Title: Adventures with Camille, a Computational Simulation of a Minimalist Language Learner
1Adventures with Camille, a Computational
Simulation of a Minimalist Language Learner
- Peter W. Culicover (The Ohio State University)
- collaborating with
- Andrzej Nowak (Warsaw / FAU )
- Wojciech Borkowski (Warsaw)
- Piotr Kochanski (Warsaw)
2- Thanks to the James S. Mcdonnell Foundation, the
Center for Cognitive Science at OSU the Center
for Complex Systems at U. of Warsaw for their
support.
3Questions
- How should we think about what knowledge of
language consists of? - What is the architecture of the learner that
explains how we have this knowledge? - Can we simulate language acquisition to explore
this view?
4Outline
- Perspective Simpler Syntax the
Constructionalist Manifesto - Facts Syntactic Nuts and the Architecture of the
Language Faculty - Learner CAMiLLe and Dynamical Grammar
computational simulation - Social network simulation
5Perspectives
- On the standard view in linguistic theory, a
generative grammar of a language is an idealized
characterization of the linguistic knowledge, or
linguistic competence, of an idealized native
speaker. - A theory of generative grammar in turn is a
characterization of possible generative grammars.
6The standard view
- On this view, a generative grammar is a theory
something that really exists, the linguistic
competence of the native speaker. - And the theory of grammar is a theory about the
language faculty that explains why linguistic
competence takes the form that it does. - Both the grammar and the theory of grammar are in
some sense "in the mind" and form a part of the
explanation of human linguistic abilities.
7An alternative view
- What resides in the language faculty is actually
something of a different sort that reflects in
its architecture the dynamical character of
language. - In particular, it directly reflects the fact that
the grammar is acquired over time, that it is a
psychological mechanism used for speaking and
understanding, and that it undergoes change over
time.
8The challenge...
- ...for an approach to human language that does
not incorporate a standard version of generative
grammar is to account for the linguistic
phenomena themselves. - Such an alternative must be able to explain what
generative grammar explains, and more.
9The challenge, contd
- Minimally, the linguistic content of the
alternative formulation must be at least
equivalent to the content of the generative
grammar. - That is, both must provide the same (or
equivalent) empirically adequate structural
descriptions of words, phrases, idioms,
constructions and sentences. - Moreover, such an account should explain the
dynamical properties of language.
10A Constructionist Manifesto
- The Neural Basis of Cognitive Development A
Constructionist Manifesto. S. R. Quartz and T. J.
Sejnowski. 1997. Behavioural and Brain Sciences
20(4) 537 -596. - In contrast to learning as selective induction,
the central component of the constructivist model
is that it does not involve a search through an a
priori defined hypothesis space, and so is not an
instance of model-based estimation, or parametric
regression. Instead, the constructivist learner
builds this hypothesis as a process of
activity-dependent construction of the
representations that underlie mature skills.
11Constructionalism
- We adopt a Jackendoffian perspective on grammar
(e.g. Culicover, Syntactic Nuts, Oxford, 1999
Jackendoff, Foundations of Language, Oxford,
2001 Culicover Jackendoff, Simpler Syntax
Oxford, to appear), which is a constructionalist
one - a. The job of grammar is to describe the
sound-meaning correspondences. - b. Some of these correspondences are
unanalyzable (words). - c. Some have linguistic structure but are simple
or not entirely transparent on the meaning side
(idioms) (no nice structure/meaning matchups). - d. Some have structure and are transparent on
the meaning side (compositional semantics
interpreting canonical phrase structure). - e. Some are a combination of the above
('constructions'), ranging from quasi-idioms,
double-objects, movement along a path
expressions, syntactic nuts (see above), various
operator-trace binding constructions, etc. Each
has some degree of predictability and generality,
and some idiosyncrasies.
12Simple Syntax
- Simple(r) Syntax Hypothesis (SSH)
- The most explanatory syntactic theory is one that
imputes the minimum structure necessary to
mediate between sound and meaning.
13The Correspondence Spectrum
- Ill go through these quickly just to give you a
feel for what Im talking about...
14Words
- Many words are unanalyzable correspondences
between sound and meaning. - (although some (e.g. Hale Keyser) have argued
that apparently simple words are syntactically
complex and are the product of derivations
involving movement and deletion.) - (but the relations captured by such derivations
can be captured in non-derivational
(constructionist) ways, and the latter are
required for certain aspects of the
correspondences.)
