GRS LX 865
Topics in Linguistics
Week 6. Optimality Theory and
acquisition.
Optimality Theory



Grammar involves constraints on the
representations (e.g., SS, LF, PF, or
perhaps a combined representation).
The constraints exist in all languages.
Where languages differ is in how important
each constraint is with respect to each
other constraint.
Optimality Theory


In our analysis, one constraint is Parse-T,
which says that tense must be realized in
a clause. A structure without tense (where
TP has been omitted, say) will violate this
constraint.
Another constraint is *F (“Don’t have a
functional category”). A structure with TP
will violate this constraint.
Optimality Theory


Parse-T and *F are in conflict—it is
impossible to satisfy both at the same
time.
When constraints conflict, the choice
made (on a language-particular basis) of
which constraint is considered to be “more
important” (more highly ranked)
determines which constraint is satisfied
and which must be violated.
Optimality Theory


So if *F >> Parse-T, TP will be omitted.
and if Parse-T >> *F, TP will be included.
Optimality Theory—big
picture



Universal Grammar is the constraints
that languages must obey.
Languages differ only in how those
constraints are ranked relative to one
another. (So, “parameter” = “ranking”)
The kid’s job is to re-rank constraints
until they match the order which
generated the input that s/he hears.
French kid data


This means if a kid uses 3sg or present
tense, we can’t tell if they are really using
3sg (they might be) or if they are not using
agreement at all and just pronouncing the
default.
So, we looked at non-present tense forms
and non-3sg forms only to avoid the
question of the defaults.
The idea

Kids are subject to conflicting constraints:
Parse-T
 Parse-Agr
 *F


*F2
Include a projection for tense
Include a project for agreement
Don’t complicate your tree with
functional projections
Don’t complicate your tree so
much as to have two functional
projections.
The idea

Sometimes Parse-T beats out *F, and then
there’s a TP. Or Parse-Agr beats out *F,
and then there’s an AgrP. Or both Parse-T
and Parse-Agr beat out *F2, and so there’s
both a TP and an AgrP.

But what does sometimes mean?
Floating constraints

The innovation in Legendre et al. (2000)
that gets us off the ground is the idea that
as kids re-rank constraints, the position of
the constraint in the hierarchy can get
somewhat fuzzy, such that two positions
can overlap.
*F
Parse-T
Floating constraints
*F
Parse-T

When the kid evaluates a form in the
constraint system, the position of ParseT is fixed somewhere in the range—and
winds up sometimes outranking, and
sometimes outranked by, *F.
Floating constraints
*F
Parse-T

(Under certain assumptions) this
predicts that we would see TP in the
structure 50% of the time, and see
structures without TP the other 50% of
the time.
French kid data



Looked at 3 French kids from CHILDES
Broke development into stages based on a modified
MLU-type measure based on how long most of their
utterances were (2 words, more than 2 words) and
how many of the utterances contain verbs.
Looked at tense and agreement in each of the three
stages represented in the data.
French kid data


Kids start out using 3sg agreement and
present tense for practically everything
(correct or not).
We took this to be a “default”

(No agreement? Pronounce it as 3sg. No
tense? pronounce it as present. Neither?
Pronounce it as an infinitive.).
French kid data


This means if a kid uses 3sg or present
tense, we can’t tell if they are really using
3sg (they might be) or if they are not using
agreement at all and just pronouncing the
default.
So, we looked at non-present tense forms
and non-3sg forms only to avoid the
question of the defaults.
French kids data


We found that tense and agreement
develop differently—specifically, in the first
stage we looked at, kids were using tense
fine, but then in the next stage, they got
worse as the agreement improved.
Middle stage: looks like
competition between T
and Agr for a single node.
A detail about counting




We counted non-3sg and non-present verbs.
In order to see how close kids’ utterances were to adult’s
utterances, we need to know how often adults use non3sg and non-present, and then see how close the kids
are to matching that level.
So, adults use non-present tense around 31% of the
time—so when a kid uses 31% non-present tense, we
take that to be “100% success”
In the last stage we looked at, kids were basically right at
the “100% success” level for both tense and agreement.
Proportion of non-present
and non-3sg verbs
40%
35%
30%
25%
non-present
non-3sg
adult non-pres
adult non-3sg
20%
15%
10%
5%
0%
3b
4b
4c
Proportion of non-finite root
forms
35%
30%
25%
20%
NRFs
15%
10%
5%
0%
3b
4b
4c
A model to predict the
percentages



