesting finding by Hsiao and
Gibson.
First, we examine connectionist models, which have been cited (Hsiao
and Gibson 2003: 15) as an instantiation of the word order theory. The
predictions of a simple recurrent network model such as Christiansen
and MacDonald's have never been compared to reading times seen in
Chinese relative clauses -- it is merely assumed that the word-order
theories (like the connectionist model) make the correct predictions
for the results Hsiao and Gibson found. We present the predictions of
a simple recurrent network for Chinese relative clause processing, and
show that the predictions are rather mixed and do not point to clear
support for an SRN-based explanation. In our simulations, we ensured
that our SRN displays identical behavior to published connectionist
simulations involving Dutch and German center embeddings and
ungrammatical English center embeddings -- the model architecture is
thus identical to previous published models. The difficulties faced by
the connectionist model in predicting the Chinese reading times cast
doubt on this instantiation of the word order regularity explanation.
Second, we repeated the Hsiao and Gibson study using their stimuli in
a self-paced reading study. We found, contrary to their results, that
Chinese in fact shows a subject preference; independent research on
Chinese relatives supports our findings (Lin and Bever 2006), (Su et
al 2007).
We suggest that the best explanation for the subject-object relative
clause asymmetry lies in the relative frequencies of syntactic frames:
subject relatives in Chinese are, as in many other languages, more
frequent than object relatives, and this frequency-pattern predicts
the subject-relative preference correctly. Using Sinica Corpus (5
million words), we present evidence supporting this view. In addition,
we argue that the frequency patterns of noun-phrases animacy (animates
tend to occur before inanimates more frequently) may exert an
independent influence on processing difficulty. Independent evidence
consistent with this idea is presented by Lin and Garnsey (2007).
In sum, based on a series of simulations and a self-paced reading
experiment, we argue that (a) the word-order explanation for relative
clauses as instantiated by connectionist models may be problematic,
and (b) theories that predict the subject-preference
cross-linguistically (structural-frequency, accessibility, and
perspective) may be more compelling.
1a. Subject relative
[GAP yaoqing fuhao de] guanyuan xinhuaibugui
invite tycoon gen official have bad intentions
'The official who invited the tycoon has bad intentions.'
1b. Object relative
[fuhao yaoqing GAP de] guanyuan xinhuaibugui
tycoon invite gen official have bad intentions
'The official who the tycoon invited has bad intentions.'
MacDonald,M. C., & Christiansen, M. H. (2002). Reassessing working
memory: A reply to Just
and Carpenter andWaters and
Caplan. Psychological Review, 109(1), 35-54.
Hsiao, F. P.-F., & Gibson, E. (2003). Processing relative clauses in
Chinese. Cognition, 90, 3-27.
Lin, Y. B., & Garnsey, S.M. (2007). Plausibility and the resolution of
temporary ambiguity in relativeclause comprehension in mandarin. In
Proceedings of the CUNY Sentence Processing Conference.
Gerth, beim Graben, and Vasishth (Version 2)
COMMENT:
1. Minor changes in wording, and reduced the no. of lines of examples+refs to 15 as required.
2. Author order changed.
Toward dynamical system models of language-related brain potentials
Sabrina Gerth (1), Peter beim Graben (2), & Shravan Vasishth (1)
(1) Institute for Linguistics, University of Potsdam, Germany
(2) School of Psychology and Clinical Language Sciences, University of Reading, UK
A central question in computational psycholinguistics is how symbolic processing capabilities such
as language are realized by neurodynamics in the human brain. An important online measure of language
processing is event-related brain potentials (ERPs). However, ERPs are currently interpreted purely phenomenologically, and few computational accounts exist (see Hagoort 2005 for a recent proposal). We use tensor product representations of symbolic structures (Smolensky 1990; Smolensky and Legendre 2006; Mizraji 1989, 1992) to model syntactic parsing as nonlinear dynamics in neural activation space, and argue that ERP differences are due to trajectories which explore different regions in that space.
In an ERP experiment on the processing of German subject-object ambiguities, the sentence (1b) evoked a P600 ERP compared to (1a) reflecting an initial garden path interpretation. Starting (for simplicity) with a Government and Binding formulation (Haegeman 1994) of the sentences (1a,b), we construct a locally ambiguous context-free grammar from the phrase structure trees. This grammar is decomposed into its unambiguous parts representing the two alternative processing strategies, namely subject preference against object preference in order to construct two appropriate deterministic pushdown recognizers (Aho & Ullman 1972).
Using the tensor product representation, the syntactic categories of the disambiguated grammars are mapped
onto linearly independent filler vectors, while positions in a labeled binary tree are given as a basis of
three-dimensional space. In order to built a parallel processor (Lewis 1998) the two parses for the
subject-object sentence (regular vs. garden path) and the other two for the object-subject sentence
respectively were linearly superimposed in activation space. Then, model ERPs are obtained as the first
principal component.
Our model is able to describe, at least qualitatively, the obtained ERP results by trajectories that explore
functionally and causally different regions in activation space while pursuing different language processing
strategies. During its transient evolution, the trajectories of the model diverged exactly when the garden
path was encountered which shows remarkable resemblance with the P600 effect in the ERP.