15Idioms
- Idioms have recognizable syntactic structure but
unpredictable meaning - by and large
- lo and behold
- beat a dead horse
- make amends
- cast aspersions on (at / to)
- a flash in the pan
- put up with
16VP constructional idioms
- a. Way-construction (Jackendoff 1990, Goldberg
1995) - Elmer hobbled/laughed/joked his way to the
bank. - (? Elmer went/made his way to the bank
hobbling/laughing /joking) - b. Time-away construction (Jackendoff 1997b)
- Hermione slept/drank/sewed/programmed three
whole evenings away. - (? Hermione spent three whole evenings
sleeping/drinking/sewing /programming) - c. Soundmotion construction (Levin and
Rappaport Hovav 1995) - The car whizzed/rumbled/squealed past Harry.
- (? the car went past Harry, making
whizzing/rumbling/squealing noises) - d. Resultative construction
- The chef cooked the pot black.
- (? the chef made the pot black by cooking
in/with it)
17Syntax-semantics mismatches
- All of these constructions share the same basic
syntax (not surprisingly, since they are all
English) what is idiosyncratic is the way in
which their meanings are related to the meanings
of the parts and to the structure in which they
(the parts) appear.
18Motto Construction of language produces
constructions in language
- which means...
- as knowledge of language is constructed
dynamically by a learner, - what emerges are constructions that may
ultimately become rules, but only if given
enough evidence and a suitable generalization
mechanism, - otherwise, they remain constructions.
19How does the learner know what she/he is dealing
with?
- Since there is no way for the learner to know
where on the spectrum a correspondence really is,
the conservative strategy is to start at the
word/idiom end, and then move away as the weight
of the evidence warrants generalization.
(Tomasello)
20CAMiLLe
- C onservative (or Concrete)
- (dont generalize much beyond the evidence)
- A ttentive
- (all input is potentially relevant)
- Mi nimalist
- L anguage
- Le arner
21Goals of CAMiLLe
- Pursuing the logic of Concrete Minimalism (and
Simpler Syntax), we constructed CAMiLLe with
minimal prior knowledge of linguistic structure. - Language acquisition by CAMiLLe is intended to
simulate the formation of trajectories and flows,
and self-organization, in a dynamical system. - Our experiments with CAMiLLe are intended to
determine how much grammatical knowledge such a
minimalist learner is capable of acquiring
strictly from sound/meaning pairings.
22The spatial metaphorRegions categories/features
(slice of a many dimensional space)
- Sound/meaning pairs are computed as the system
changes states, represented as trajectories in a
space. - Structure guides movement from one region of the
space to another.
23N -gt A N
24(No Transcript)
25N -gt A N
Recursion
26Flow
- Multiple trajectories from one region to another
create flows. - If the flow is restricted in the space it is a
construction.
27Flow -gt Rule
- The individual trajectories may carve out a broad
region of the entire trajectory space. - If there are enough of them, we could fill in the
empty spaces make a rule.
28Self-organization
- CAMiLLe, as a dynamical system, should
self-organize when it is possible to collapse
individual rules or representations. This
produces generalization and over-generalization. - (Self-organization is limited in the current
implementation of CAMiLLe, it should be noted.)
29Concrete minimalism
- The computational system should be maximally
simple - not in terms of abstract computational
simplicity, - but in terms of the criterion of
- learning on the basis of the concrete
evidence. - That is, it should be the simplest system that
can arrive at an adequate account of the language
given a large but finite sample of experience.
(Simpler Syntax)
30Strategy
- Pursuing the logic of Concrete Minimalism, we
constructed CAMiLLe with minimal prior knowledge
of linguistic structure. - Our experiments with CAMiLLe are intended to
determine how much grammatical knowledge such a
minimalist learner is capable of acquiring.
31- Success would be nice, but some failure can be
quite informative.
32A Concrete Minimalist Learner
- We assume that the learner has access to
- sounds and the phonological system
- meanings and the Conceptual Structure system
- the minimal prior information about grammar that
is necessary to acquire a descriptively adequate
grammar given the paired sounds and meanings of a
language. - To begin, we assume, counterfactually, that this
information is ZERO, and see what needs to be
added.