Stage 3b (first stage)
no agreement
about 1/3 NRFs, 2/3 tensed forms
*F2
ParseT
ParseA
*F
A model to predict the
percentages



Stage 4b (second stage)
non-3sg agreement and non-present tense
each about 15% (=about 40% agreeing,
50% tensed)
about 20% NRFs
*F2
ParseT
ParseA
*F
A model to predict the
percentages


Stage 4c (third stage)
everything appears to have tense and
agreement (adult-like levels)
*F2
ParseT
ParseA
*F
Predicted vs. observed—
tense
40%
35%
30%
25%
non-present
predicted non-pres
20%
15%
10%
5%
0%
3b
4b
4c
Predicted vs. observed—agr’t
40%
35%
30%
25%
non-3sg
predicted non-3sg
20%
15%
10%
5%
0%
3b
4b
4c
Predicted vs. observed—
NRFs
35%
30%
25%
20%
NRFs
predicted NRFs
15%
10%
5%
0%
3b
4b
4c
Various things (homework)

Is the OT model just proposed a structurebuilding or full competence model?

How does the OT model fit in the overall
big picture with the ATOM model?
Various things (homework)


For French, we assumed that NRFs
appear when both TP and AgrP are
missing. Yet, Schütze & Wexler 1996
claimed the root infinitives appeared with
either TP or AgrP were missing.
Which one is it?
French v. English

English: T+Agr is pronounced like
/s/ if we have features [3, sg, present]
 /ed/ if we have the feature [past]
 /Ø/ otherwise


French: T+Agr is pronounced like:
danser
 a dansé
 je danse
 j’ai dansé

NRF
(3sg) past
1sg (present)
1sg past










What we’re doing





The driver who my neighbor who I trust
suggested took me to the airport.
The driver who my neighbor who my boss trusts
suggested took me to the airport.
Overarching hypothesis: Sentence difficulty has
to do with holding onto several unsatisfied
dependencies. Longer ones are harder to hold.
Question: What measures length?
Hypothesis: New referents.
How do we see if that’s right?

Center-embedded sentences are the most
taxing, several started dependencies,
center-most element triple-counted.

The driver who my neighbor who I trust …

That’s the most sensitive point, seems to
be near critical point of processability.
Experimenting

Does it matter whether we have a known
referent (I, you) or a new referent (my
neighbor)?

To know for sure, we try holding everything
constant except the most embedded subject and
see if there are differences (which can then be
attributed to the only thing that’s different, the
properties of the most embedded subject).
Building the items




The driver who my neighbor who I trust
suggested took me to the airport.
The driver who my neighbor who John
trusts suggested took me to the airport.
The driver who my neighbor who the
housekeeper trusts suggested took me to
the airport.
The driver who my neighbor who they trust
suggested took me to the airport.
Planning the experiment



Each set of four sentences constitutes a token
set (a.k.a. item)
Each item are four conditions (1/2 pronoun,
name, definite description, 3 pronoun).
Counterbalancing rules:



Each subject will judge no more than one sentence
from each token set.
Each subject will judge all conditions and will see
equal numbers of sentences from each condition
Every sentence in every token set will be judged by
some subject.
Trial lists

We have four conditions, so we need:
Four different “scripts” (versions of the lists)
 Some number of fourples of token sets.


E.g., items 1-4, each with conds a-d
Subj W: 1a, 2b, 3c, 4d
 Subj X: 1b, 2c, 3d, 4a
 Subj Y: 1c, 2d, 3a, 4b
 Subj Z: 1d, 2a, 3b, 4c

(script 1)
(script 2)
(script 3)
(script 4)
Our experiment

We will have 20 items (picked from the ones you
submitted) and 20 fillers.


(Note: That’s on the small side for a real experiment)
Next steps:





Create the lists of test sentences for the four different
scripts.
Spec out and pseudocode our experiment
Investigate PsyScript
Run the experiment
Deal with the data
Creating the scripts

Our sentences are made of very
predictable components:



The X who/that the Y who/that Z VP1 VP2
VP3
The only thing that changes across
conditions is Z, while the rest changes
across token sets.
We can use Excel to build these from their
pieces, to avoid unnecessary errors.
Worksheets

Components












Subj1
Rel1
Subj2
Rel2
Subj3a
Subj3b
Subj3c
VP3
VP2
VP1
Answer
Question

Fillers




Question
Answer
Regions…
The way I’ve set it up,
everything needs to
be exactly 8 regions
long (even the fillers)
Worksheets

Constructed

Computes item (token
group) and condition
based on row number,
comes up with a code
like I5V2 (fifth token
group, version 2).
Builds the sentence
region by region based
on the condition
number.