(1a) Die Rednerin hat den Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][SUB] at the congress sought
'The speaker has sought the advisor at the congress'
(1b) Die Rednerin hat der Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][OBJ] at the congress sought
'The speaker has been sought by the advisor'
Haegeman L. (1994) Introduction to Government & Binding Theory, Blackwell Textbooks in Linguistics, vol1, 2nd
edn. Blackwell Publishers, Oxford, 1st edition 1991.
Hagoort P. (2005) On Broca, brain, and binding: a new framework. Trends in Cognitive Science 9(9):416 - 423.
Lewis RL (1998) Reanalysis and limited repair parsing: Leaping off the garden path. In: Fodor and Ferreira (1998), pp 147 - 285.
Mizraji E. (1989) Context-dependent association in linear distributed memories. Bulletin of Mathematical
Biology 51(2):195 - 205.
Smolensky P. (1990) Tensor product variable binding and the representation of symbolic structures in
connectionist systems. Artificial Intelligence 46:159 - 216.
1. Minor changes in wording, and reduced the no. of lines of examples+refs to 15 as required.
2. Author order changed.
Toward dynamical system models of language-related brain potentials
Sabrina Gerth (1), Peter beim Graben (2), & Shravan Vasishth (1)
(1) Institute for Linguistics, University of Potsdam, Germany
(2) School of Psychology and Clinical Language Sciences, University of Reading, UK
A central question in computational psycholinguistics is how symbolic processing capabilities such
as language are realized by neurodynamics in the human brain. An important online measure of language
processing is event-related brain potentials (ERPs). However, ERPs are currently interpreted purely phenomenologically, and few computational accounts exist (see Hagoort 2005 for a recent proposal). We use tensor product representations of symbolic structures (Smolensky 1990; Smolensky and Legendre 2006; Mizraji 1989, 1992) to model syntactic parsing as nonlinear dynamics in neural activation space, and argue that ERP differences are due to trajectories which explore different regions in that space.
In an ERP experiment on the processing of German subject-object ambiguities, the sentence (1b) evoked a P600 ERP compared to (1a) reflecting an initial garden path interpretation. Starting (for simplicity) with a Government and Binding formulation (Haegeman 1994) of the sentences (1a,b), we construct a locally ambiguous context-free grammar from the phrase structure trees. This grammar is decomposed into its unambiguous parts representing the two alternative processing strategies, namely subject preference against object preference in order to construct two appropriate deterministic pushdown recognizers (Aho & Ullman 1972).
Using the tensor product representation, the syntactic categories of the disambiguated grammars are mapped
onto linearly independent filler vectors, while positions in a labeled binary tree are given as a basis of
three-dimensional space. In order to built a parallel processor (Lewis 1998) the two parses for the
subject-object sentence (regular vs. garden path) and the other two for the object-subject sentence
respectively were linearly superimposed in activation space. Then, model ERPs are obtained as the first
principal component.
Our model is able to describe, at least qualitatively, the obtained ERP results by trajectories that explore
functionally and causally different regions in activation space while pursuing different language processing
strategies. During its transient evolution, the trajectories of the model diverged exactly when the garden
path was encountered which shows remarkable resemblance with the P600 effect in the ERP.
(1a) Die Rednerin hat den Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][SUB] at the congress sought
'The speaker has sought the advisor at the congress'
(1b) Die Rednerin hat der Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][OBJ] at the congress sought
'The speaker has been sought by the advisor'
Haegeman L. (1994) Introduction to Government & Binding Theory, Blackwell Textbooks in Linguistics, vol1, 2nd
edn. Blackwell Publishers, Oxford, 1st edition 1991.
Hagoort P. (2005) On Broca, brain, and binding: a new framework. Trends in Cognitive Science 9(9):416 - 423.
Lewis RL (1998) Reanalysis and limited repair parsing: Leaping off the garden path. In: Fodor and Ferreira (1998), pp 147 - 285.
Mizraji E. (1989) Context-dependent association in linear distributed memories. Bulletin of Mathematical
Biology 51(2):195 - 205.
Smolensky P. (1990) Tensor product variable binding and the representation of symbolic structures in
connectionist systems. Artificial Intelligence 46:159 - 216.
Gerth , Vasishth & beim Graben (Version 1)
Toward dynamical system models of language-related brain potentials
Sabrina Gerth (1), Shravan Vasishth (1) & Peter beim Graben (2)
(1) Institute for Linguistics, University of Potsdam, Germany
(2) School of Psychology and Clinical Language Sciences, University of Reading, UK
One of the crucial questions in computational psycholinguistics is how symbolic processing capabilities such
as language are realized by neurodynamics in the human brain. An important online measure of language
processing are event-related brain potentials (ERPs). Yet these are understood purely phenomenologically,
while computational accounts are by and large lacking (see Hagoort 2005 for a recent proposal). We shall use
tensor product representations of symbolic structures (Smolensky 1990; Smolensky and Legendre 2006; Mizraji 1989, 1992) to
describe syntactic parsing as nonlinear dynamics in neural activation space and argue that ERP differences
are due to trajectories which explore different regions in that space.
In an ERP experiment on the processing of German subject-object ambiguities, the sentence (1b) evoked a P600
ERP compared to (1a) reflecting an initial garden path interpretation. Starting from a Government and Binding
description (Haegeman 1994) of the sentences (1a,b), we construct a locally ambiguous context-free grammar
from the phrase structure trees. This grammar is decomposed into its unambiguous parts representing the two
alternative processing strategies, namely subject preference against object preference in order to construct
two appropriate deterministic pushdown recognizers (Aho & Ullman 1972).