33Representations in CAMiLLe
- There are two representational systems in
CAMiLLe - the system that encodes meaning,
- and the system that encodes form.
- Meaning is encoded as a structured list of
arguments and adjuncts, where thematic roles and
modifiers are explicitly specified. - Syntax (form) is simply the linear arrangement of
elements (words and morphemes)
34Meaning
- A meaning in the CS presented to CAMiLLe is
expressed in a simple attribute-value language.
E.g., - TOUCH(AGENTMAN,THEMEANIMAL)
- Relations typically expressed by verbs are
represented as constants with an associated
argument structure. - Arguments are given as thematic roles with their
values. (Like AGENTMAN) - We assume that the meaning that CAMiLLe is
presented with contains only primitives that are
cognitively accessible to CAMiLLe at a given
stage of development.
35Cognitive Development
- At the earliest stage of development CAMiLLe
(simulating an actual child) may only perceive
that some man touches some animal. - For example, John touches the cat could have the
meaning - TOUCH(AGENTMAN,THEMEANIMAL)
- at an early stage.
36- Meanings become more sophisticated as a
consequence of development of cognition and
perception. - E.g., later, the learner may perceive that there
is John, a distinct male person, that there is a
particular type of animal (a cat), that both are
singular in this context, and that they
participate in this relation. - TOUCH(AGENTJOHN(TYPEPERSON,
- GENDERMALE,NUMSG),
- THEMECAT(TYPEANIMAL,NUMSG))
37Capacities of CAMiLLe
- CAMiLLe is conservative, in that it does not form
hypotheses for which it does not have some
evidence. - In this implementation, CAMiLLe is assumed to
have prior knowledge of what the words are in a
sentence. It does not perform word segmentation
(although in principle it could).
38Knowledge about categories
- CAMiLLe must know that lexical categories exist
(but not which ones) and tries to determine what
categories there are (that is, which words are
similar in syntactic, semantic or morphological
characteristics). - CAMiLLe will generalize elements into a category
when it appears that they share sufficiently many
characteristics. - Distributional characteristics
- Similar meaning (e.g. cat dog both refer to
similar animals) - CAMiLLe does not know about the specific
syntactic categories such as Noun, Verb,
Adjective, etc.
39Knowledge about structure
- CAMiLLe must know that there are heads and
phrases and knows that a phrase consists of at
least a head and possibly other material that
bears some relationship to it. - there does not appear to be evidence in the raw
data that would tell you that there are such
things if you werent looking for them. - Syntactic Nuts (Culicover 1999) A lot of core
syntactic structure is in CS. Unpredictable
structure and particular linear relationships are
in the correspondence rules specific to the
language.
40Ignorance is bliss
- CAMiLLe is a concrete minimalist, in that it only
makes use of information about the linear order
of formatives and corresponding meanings
presented to it in the course of learning. - CAMiLLe does not know about functional heads.
There are no purely grammatical formatives in
CAMiLLes implementation that CAMiLLe tries to
match against linguistic input. - CAMiLLe does not know about transformations per
se. - And lots more...
41More bliss
- CAMiLLe does not know about traces or other empty
categories (but should, and will, in the next
implementation). - CAMiLLe does not have grammatical indices.
- CAMiLLe does not know about constraints, such as
Subjacency and the ECP, and in fact lacks all
comparable prior knowledge of syntactic theory. - Where do these come from?
42Even more bliss
- CAMiLLe cannot compute morphological structure,
and must have morphological structure presented
to it explicitly in order to make use of it in
forming grammatical hypotheses. - CAMiLLe does not know about classical X-bar
theory. That is, CAMiLLe does not know about
specifiers and complements per se, zero-level and
maximal projections, and so on. - CAMiLLe does not know about government.
43Correspondences
- The primary task of CAMiLLe is to construct
correspondences, that is, mappings of strings of
words into meanings (... based purely on the
pairing of strings and meanings). - dog ? DOG(TYPEANIMAL)
- see Robin ? SEE(THEMEROBIN)
- see the big bird ? SEE(THEMEBIRD(ATTRIBBIG,RE
FDEF)) - see category ? SEE(THEMECATEGORY)
44Rule formation
- CAMiLLe incorporates the information provided by
every sentence into a new rule or into an update
of existing rules. - The relevant information consists of the linear
ordering of words and other formatives in the
string . - The meaning of each sentence is compared with all
existing rules in terms of the meaning features
mentioned in the rule. - CAMiLLe extracts those features of meaning that
are possibly relevant to the correspondence.