Tables



Keeps track of what
will be on each script.
Scripts are divided into
“blocks”, and each
block has one of each
condition and four
fillers, randomized.
Sort column is 2*block
plus a random number
(to order the blocks,
but randomly within)
Worksheets


Script


The master script sheet
This generates a script
based on the columns you
put into I1 and J1. (The
column refer to the tables
sheet, where the item and
condition numbers will be
found)




B and C for script 1
D and E for script 2
F and G for script 3
H and I for script 4
Script a, … script d






Actual scripts.
Select the part of script
sheet that has data
(A1:O41) and copy.
Go to script a sheet
Paste special… and
choose Value (so we don’t
copy formulas, only
results).
Delete column B-D (item,
cond, row), select rows 241, hit sort button, delete
column A (sort), and row 1
(labels)
Save as tab-delimited text.
The scripts are ready



So, we have the data that we’re going to
use.
The next thing is to figure out how we’re
going to test these.
The goal is to test reading time on each
region of the sentence by presenting the
sentence region by region.
Thinking through the
experiment

What do we want to have happen?






Display some instructions
Do some practice trials
Display “practice is over” message
Do some real trials
Display “thanks!”
The trials:




Show fully obscured sentence, wait for a key
Reveal next word, wait for a key, until done
Ask question, wait for response
Give sound feedback about correctness
PsyScript

To do this, we’ll use PsyScript, an
environment for creating psychology
experiments on the Mac.

(It’s basically the only freely available
software of this type that has promise for
working in the future – if PsyScope had
not become commercial as E-Prime, we’d
be learning that instead).
AppleScript

The underlying machinery behind PsyScript is
something called AppleScript.


This has been part of the Mac OS for about the past
10 years, although it is gaining power and popularity
recently.
AppleScript is a means by which you can tell
other programs what to do.


For example, tell Internet Explorer to go to a particular
web page, tell Word to create a new document and
type the date, …
Until you have an actual need for this, it doesn’t seem
very exciting…
AppleScript


AppleScript is a sophisticated high-level
programming language designed to be human
readable (and kind of human writable). It’s
supposed to look a lot like English.
PsyScript itself is an application that can be
bossed around by AppleScript, and has the
features that are useful in psycholinguistic
experiments, such as timing, drawing, input,
data recording functions.
Getting started

To write (and use) AppleScript, we use
Script Editor.

Easiest way to do this: Find the end
experiment script and double-click on it.
tell application “PsyScript”
end experiment
end tell
Note about PsyScript


PsyScript runs faster from the Script Editor
If you run PsyScript from the Script Editor you
have to manually tell it where your script is.


To do this, find the line that says tell fileHelper to
setContainer and change the thing in parentheses to
what you see when you Command-click on the name
of the script in the title bar of the Script Editor
Window, bottom to top, each separated by : and not
including the actual name of the script. E.g.,
setContainer(“Station 5:Desktop Folder:PsyScript:”)
Movingwindow

I wrote a script called movingwindow to do what
we’re going to do today.

The stimuli and instructions files are in a folder called
“resources” in the same folder as the script is. The
names of these files are set at the top of the script, in
mine, they are:






Mwstimuli.txt : sentence list as exported from Excel (tabdelimited text, exporting e.g., script a)
Mwpractice.txt : sentence list for the practice items
Mwinstruc.txt : initial instructions
Mwready.txt : post-practice instructions
Mwthanks.txt : end of experiment debriefing.
Results are stored in “results” folder.
Sentence lists

To generate the sentence lists in the right format
for movingwindow, go to one of the script a-d
pages, do Save As… from Excel, and choose
tab-delimited text.


Columns should be code, question, answer, sentence
(in eight columns)
The end results will come out in a file that you
can load back into Excel (a tab-delimited file):

Columns are: code, region number, time for region,
correct answer 1/0, text of region










Descargar

GRS LX 700 Language Acquisition and Linguistic Theory