Using the tensor product representation, the syntactic categories of the disambiguated grammars are mapped
onto linearly independent filler vectors, while positions in a labeled binary tree are given as a basis of
three-dimensional space. In order to built a parallel processor (Lewis 1998) the two parses for the
subject-object sentence (regular vs. garden path) and the other two for the object-subject sentence
respectively were linearly superimposed in activation space. Then, model ERPs are obtained as the first
principal component.
Our model is able to describe, at least qualitatively, the obtained ERP results by trajectories that explore
functionally and causally different regions in activation space in pursuing different language processing
strategies. During its transient evolution, the trajectories of the model diverged exactly when the garden
path was encountered which shows remarkable resemblance with the P600 effect in the ERP.
(1a) Die Rednerin hat den Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][SUB] at the congress sought
'The speaker has sought the advisor at the congress'
(1b) Die Rednerin hat der Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][OBJ] at the congress sought
'The speaker has been sought by the advisor'
References:
Aho AV, Ullman JD (1972) The Theory of Parsing, Translation and Compiling, vol I: Parsing. Prentice Hall,
Englewood Cliffs (NJ).
Haegeman L. (1994) Introduction to Government & Binding Theory, Blackwell Textbooks in Linguistics, vol1, 2end
edn. Blackwell Publishers, Oxford, 1st edition 1991.
Hagoort P. (2005) On Broca, brain, and binding: a new framework. Trends in Cognitive Science 9(9):416 - 423.
Lewis RL (1998) Reanalysis and limited repair parsing: Leaping off the garden path. In: Fodor and Ferreira
(1998), pp 147 - 285.
Mizraji E. (1992) Vector logics: The matrix-vector representation of logical calculus. Fuzzy Sets and Systems 50:179 - 185.
Mizraji E. (1989) Context-dependent association in linear distributed memories. Bulletin of Mathematical
Biology 51(2):195 - 205.
Smolensky P. (1990) Tensor product variable binding and the representation of symbolic structures in
connectionist systems. Artificial Intelligence 46:159 - 216.
Smolensky P., Legendre G. (2006) The Harmonic Mind. From Neural Computation to Optimality-Theoretic Grammar,
vol 1: Cognitive Architecture. MIT Press, Cambridge (MA).
Sabrina Gerth (1), Shravan Vasishth (1) & Peter beim Graben (2)
(1) Institute for Linguistics, University of Potsdam, Germany
(2) School of Psychology and Clinical Language Sciences, University of Reading, UK
One of the crucial questions in computational psycholinguistics is how symbolic processing capabilities such
as language are realized by neurodynamics in the human brain. An important online measure of language
processing are event-related brain potentials (ERPs). Yet these are understood purely phenomenologically,
while computational accounts are by and large lacking (see Hagoort 2005 for a recent proposal). We shall use
tensor product representations of symbolic structures (Smolensky 1990; Smolensky and Legendre 2006; Mizraji 1989, 1992) to
describe syntactic parsing as nonlinear dynamics in neural activation space and argue that ERP differences
are due to trajectories which explore different regions in that space.
In an ERP experiment on the processing of German subject-object ambiguities, the sentence (1b) evoked a P600
ERP compared to (1a) reflecting an initial garden path interpretation. Starting from a Government and Binding
description (Haegeman 1994) of the sentences (1a,b), we construct a locally ambiguous context-free grammar
from the phrase structure trees. This grammar is decomposed into its unambiguous parts representing the two
alternative processing strategies, namely subject preference against object preference in order to construct
two appropriate deterministic pushdown recognizers (Aho & Ullman 1972).
Using the tensor product representation, the syntactic categories of the disambiguated grammars are mapped
onto linearly independent filler vectors, while positions in a labeled binary tree are given as a basis of
three-dimensional space. In order to built a parallel processor (Lewis 1998) the two parses for the
subject-object sentence (regular vs. garden path) and the other two for the object-subject sentence
respectively were linearly superimposed in activation space. Then, model ERPs are obtained as the first
principal component.
Our model is able to describe, at least qualitatively, the obtained ERP results by trajectories that explore
functionally and causally different regions in activation space in pursuing different language processing
strategies. During its transient evolution, the trajectories of the model diverged exactly when the garden
path was encountered which shows remarkable resemblance with the P600 effect in the ERP.
(1a) Die Rednerin hat den Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][SUB] at the congress sought
'The speaker has sought the advisor at the congress'
(1b) Die Rednerin hat der Berater beim Kongress gesucht
The speaker [AMBIG] has [the advisor][OBJ] at the congress sought
'The speaker has been sought by the advisor'
References:
Aho AV, Ullman JD (1972) The Theory of Parsing, Translation and Compiling, vol I: Parsing. Prentice Hall,
Englewood Cliffs (NJ).
Haegeman L. (1994) Introduction to Government & Binding Theory, Blackwell Textbooks in Linguistics, vol1, 2end
edn. Blackwell Publishers, Oxford, 1st edition 1991.
Hagoort P. (2005) On Broca, brain, and binding: a new framework. Trends in Cognitive Science 9(9):416 - 423.
Lewis RL (1998) Reanalysis and limited repair parsing: Leaping off the garden path. In: Fodor and Ferreira
(1998), pp 147 - 285.