45CAMiLLe constructs...
- word/meaning correspondences
- string/meaning correspondences (idioms)
- category/category correspondences (based on
identical or very similar distribution) - category/category correspondences in strings
(limited constructions or templates)
46- But to go beyond templates, CAMiLLe needs to
generalize. - And to learn the grammar of English, CAMiLLe
needs a more sophisticated understanding of what
kinds of relations may hold across a string of
words than it currently has. - We believe that this is an achievable goal.
47So where do constraints/universals come from?
- Competing formulations of the sound/meaning
correspondences in a social network
48The social network simulation
- Agents
- Social impact function
- Parameter Interaction partners
- Parameter Interaction distance
- Knowledge of language
- Features (in this case, 3)
- and feature values (2, 4, 8, etc)
- Noise (all other factors lumped together)
49Assumptions about learning
- Each learner interacts with a number of
individuals at each time t - Each learner is influenced by the individuals
that it interacts with dependent on their
relative strength. - Majority wins
- There is no review or evaluation by the learner
of its own internal state
50Display (start)
51Language distribution (start)
The Tower of Babel
52Strengths of individuals
53Gaps
- It is well-known that there are gaps in the
possible languages, that is, languages that are
logically possible but do not exist. Is this a
deep fact about cognition (explained by UG), or
not? - maybe, maybe not.
54Hypothesis
- Differential complexity of different ways of
expressing a CS introduces a bias into the
network against some alternatives - These will be disfavored in the network, perhaps
even disappear.
55Initial random distribution of feature values
56Initial population of the eight languages
57Distribution of languages and features after 150
steps
58Language distribution after 150 steps
59Some examples of what CAMiLLe does
60Identifying and categorizing nouns(Input Nouns
1, Eve)
- you xxx more cookies ? COOKIE(TYPEFOOD)
- how about another graham cracker ?
COOKIE(TYPEFOOD) - would that do just as well ?
- here .
- here you go .
- you have another cookie right on the table .
COOKIE(TYPEFOOD) - more juice ? JUICE(TYPEDRINK)
- would you like more grape juice ?
JUICE(TYPEDRINK) - where's your cup ? CUP(TYPEUTENSIL)
- oh I took it .
- I think that was Fraser . FRASER(TYPEPERSON)
- I'm not sure .
- what ?
- are you saying Fraser ? FRASER(TYPEPERSON)
- Mr Fraser ? FRASER(TYPEPERSON)
- yes that's much better .
- Mr Fraser ? FRASER(TYPEPERSON)
- what is that ?
- huh ?
61Sample of noun correspondences identified by
CAMilLLe in Nouns 1
- CHAIR ? chair
- CHAIR(TYPEFURNITURE) ? chair
- CLOTHES ? hat
- DIAPER ? diaper
- DIAPER(TYPECLOTHES) ? diaper
- FRASER ? Fraser
- FRASER(TYPEPERSON) ? Fraser
- FURNITURE ? chair
- HAT ? hat
- HAT(TYPECLOTHES) ? hat
- HEAD ? head
- HEAD(TYPEBODYPART) ? head
- JUICE ? juice
- JUICE(TYPEDRINK) ? juice
- PENCIL ? pencil
- PENCIL(TYPEUTENSIL) ? pencil
- PUDDING ? pudding
- PUDDING(TYPEFOOD) ? pudding
- STOOL ? stool
62Nouns exemplified in Nouns 1
- baby
- book
- bottle
- box
- chair
- cheese
- coffee
- cookie
- cup
- diaper
- duck
- Eve
- eye
- Fraser
- hat
-
- head
- juice
- man
- milk
- mommy
- paper
- pencil
- pudding
- radio
- shoe
- fly
- soldiers
- stool
- telephone
- train
- water
63Nouns learned in Nouns 1
- baby
- book
- bottle
- box
- chair
- cheese
- coffee
- cookie
- cup
- diaper
- duck
- Eve
- eye
- Fraser
- hat
-
- head
- juice
- man
- milk
- mommy
- paper
- pencil
- pudding
- radio
- shoe
- fly
- soldiers
- stool
- telephone
- train
- water
64Why CAMiLLe thinks that telephone is a sentence
final element
- well go and get your telephone .
