HAL Id: hal-02004264
https://hal.archives-ouvertes.fr/hal-02004264
Submitted on 1 Feb 2019
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Reconciling opposite neighborhood frequency effects in lexical decision: Evidence from a novel probabilistic
model of visual word recognition
Thierry Phénix, Sylviane Valdois, Julien Diard
To cite this version:
Thierry Phénix, Sylviane Valdois, Julien Diard. Reconciling opposite neighborhood frequency effects
in lexical decision: Evidence from a novel probabilistic model of visual word recognition. Conference
of the Society of Scientific Studies of Reading (SSSR), Jul 2018, Brighton, United Kingdom. �hal-
02004264�
A
The general structure of the BRAID model relies on three levels of processing as featured in most previous word recogni:on models.
1. Thele#er sensory submodelimplements low-level visual mechanisms involved in le>er iden:fica:on and le>er posi:on coding. Feature extrac:on is parallel over the input string, an acuity gradient is implemented symmetrically around fixa:on and loca:on is distributed over adjacent le>er posi:ons, implemen:ng lateral interference between le>ers.
2. Thele#er perceptual submodelimplements how informa:on extracted from the sensory input accumulates gradually over :me to create a percept, i.e. an internal representa:on of the input le>er string.
3. Thelexical knowledge submodelimplements knowledge about the 40.481 English words of the Bri:sh Lexicon Project. The probability to recognize a word is modulated by its frequency. This level is enriched by a module, called thelexical membership submodel, that implements a mechanism to decide whether or not the input le>er-string is a known word. This submodel is cri:cal to simulate lexical decision.
One major originality of BRAID is to assume the existence of a fourth level:
4. Thevisual a#en8onal submodelimplements an a>en:onal filtering mechanism between the le>er sensory submodel and the le>er perceptual submodel. Transfer of informa:on between these two submodels is modulated by the amount of a>en:on allocated to each le>er posi:on.
References
Ginestet, E., Phenix, T., Diard, J., & Valdois, S. (submi>ed). Modelling length effect for words in lexical decision: the role of visual a>en:on.
Phénix, T., Valdois, S., & Diard, J. (submi>ed). Bayesian word recogni:on with a>en:on, interference and dynamics.
Phénix, T. (2018). Modélisa)on bayésienne algorithmique de la reconnaissance visuelle de mots et de l’a9en)on visuelle.Universite ́ Grenoble Alpes.
h>ps://tel.archives-ouvertes.fr/tel-01735606
Reconciling opposite neighborhood frequency effects in lexical decision:
Evidence from a novel probabilistic model of visual word recognition
Thierry Phénix, Sylviane Valdois and Julien Diard
Laboratoire de Psychologie et NeuroCognition
Univ. Grenoble Alpes, CNRS, LPNC UMR 5105, F-38000 Grenoble, France Contact: [email protected]
Neighborhood Frequency Effects in Lexical Decision Behavioral Task Simula:ons:
Probabilis:c Ques:ons
BRAID Bayesian Word Recogni:on with A>en:on Interference and Dynamics
Lexical Knowledge Submodel
Acuity Interference Final
Sensory Letter Submodel
C L O N E
Visual A#en8on Submodel
clone close alone
0 200 400 600 800 1000
0.0 0.2 0.4 0.6 0.8 1.0
Iteration
Recognitionprobability
Perceptual Letter Submodel 100 iterations 250 iterations
350 itera8ons 500 itera8ons
Q0PT
n=P(PnT |s1:T1:N[L1:T
1:N= 1] [ P1:T
1:N= 1]µ1:TA 1:T A g1:T)
• Letter identification with lexical knowledge
QWT
=P(WT|s1:T1:N[ L1:T
1:N= 1] [ P1:T
1:N= 1]µ1:TA 1:T A g1:T)
• Word iden8fica8on
QDT
=P(DT|s1:T1:N[ D1:T
1:N= 1] [ P1:T
1:N= 1]µ1:TA 1:T A g1:T)
• Lexical decision
Context
In the lexical decision task, participants have to indicate whether a printed stimulus is a known word (YES response) or not. Among the classical effects related to lexical decision, a poorly understood and highly controversial effect is the neighborhood frequency effect.
Latency on YES responses in lexical decision is modulated by the existence of higher frequency orthographic neighbors, namely the existence of at least one word of higher frequency that differs from the target word by a single letter (e.g., LIKE is a higher frequency neighbor of the target word BIKE).