Mizraji E. (1992) Vector logics: The matrix-vector representation of logical calculus. Fuzzy Sets and Systems 50:179 - 185.
Mizraji E. (1989) Context-dependent association in linear distributed memories. Bulletin of Mathematical
Biology 51(2):195 - 205.
Smolensky P. (1990) Tensor product variable binding and the representation of symbolic structures in
connectionist systems. Artificial Intelligence 46:159 - 216.
Smolensky P., Legendre G. (2006) The Harmonic Mind. From Neural Computation to Optimality-Theoretic Grammar,
vol 1: Cognitive Architecture. MIT Press, Cambridge (MA).
Logacev and Vasishth, Version 2
Binding interference in sentence comprehension
New data and theory
Pavel Logacev and Shravan Vasishth
Current theories of similarity-based interference (SBI) claim that
when one feature (such as animacy or NP-type) is common (matches)
between noun phrases (NPs), the discriminability of the NPs degrades
during encoding as well as retrieval (Gordon et al 2007; Van Dyke and
McElree 2006; Lewis, Vasishth, & Van Dyke 2006), (Suckow et al. 2006).
These theories predict that interference should increase when two or
more features match. However, a reading study using eyetracking
suggests otherwise: we found that when one feature matches between
NPs, processing (reading) time is slower during retrieval, but when
two features match, processing time is faster. Using German stimuli
like (1) we manipulated the match/mismatch of two NPs in NP-type
(profession/nationality) and gender (masculine/feminine).
This surprising facilitation effect can be explained through an
independently motivated theory of feature binding, as proposed by Hommel
et al. (1998) in the context of visual processing.
The assumption is that objects in memory are composed of feature
bundles that are bound pairwise to each other. In the representation of
one NP, a binding between features (e.g. gender and NP-type) is termed
conflicting if the other NP contains a binding between a matching value
of one feature to a mismatching value of another feature. So if the NPs
match in both features (complete match) or in no feature (complete
mismatch) there are no conflicting bindings. If both NPs match in one
feature (e.g., NP-type) but not in the other one (e.g. gender), then one
value for NP-type (e.g. profession) will be bound to differently-valued
(e.g., masc/fem) gender-features of the two NPs.
The central claim is that interference is caused by binding conflicts.
The amount of interference experienced is crucially dependent on the
proportion of shared features: when the proportion is high, facilitation
results, when low, traditional interference effects result.
A remarkable fact is that the binding-interference theory can not only
explain the traditional SBI results which involve the mismatch of a
single feature, and the surprising facilitation effect when two
features mismatch, but also an intiguing and seemingly unrelated
phenomenon, the case-matching effect in German.
As discussed by Bader & Meng (1998) and Schlesewsky (1996), the case-
matching effect refers to the phenomenon that a case-marked NP (e.g.
accusative) can affect the resolution of a temporary case ambiguity of
a coreferent NP. Resolving it towards the same case is easier than
towards a different one (e.g. dative). Here, the assumption that the
case feature is bound to most other features and the relatively high
proportion of shared features between the coreferent NPs results less
conflicting bindings if the case value is the same. Therefore the
binding-interference theory predicts the observed facilitation.
We also present two other eyetracking experiments that provide further
evidence consistent with the predictions of the binding-interference
theory for case-matching effects.
In sum, the binding-interference theory furnishes a unifying
explanation for (i) traditional SBI effects, (ii) the surprising
facilitation effect when feature matches are increased, and (iii)
seemingly unrelated and otherwise puzzling case-matching effects. In
addition to its wide coverage of surprising results that traditional
theories cannot explain, the binding-interference account has the
virtue of being independently motivated (Hommel et al 1998) and
generalizable to other domains.
(1)
(a) *Die Polizistin* hat den Zaun, ...
The policewoman-FEM/PROF has the fence,
(b) *Der Polizist* hat den Zaun, ...
The policewoman-MASC/PROF has the fence,
(c) *Die Spanierin* hat den Zaun, ...
The policewoman-FEM/NAT has the fence,
(d) *Die Spanier* hat den Zaun, ...
The policewoman-MASC/NAT has the fence,
... den *die Nachtw‰chterin* deutlich sah, sehr ‰rgerlicherweise ¸bersehen.
that the night guard(?)-FEM/PROF clearly saw, very unfortunately overseen.
References:
[coming soon]
New data and theory
Pavel Logacev and Shravan Vasishth
Current theories of similarity-based interference (SBI) claim that
when one feature (such as animacy or NP-type) is common (matches)
between noun phrases (NPs), the discriminability of the NPs degrades
during encoding as well as retrieval (Gordon et al 2007; Van Dyke and
McElree 2006; Lewis, Vasishth, & Van Dyke 2006), (Suckow et al. 2006).
These theories predict that interference should increase when two or
more features match. However, a reading study using eyetracking
suggests otherwise: we found that when one feature matches between
NPs, processing (reading) time is slower during retrieval, but when
two features match, processing time is faster. Using German stimuli
like (1) we manipulated the match/mismatch of two NPs in NP-type
(profession/nationality) and gender (masculine/feminine).
This surprising facilitation effect can be explained through an
independently motivated theory of feature binding, as proposed by Hommel
et al. (1998) in the context of visual processing.