TELEPHONE(TYPETOY) - yes he gave you your telephone .
TELEPHONE(TYPETOY) - yes that's the telephone . TELEPHONE(TYPETOY)
- that was the telephone . TELEPHONE(TYPETOY)
- it was Papa on the telephone .
TELEPHONE(TYPETOY) - yes the telephone . TELEPHONE(TYPETOY)
65Verb 2
1. a give b to c GIVE(AGENTA,THEMEB,RECIP
C) 2. does y give x to z GIVE(AGENTY,THEMEX
,RECIPZ) GIVE ? give to give2-gtto
givelt-2-gtto P1 GIVE ? 1.a 2.give 3.b
4.to 5.c 1.does 2.y 3.give 4.x 5.to
6.z P0.5 3. tell s to give r to q please
GIVE(AGENTS,THEMER,RECIPQ) 4. give j to e
now GIVE(THEMEJ,RECIPE) Same results,
stronger rules 5. c doesn't give k to a
NEG(GIVE(AGENTC,THEMEK,RECIPA)) 6. can q
give the w to s ? QU(GIVE(AGENTQ,THEMEW,RE
CIPS)) GIVE ? give to P1 GIVE ?
give to give2-gtto givelt-2-gtto P0.833333 (
And with richer input we get GIVE(RECIPE) ?
give to give2-gtto givelt-2-gtto e to1-gte
elt-1-gtto)
66Verbal inflection
- Input
- Mary have s sleep en soundly
SLEEP(AGENTMARY,ASPECTCOMPLETE,TIMENOW) - John be s really snore ing SNORE(AGENTJOHN,
ASPECTPROGRESSIVE,TIMENOW) - John probably have s fall en
FALL(EXPJOHN,ASPECTCOMPLETE,TIMENOW) - John might have s see en SEE(EXPJOHN,ASPEC
TCOMPLETE,TIMENOW) - John suddenly see ed it SEE(EXPJOHN,PAST)
- and then Mary fall ed FALL(THEMEMARY,PAST)
67Output (ranked in order of strength)
- JOHN MARY ?john mary
- COMPLETE ? have en
- NOW ? s
- PAST ? ed
- PROG ? be ing
- DIE(ASPECTCOMPLETE) ? die en die1-gten
dielt-1-gten - EAT(ASPECTCOMPLETE) ? 5.eat en eat1-gten
eatlt-1-gten - FALL(ASPECTPROG) ? fall ing fall1-gting
falllt-1-gting - FALL(ASPECTCOMPLETE) ? fall en fall1-gten
falllt-1-gten - FIND ? find s findlt-1-gts
- FIND(ASPECTCOMPLETE) ? have find have2-gtfind
findlt-2-gthave s s1-gtfind findlt-1-gts en
find1-gten findlt-1-gten - FIND(TIMENOW) ? find s findlt-1-gts
- JUMP(ASPECTPROG) ? jump ing jump1-gting
jumplt-1-gting - JUMP(ASPECTCOMPLETE) ? have jump have2-gtjump
havelt-2-gtjump s s1-gtjump jumplt-1-gts en
jump1-gten jumplt-1-gten - JUMP(TIMENOW) ? jump
- JUMP(TIMEPAST) ? jump ed jumplt-1-gted
- LOOK(ASPECTCOMPLETE) ? look en look1-gten
looklt-1-gten - LOOK(TIMENOW) ? look
- RUN(ASPECTCOMPLETE) ? run en run1-gten
runlt-1-gten
68- SEE(TIMENOW) ? see s seelt-1-gts
- SLEEP(ASPECTCOMPLETE) ? sleep en sleep1-gten
sleeplt-1-gten - SLEEP(ASPECTCOMPLETE PROG ? sleep en
ing sleep1-gten ing sleeplt-1-gten
ing - SLEEP(ASPECTPROG) ? sleep ing sleep1-gting
sleeplt-1-gting - SNORE(ASPECTPROG) ? snore ing snore1-gting
snorelt-1-gting - SNORE(ASPECTCOMPLETE) ? snore en snore1-gten
snorelt-1-gten - DIE LOOK(TIMENOW) ? die look en die
look1-gten die looklt-1-gten - DIE EAT LOOK ? die eat look en die
eat look 1-gten die eat look lt-1-gten - DIE FALL LOOK RUN SLEEP SNORE
(ASPECTCOMPLETE) ?die fall look run
sleep snore en die fall look run sleep
snore 1-gten die fall look run sleep
snore lt-1-gten
69- FALL SLEEP SNORE(ASPECTCOMPLETE PROG
? fall sleep snore en ing fall
sleep snore1-gten ing - FALL JUMP SLEEP SNORE (ASPECTPROG) ?