Contradictory behavioral experiments Perea & Pollatsek (1998) Siakaluk, Sears, & Lupker (2002) Material and procedure:
The two experiments manipulate neighborhood frequency of English words, considering words with one (1HF) or zero (OHF) higher frequency neighbor. Experiment A used small density N and manipulated frequency whereas experiment B manipulated N for low frequency words. Ninety-two lowercase 5-to-6 letter words are used in Experiment A, 60 uppercase 4-to-5 letter words in Experiment B.
Results:
A significant inhibitory neighborhood frequency effect is reported in Experiment A but only for the low frequency words. In contrast, Experiment B reports a significant facilitatory neighborhood frequency effect for low-to- medium frequency words.
Lexical Membership
Submodel Word Is word?
Behavioral experiment:A standard finding is that high frequency words are processed faster than low frequency words and that identification time is proportional to log- frequency.
Simulation:To simulate this effect, we reproduce the simulation design first proposed by Norris (2006). The frequency of occurrence of each word rotates in a round- robin manner over all experimental conditions, so that only frequency affects identification time. The BRAID model reproduces the log-frequency effect.
Frequencies 1 2 5 10 20 50 100 200 500 1000 2000 5000 10 000
0 200 400 600 800 1000
0.0 0.2 0.4 0.6 0.8 1.0
Iteration
Recognitionprobability
1 10 100 1000 10 000
0 50 100 150 200 250 300
Log frequency
Iteration
Behavioral experiment:Johnston (1978, experiment 2) showed that le>ers are recognized significantly more oken when they are presented in a word context rather than in isola:on and when they are presented in isola:on rather than in a non-word context.
Simula8on:BRAID reproduces behavioral results (e.g., aker 62 itera:ons). The mean :me course of le>er iden:fica:on shows that the word effect over isolated le>er and non-word is a strong effect and that the advantage of isolated le>ers over a non-word context, significant from itera:on 43, is durable.
Data BRAID
Mot Non-mot Lettre isolée
0.0 0.2 0.4 0.6 0.8 1.0
Contexte
Probabilitéd'identificationdelettre
Context Word Nonword Isolated letter
0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
iteration
Letteridentificationprobability
Behavioral experiment:Andrews (1989, experiment 1 and 3) showed that high density neighborhood slows down le>er iden:fica:on significantly only for low frequency s:muli and has no effect on high frequency s:muli.
Simula8on:BRAID reproduces behavioral results with a threshold of YES response at 0.77. More interes:ngly, it shows that having a high density neighborhood is an advantage at the beginning of the iden:fica:on process but becomes a drawback when the decision is based on a high level of precision.
Data Braid
LF/SD LF/LD HF/SD HF/LD
200 300 400 500 600 700 800
Freq./N
Latency(ms)/iteration
Freq./N LF/SD LF/LD HF/SD HF/LD
0 100 200 300 400
0.6 0.7 0.8 0.9 1.0
Iteration
Identificationprobability
Uniform Standard Reduced
C L O N E
0.0 0.2 0.4 0.6 0.8
Probability
C L O N E
0.0 0.2 0.4 0.6 0.8
Probability
C L O N E
0.0 0.2 0.4 0.6 0.8
Probability
Simula8on
Material and procedure:
The s:muli used in the simula:ons are the same as in the behavioral experiments. We first simulate lexical decision in BRAID for 1,000 itera:ons. Words that did not reach 0.97 iden:fica:on probability aker 1,000 itera:ons were removed, resul:ng in an error rate of 1%
and 2.5% for the s:muli used in Simula:on A and B, respec:vely.
Results:
The simulated data successfully capture both inhibitory and facilitatory effects of human data, with the same set of parameters. The dynamic curves reveal an inversion of the neighborhood frequency effect with :me. The existence of a higher frequency neighbor is facilitatory at the beginning of processing but turns inhibitory over :me. In simula:on A, the pa>ern inversion occurs earlier for the low frequency words (itera:on 188; threshold value=0.68) than for the medium frequency words (itera:on 213; threshold value=0.81), and, in simula:on B, earlier for the words that have a small vs. large neighborhood density (itera:on 217 vs. 239;
threshold value = 0.81 vs. 0.85).
Data Braid
LF/0HF LF/1HF MF/0HF MF/1HF SD/0HF SD/1HF LD/0HF LD/1HF 400
450 500 550 600 650 700
Condition
Latency(ms)/Timesteps
A: Perea and Pollatsek (1998)Condition B: Siakaluk et al. (2002)
Condition SD/0HF SD/1HF LD/0HF LD/1HF
0 100 200 300 400
0.6 0.7 0.8 0.9 1.0
Time steps
ProbabilityofYESresponse
Condition LF/0HF LF/1HF MF/0HF MF/1HF
0 100 200 300 400 500
0.6 0.7 0.8 0.9 1.0
Time steps
ProbabilityofYESresponse
A
B