The assumption is that objects in memory are composed of feature
bundles that are bound pairwise to each other. In the representation of
one NP, a binding between features (e.g. gender and NP-type) is termed
conflicting if the other NP contains a binding between a matching value
of one feature to a mismatching value of another feature. So if the NPs
match in both features (complete match) or in no feature (complete
mismatch) there are no conflicting bindings. If both NPs match in one
feature (e.g., NP-type) but not in the other one (e.g. gender), then one
value for NP-type (e.g. profession) will be bound to differently-valued
(e.g., masc/fem) gender-features of the two NPs.
The central claim is that interference is caused by binding conflicts.
The amount of interference experienced is crucially dependent on the
proportion of shared features: when the proportion is high, facilitation
results, when low, traditional interference effects result.
A remarkable fact is that the binding-interference theory can not only
explain the traditional SBI results which involve the mismatch of a
single feature, and the surprising facilitation effect when two
features mismatch, but also an intiguing and seemingly unrelated
phenomenon, the case-matching effect in German.
As discussed by Bader & Meng (1998) and Schlesewsky (1996), the case-
matching effect refers to the phenomenon that a case-marked NP (e.g.
accusative) can affect the resolution of a temporary case ambiguity of
a coreferent NP. Resolving it towards the same case is easier than
towards a different one (e.g. dative). Here, the assumption that the
case feature is bound to most other features and the relatively high
proportion of shared features between the coreferent NPs results less
conflicting bindings if the case value is the same. Therefore the
binding-interference theory predicts the observed facilitation.
We also present two other eyetracking experiments that provide further
evidence consistent with the predictions of the binding-interference
theory for case-matching effects.
In sum, the binding-interference theory furnishes a unifying
explanation for (i) traditional SBI effects, (ii) the surprising
facilitation effect when feature matches are increased, and (iii)
seemingly unrelated and otherwise puzzling case-matching effects. In
addition to its wide coverage of surprising results that traditional
theories cannot explain, the binding-interference account has the
virtue of being independently motivated (Hommel et al 1998) and
generalizable to other domains.
(1)
(a) *Die Polizistin* hat den Zaun, ...
The policewoman-FEM/PROF has the fence,
(b) *Der Polizist* hat den Zaun, ...
The policewoman-MASC/PROF has the fence,
(c) *Die Spanierin* hat den Zaun, ...
The policewoman-FEM/NAT has the fence,
(d) *Die Spanier* hat den Zaun, ...
The policewoman-MASC/NAT has the fence,
... den *die Nachtw‰chterin* deutlich sah, sehr ‰rgerlicherweise ¸bersehen.
that the night guard(?)-FEM/PROF clearly saw, very unfortunately overseen.
References:
[coming soon]
Tuesday, December 4, 2007
von der Malsburg and Vasishth Version 1
A Similarity Measure for Scanpaths Sensitive to Fixation Durations
and the Location of Fixation Targets
Titus von der Malsburg and Shravan Vasishth
In the investigation of reading, word-based measures such as first fixation
duration, gaze duration and regression probability, are commonly used for
analyzing the empirical data. However, these measures are not always adequate
when the effect of an experimental manipulation is expected to appear in the
patterns of fixations rather than in local measures, e.g. when reanalysis
strategies in temporarily ambiguous sentences are studied. We address this
shortcoming by proposing alternative methods for analyzing eye movements
that are based on the notion of similarity of scanpaths. For this
purpose we introduce a measure that quantifies the similarity of two scanpaths
with respect to their fixation patterns and their temporal dynamics.
Our metric is a member of the class of edit-distances of which the Levenshtein
metric is a well-known instance (Levenshtein 1968). Edit-distances have already
been used for analysing eye tracking data in a variety of studies outside
reading research, for example in experiments studying scene perception or the
usability of graphical user interfaces [citation]. While edit-distances capture
the topological similarities of scanpaths, they operate on a representation of
scanpaths that neglects properties that are crucial when studying reading. The
first is fixation duration, which is one of the most important and best studied
sources of evidence in online reading research. Secondly, the available
edit-distance measures are insensitive to the proximity of regions of interest.
However, the observation of parafoveal preview effects suggest that the spatial
relations of regions of interest do play an important role and should be
accounted for [cite something that demonstrates preview effects].
We addressed these deficits in standard edit-distance measures by replacing the
cost function of the Levenshtein metric. Our new cost function compares
fixations with respect to their duration and spatial distance. Thus, the way
distance is incorporated is motivated by findings about the distribution of
acuity in the visual field (cortical magnification, Rovamo, Virsu, N‰s‰nen,
1978). The resulting measure operates on a parsimonious representation of eye
movement data and can efficiently be computed (O(m*n), where n, m are the
numbers of fixations of the two compared scanpaths).
We suggest several approaches for analysing eye movement data based on this new
measure: 1.) Cluster analysis can be applied to scanpaths in order to identify
reading strategies among experimental conditions and among subjects. 2.) The
variance of the pairwise similarities in sets of scanpaths can be compared to
the variance of the union of those sets in order to determine whether they
represent different populations. 3.) String interpolation methods and
optimization techniques can be used for computing prototypical scanpaths for
sets of scanpaths.
Areas of investigation where these methods might also be useful comprise the
evaluation of models of oculo-motor control, classification of reading
disorders, analysis of cognitive strategies in experiments using the visual
world paradigm.
As opposed to approaches using word-based measures, the proposed similarity-based
methods enable the analysis of the full available data and do not necessitate
averaging of variables thereby preserving possibly crucial information. The
main benefit of our similarity measure is that it is sensitive to fixation
patterns and the temporal dynamics in a scanpaths at the same time.