fall jump sleep snore ing fall jump
sleep snore 1-gting - FALL RUN (TIMENOW) ? 5.fall run
- FIND JUMP ? have find jump have2-gtfind
jump s s1-gtfind jump en find jump
1-gten - FIND JUMP(TIMENOW) ? have find jump
have2-gtfind jump s1-gtfind jump find
jump1-gten - FIND JUMP SLEEP (ASPECTCOMPLETE) ? have
find jump sleep have2-gtfind jump
sleep s1-gtfind jump sleep find jump
sleep 1-gten - FIND SEE (TIMENOW) ? find see s
- JUMP RUN (TIMEPAST) ? jump run ed
jump run lt-1-gted
70DP Structure
- Input, DP3
- See the man SEE(THEMEMAN(TYPEPERSON))
- See the woman SEE(THEMEWOMAN(TYPEPERSON))
- look at the baby LOOK(THEMEBABY(TYPEPERSON))
- that 's a cat CAT(TYPEANIMAL)
- and this is a dog DOG(TYPEANIMAL)
- show me the horse SHOW(THEMEHORSE(TYPEANIMAL
)) - what a tall man MAN(TYPEPERSON,ATTRTALL)
- he is a nice man MAN(TYPEPERSON,ATTRNICE)
- and this is a tall woman WOMAN(TYPEPERSON,ATT
RTALL) - look at the little baby LOOK(THEMEBABY(TYPEP
ERSON,ATTRLITTLE)) - see the nice cat SEE(THEMECAT(TYPEANIMAL,AT
TRNICE)) - see the big dog SEE(THEMEDOG(TYPEANIMAL,ATT
RBIG)) - and the big horse HORSE(TYPEANIMAL,ATTRBIG)
- that 's a little dog DOG(TYPEANIMAL,ATTRLITT
LE)
71Rule formed in DP3
- BABY DOG HORSE
- big little baby dog horsebiglittle1-
gtbaby dog horse
72Argument structure
- 1. 7704 SEE ?see
- 7. 392 SEE(EXPME) ? see . see1-gt
2-gt. i1-gt 2-gt see - 9. 250 SEE(THEMEIT YOU ? see it
you see1-gtit you - 10. 216 SEE(THEMEBECKY EVE ? see
see1-gtbecky eve - Rules for see based on 115 sentences with see
spoken to one actual child
73Word order argument structure
- Input
- 18. I see mary SEE(EXPME,THEMEMARY(TYPEPE
RSON)) - 19. do you see john ? SEE(EXPYOU,THEMEJOHN(
TYPEPERSON)) - 20. I see john SEE(EXPME,THEMEJOHN(TYPEPE
RSON)) - 21. and I see a boy too SEE(EXPME,THEMEBOY(
TYPEPERSON)) - Rules
- SEE(EXPME) ? i see i1-gtsee
- SEE(THEMEJOHN MARY ? see john mary
see1-gtjohn mary - SEE(THEMEBOY) ? see boy see2-gtboy
74Discontinuous dependencies
- Input
- you are say-ing something SAY(AGENTYOU,THEME
THING(REFINDEF)) - you will say something SAY(AGENTYOU,THEMETH
ING(REFINDEF)) - you did say something SAY(AGENTYOU,THEMETHI
NG(REFINDEF)) - you are say-ing nonsense SAY(AGENTYOU,THEME
NONSENSE) - you will say nonsense SAY(AGENTYOU,THEMENONS
ENSE) - you did say nonsense SAY(AGENTYOU,THEMENONSE
NSE) - what are you say-ing QU(SAY(AGENTYOU,THEME
WHTHING)) - what will you say QU(SAY(AGENTYOU,THEMEWH
THING)) - what did you say QU(SAY(AGENTYOU,THEMEWHT
HING)) - and similarly with do, eat
- Rules
- DO EAT SAY(THEMEWHTHING) ? 1.what 4.do
eat say
75Mock-Japanese wh-questions
- Input
- 400 sentences of the general form
- this ACC do IMP . IMP(DO(THEMETHING(REFDEF
))) - sing IMP . IMP(SING(AGENTPERSON))
- sleep PRES Q . QU(SLEEP(AGENTPERSON))
- what ACC do PAST Q . QU(DO(AGENTPERSON,THEM
EWHTHING)) - where TOP men NOM carrots ACC eat FUT Q .