References:
[coming soon]
and the Location of Fixation Targets
Titus von der Malsburg and Shravan Vasishth
In the investigation of reading, word-based measures such as first fixation
duration, gaze duration and regression probability, are commonly used for
analyzing the empirical data. However, these measures are not always adequate
when the effect of an experimental manipulation is expected to appear in the
patterns of fixations rather than in local measures, e.g. when reanalysis
strategies in temporarily ambiguous sentences are studied. We address this
shortcoming by proposing alternative methods for analyzing eye movements
that are based on the notion of similarity of scanpaths. For this
purpose we introduce a measure that quantifies the similarity of two scanpaths
with respect to their fixation patterns and their temporal dynamics.
Our metric is a member of the class of edit-distances of which the Levenshtein
metric is a well-known instance (Levenshtein 1968). Edit-distances have already
been used for analysing eye tracking data in a variety of studies outside
reading research, for example in experiments studying scene perception or the
usability of graphical user interfaces [citation]. While edit-distances capture
the topological similarities of scanpaths, they operate on a representation of
scanpaths that neglects properties that are crucial when studying reading. The
first is fixation duration, which is one of the most important and best studied
sources of evidence in online reading research. Secondly, the available
edit-distance measures are insensitive to the proximity of regions of interest.
However, the observation of parafoveal preview effects suggest that the spatial
relations of regions of interest do play an important role and should be
accounted for [cite something that demonstrates preview effects].
We addressed these deficits in standard edit-distance measures by replacing the
cost function of the Levenshtein metric. Our new cost function compares
fixations with respect to their duration and spatial distance. Thus, the way
distance is incorporated is motivated by findings about the distribution of
acuity in the visual field (cortical magnification, Rovamo, Virsu, N‰s‰nen,
1978). The resulting measure operates on a parsimonious representation of eye
movement data and can efficiently be computed (O(m*n), where n, m are the
numbers of fixations of the two compared scanpaths).
We suggest several approaches for analysing eye movement data based on this new
measure: 1.) Cluster analysis can be applied to scanpaths in order to identify
reading strategies among experimental conditions and among subjects. 2.) The
variance of the pairwise similarities in sets of scanpaths can be compared to
the variance of the union of those sets in order to determine whether they
represent different populations. 3.) String interpolation methods and
optimization techniques can be used for computing prototypical scanpaths for
sets of scanpaths.
Areas of investigation where these methods might also be useful comprise the
evaluation of models of oculo-motor control, classification of reading
disorders, analysis of cognitive strategies in experiments using the visual
world paradigm.
As opposed to approaches using word-based measures, the proposed similarity-based
methods enable the analysis of the full available data and do not necessitate
averaging of variables thereby preserving possibly crucial information. The
main benefit of our similarity measure is that it is sensitive to fixation
patterns and the temporal dynamics in a scanpaths at the same time.
References:
[coming soon]
Engelmann, Kuo and Vasishth Version 1
The Chinese subject-object relative asymmetry:
A computational and corpus investigation
Felix Engelmann, Kueilan Kuo and Shravan Vasishth
Cross-linguistically, subject relative clauses tend to be easier to
process than object relatives. However, Hsiao and Gibson (2003)
presented a surprising result that was recently replicated by Lin and
Garnsey (2007): in Chinese, subject relatives like (1a) are harder to
process than object relatives (1b).
This result is striking because it dramatically shrinks the space of
theoretical explanations for relative-clause processing. Since word
order in object relatives is SVO (the canonical order for Chinese) and
in object relatives it is VOS, theories that rely on word-order
regularities (Christiansen and MacDonald 2002, Bever 1970) can predict
the Hsiao and Gibson finding correctly, but other theories fail, such
as Keenan and Comrie's (1977) which relies on the different degrees of
accessibility of subjects versus objects, MacWhinney's (1977), which
invokes the need to shift perspective from subject to object in object
relatives (MacWhinney 1977), and structural frequency (Mitchell 1995),
which asssumed that the frequency of syntactic frames determines
parsing difficulty.
We make two contributions to this interesting finding by Hsiao and
Gibson.
First, although connectionist models have been cited (Hsiao
and Gibson 2003: 15) as an instantiation of the word order theory, the
predictions of a simple recurrent network model such as Christiansen
and MacDonald's have never been compared to moment-by-moment reading
times seen in Chinese relative clauses -- it is merely assumed that
the word-order theories (like the connectionist model) would be correct for
the results Hsiao and Gibson found. We present the predictions of a
simple recurrent network for Chinese relative clause processing, and
show that the predictions are rather mixed and do not point to clear
support for an SRN-based explanation. In our simulations, we took the
precaution of ensuring that our SRN displays identical behavior to
published connectionst simulations involving Dutch and German center
embeddings and ungrammatical English center embeddings -- the model
architecture is thus identical to previous published models. The mixed
nature of the connectionist model's predictions thus poses considerable
problems for the word-order based accont.
Second, we repeated the Hsiao and Gibson study using their stimuli in
a self-paced reading study. We found, contrary to their results, that
Chinese in fact shows a subject preference; independent research on
Chinese relatives supports our findings (Lin and Bever 2006), (Su et
al 2007).
We suggest that the best explanation for the subject-object relative
clause asymmetry lies in the relative frequencies of syntactic frames:
subject relatives in Chinese are, as in other languages, more frequent
than object relatives and this frequency-pattern predicts the
subject-relative preference correctly. Using Sinica Corpus (5 million
words), we present evidence supporting this view. In addition, we
argue that the frequency patterns of noun-phrases animacy (animates
tend to occur before inanimates more frequently) may exert an
independent influence on processing difficulty.
In sum, we argue that (a) the word-order explanation for relative
clauses as instantiated by connectinoist models may be problematic,
and (b) the data available so far argues in favor of the
structural-frequency account of relative-clause processing.
1a. Subject relative
[GAP yaoqing fuhao de] guanyuan xinhuaibugui
invite tycoon gen official have bad intentions
'The official who invited the tycoon has bad intentions.'
1b. Object relative
[fuhao yaoqing GAP de] guanyuan xinhuaibugui
tycoon invite gen official have bad intentions
'The official who the tycoon invited has bad intentions.'
References:
[Coming soon]
A computational and corpus investigation
Felix Engelmann, Kueilan Kuo and Shravan Vasishth
Cross-linguistically, subject relative clauses tend to be easier to
process than object relatives. However, Hsiao and Gibson (2003)
presented a surprising result that was recently replicated by Lin and
Garnsey (2007): in Chinese, subject relatives like (1a) are harder to
process than object relatives (1b).
This result is striking because it dramatically shrinks the space of
theoretical explanations for relative-clause processing. Since word
order in object relatives is SVO (the canonical order for Chinese) and
in object relatives it is VOS, theories that rely on word-order
regularities (Christiansen and MacDonald 2002, Bever 1970) can predict
the Hsiao and Gibson finding correctly, but other theories fail, such
as Keenan and Comrie's (1977) which relies on the different degrees of
accessibility of subjects versus objects, MacWhinney's (1977), which
invokes the need to shift perspective from subject to object in object
relatives (MacWhinney 1977), and structural frequency (Mitchell 1995),
which asssumed that the frequency of syntactic frames determines
parsing difficulty.
We make two contributions to this interesting finding by Hsiao and
Gibson.
First, although connectionist models have been cited (Hsiao
and Gibson 2003: 15) as an instantiation of the word order theory, the
predictions of a simple recurrent network model such as Christiansen
and MacDonald's have never been compared to moment-by-moment reading
times seen in Chinese relative clauses -- it is merely assumed that
the word-order theories (like the connectionist model) would be correct for
the results Hsiao and Gibson found. We present the predictions of a
simple recurrent network for Chinese relative clause processing, and
show that the predictions are rather mixed and do not point to clear
support for an SRN-based explanation. In our simulations, we took the
precaution of ensuring that our SRN displays identical behavior to
published connectionst simulations involving Dutch and German center
embeddings and ungrammatical English center embeddings -- the model
architecture is thus identical to previous published models. The mixed
nature of the connectionist model's predictions thus poses considerable
problems for the word-order based accont.
Second, we repeated the Hsiao and Gibson study using their stimuli in
a self-paced reading study. We found, contrary to their results, that
Chinese in fact shows a subject preference; independent research on
Chinese relatives supports our findings (Lin and Bever 2006), (Su et
al 2007).
We suggest that the best explanation for the subject-object relative
clause asymmetry lies in the relative frequencies of syntactic frames:
subject relatives in Chinese are, as in other languages, more frequent
than object relatives and this frequency-pattern predicts the
subject-relative preference correctly. Using Sinica Corpus (5 million
words), we present evidence supporting this view. In addition, we
argue that the frequency patterns of noun-phrases animacy (animates
tend to occur before inanimates more frequently) may exert an
independent influence on processing difficulty.
In sum, we argue that (a) the word-order explanation for relative
clauses as instantiated by connectinoist models may be problematic,
and (b) the data available so far argues in favor of the
structural-frequency account of relative-clause processing.
1a. Subject relative
[GAP yaoqing fuhao de] guanyuan xinhuaibugui
invite tycoon gen official have bad intentions
'The official who invited the tycoon has bad intentions.'
1b. Object relative
[fuhao yaoqing GAP de] guanyuan xinhuaibugui
tycoon invite gen official have bad intentions
'The official who the tycoon invited has bad intentions.'
References:
[Coming soon]
Logacev and Vasishth Version 1
Binding interference in sentence comprehension
New data and theory
Pavel Logacev and Shravan Vasishth
Current theories of similarity-based interference (SBI) claim that
when one feature (such as animacy or NP-type) is common (matches)
between noun phrases (NPs), the discriminability of the NPs degrades
during encoding as well as retrieval (Gordon et al 2007; Van Dyke and
McElree 2006; Lewis, Vasishth, & Van Dyke 2006), (Suckow et al. 2006).
These theories predict that interference should increase when two or
more features match. However, a German reading study using eyetracking
suggests otherwise: we found that when one feature matches between
NPs, processing (reading) time is slower during retrieval, but when
two features match, processing time is faster.
This surprising facilitation effect can be explained through an
independently motivated theory of feature binding proposed by Hommel
et al. (1998) in the context of visual processing.