QU(EAT(AGENTMAN,THEMECARROT,PLACEWHPLACE))
76Sample of rules formed
- QU ? q .
- SAY ? say X .
- DO ? do X .
- IMP ? imp .
- WHTHING ? what acc X X q .
- WOMAN ? woman1-gtnom
- SAY ? say
- EAT ? eat
- IMP QU ? imp q .
- DO ? do
- QU(DO EAT SLEEP ? do eat sleep X q .
- DO(THEMEWHTHING) ? what1-gtacc do2-gtq .
- CARROT ? carrots1-gtacc
- SAY(THEMEWHTHING) ? what1-gtacc q say
say2-gtq - EAT ? eat
- NONSENSE ? nonsense1-gtacc
- DO SAY(THEMEWHTHING) ? what1-gtacc q do
say2-gtq - DO EAT SAY ? .do eat say
- EAT(AGENTCHILD MAN WOMAN) ? child men
woman nom
77Scrambling
- hit bill-acc HIT(AGENTHIM,THEMEBILL)
- (He) hits Bill.
- bob-nom see SEE(EXPBOB)
- Bob sees.
- See jim-acc SEES(EXPHIM,THEMEJIM)
- (He) sees Jim.
- ...
- CAMiLLe correctly determines that the EXP role
of SEE is assigned to the (interpretation of the)
DP marked -nom, that the THEME role of SEE is
assigned to the (interpretation of the) DP marked
-acc, and that the AGENT role of HIT is assigned
to the (interpretation of the) DP marked -nom. - SEE(EXPARTHUR BOB) ? 1. arthur bob 2.nom
3.see - SEE(THEMEARTHUR BOB) ? arthur bob acc
arthur bob1-gtacc arthur boblt-1-gtacc - HIT(AGENTARTHUR BOB) ? 1.hit 2. arthur
bob 3.nom
78Inversion (English type)
- Input
- John JOHN(TYPEPERSON)
- Bird BIRD(TYPEANIMAL)
- John see s the bird SEE(EXPJOHN,THEMEBIRD)
- do s John see a bird QU(SEE(EXPJOHN,THEMEB
IRD)) - the bird see s John SEE(EXPBIRD,THEMEJOHN)
- do s the bird see John QU(SEE(EXPBIRD,THEME
JOHN)) - a bird see s John SEE(EXPBIRD,THEMEJOHN)
- ...
- Rules
- QU ? 1.do 2.s
79Inversion (non-English type)
- Input
- Sandy eat s the cake EAT(AGENTSANDY,THEMECA
KE) - eat s Sandy the cake QU(EAT(AGENTSANDY,THEM
ECAKE)) - Sandy come s COME(AGENTSANDY)
- come s Sandy QU(COME(AGENTSANDY))
- Sandy go s GO(AGENTSANDY)
- go s Sandy QU(GO(AGENTSANDY))
- ...
- Rules
- QU ? 2.s
80Imperatives
- Sample Input
- eat IMP(EAT(AGENTYOU))
- eat the cereal IMP(EAT(AGENTYOU,THEMECEREA
L)) - drink IMP(DRINK(AGENTYOU))
- drink the milk IMP(DRINK(AGENTYOU,THEMEMIL
K)) - Rules
- DRINK EAT(AGENTYOU) ? 1.drink eat
- HEAR SEE(EXPYOU) ? 1.hear see
- IMP(DRINK EAT HEAR LOOK SEE) ? 1.drink
eat hear look see
81Templates
- These are all templates they are not
correspondence rules that take into account
grammatical structure, and hence cannot be fully
general. - They are good for local constructions (like
inversion), but not for those that hold over
unbounded strings (like wh-questions and many
others).
82- CAMiLLe needs to know about grammatical structure
in order to find grammatical structure in the
input. - CAMiLLe needs to know about relations that hold
across syntactic structures in order to find such
relations in the input.