The assumption is that objects in memory are composed of feature
bundles that are bound pairwise to each other. A pair of NP-features
(e.g., gender and NP-type) is termed non-conflicting if both are
associated with the same NP (complete match). If one of two bound
features (e.g., NP-type) has the same value for two different NPs, the
features conflict (complete mismatch): the common NP-type feature has a
conflicting binding with differently-valued (e.g., masc/fem)
gender-features of the two NPs. The central claim is that interference
is caused by binding conflicts. The amount of interference experienced
is crucially dependent on the proportion of shared features: when the
proportion is high, facilitation results, when low, traditional
interference effects result.
A remarkable fact is that the binding-interference theory can not only
explain the traditional SBI results which involve the mismatch of a
single feature, and the surprising facilitation effect when two
features mismatch, but also an intiguing and seemingly unrelated
phenomenon, the case-matching effect in German.
As discussed by Bader & Meng (1998) and Schlesewsky (1996), the
case-matching effect is that it is that, when a verb is read that
takes an accusative or dative object, it is easier to resolve a
temporary case ambiguity of an NP (e.g., of Operasaengerinnen, (1a,b))
towards accusative, respectively dative when a coreferent pronoun
intervening between the NP and the verb's argument-features has
accusative or dative case. Here, the relatively high proportion of
shared features between the coreferent NPs results in the
facilitation, as predicted by the binding-interference theory.
We also present two other eyetracking experiments that provide further
evidence consistent with the predictions of the binding-interference
theory for case-matching effects.
In sum, the binding-interference theory furnishes a unifying
explanation for (i) traditional SBI effects, (ii) the surprising
facilitation effect when feature matches are increased, and (iii)
seemingly unrelated and otherwise puzzling case-matching effects. In
addition to its wide coverage of surprising results that traditional
theories cannot explain, the binding-interference account has the
virtue of being independently motivated (Hommel et al 1998).
...dass man...
that one
(a) Opernsaengerinnen, obwohl man sie-ACC verehrt, begruflen sollte...
opera singers although one them adores greet should
(b) Opernsaengerinnen, obwohl man ihnen-DAT misstraut, begruflen sollte...
opera singers although one them distrusts greet should
We present three further eyetracking experiments that vary (a) gender and
noun type of NPs (b)
References:
[coming soon]
New data and theory
Pavel Logacev and Shravan Vasishth
Current theories of similarity-based interference (SBI) claim that
when one feature (such as animacy or NP-type) is common (matches)
between noun phrases (NPs), the discriminability of the NPs degrades
during encoding as well as retrieval (Gordon et al 2007; Van Dyke and
McElree 2006; Lewis, Vasishth, & Van Dyke 2006), (Suckow et al. 2006).
These theories predict that interference should increase when two or
more features match. However, a German reading study using eyetracking
suggests otherwise: we found that when one feature matches between
NPs, processing (reading) time is slower during retrieval, but when
two features match, processing time is faster.
This surprising facilitation effect can be explained through an
independently motivated theory of feature binding proposed by Hommel
et al. (1998) in the context of visual processing.
The assumption is that objects in memory are composed of feature
bundles that are bound pairwise to each other. A pair of NP-features
(e.g., gender and NP-type) is termed non-conflicting if both are
associated with the same NP (complete match). If one of two bound
features (e.g., NP-type) has the same value for two different NPs, the
features conflict (complete mismatch): the common NP-type feature has a
conflicting binding with differently-valued (e.g., masc/fem)
gender-features of the two NPs. The central claim is that interference
is caused by binding conflicts. The amount of interference experienced
is crucially dependent on the proportion of shared features: when the
proportion is high, facilitation results, when low, traditional
interference effects result.
A remarkable fact is that the binding-interference theory can not only
explain the traditional SBI results which involve the mismatch of a
single feature, and the surprising facilitation effect when two
features mismatch, but also an intiguing and seemingly unrelated
phenomenon, the case-matching effect in German.
As discussed by Bader & Meng (1998) and Schlesewsky (1996), the
case-matching effect is that it is that, when a verb is read that
takes an accusative or dative object, it is easier to resolve a
temporary case ambiguity of an NP (e.g., of Operasaengerinnen, (1a,b))
towards accusative, respectively dative when a coreferent pronoun
intervening between the NP and the verb's argument-features has
accusative or dative case. Here, the relatively high proportion of
shared features between the coreferent NPs results in the
facilitation, as predicted by the binding-interference theory.
We also present two other eyetracking experiments that provide further
evidence consistent with the predictions of the binding-interference
theory for case-matching effects.
In sum, the binding-interference theory furnishes a unifying
explanation for (i) traditional SBI effects, (ii) the surprising
facilitation effect when feature matches are increased, and (iii)
seemingly unrelated and otherwise puzzling case-matching effects. In
addition to its wide coverage of surprising results that traditional
theories cannot explain, the binding-interference account has the
virtue of being independently motivated (Hommel et al 1998).
...dass man...
that one
(a) Opernsaengerinnen, obwohl man sie-ACC verehrt, begruflen sollte...
opera singers although one them adores greet should
(b) Opernsaengerinnen, obwohl man ihnen-DAT misstraut, begruflen sollte...
opera singers although one them distrusts greet should
We present three further eyetracking experiments that vary (a) gender and
noun type of NPs (b)
References:
[coming soon]
Friday, November 30, 2007
Hi all, here is our new lab blog!
At this time, we invite all people in the MSU-CL blog for purposes of critiquing CUNY abstracts.
Thank you.
Thank you.
Subscribe to:
Posts (Atom)