Transcript The Nature of Words in Human Protolanguages

Protolanguage:
Between Holophrastic and
Atomic Meanings
Mike Dowman
University of Tokyo
January 23, 2007
Protolanguage
• At some time in the past humans didn’t
have language.
• What did the first human languages look
like?
• Were there simple languages before
complex ones?
Starting Small
• The first humans could only articulate
and/or perceive a limited number of
distinct words
No phonology
Little or no syntax
They only tried to convey a limited
number of simple meanings
Bickerton (1990, 1996)
• Words in the first languages were like
modern words
• Labeled preexisting concepts and
entities
• Words used in short strings
No fixed word order
Words can be omitted
Modern Day Examples of
Protolanguage
•
•
•
•
Children under two
Speakers of pidgins
Adults deprived of language in childhood
Trained apes:
Nim eat Nim eat
Tickle me Nim play
Me banana you banana me you give
Banana me me me eat
Wray (1998)
• The First Words in Protolanguages were
Holophrastic
Each word conveyed a whole complex
meaning (e.g. Give me the meat)
• Chimpanzee vocal noises and gestures are
holistic
Inform about location of food, threaten, get
another chimpanzee to do something
Syntactic Abilities
• Word categories
• Permissible and ungrammatical
sequences
• Mapping from structures to meanings
• Breaking meanings into bits
Making sure all the bits are expressed
once
No bits are expressed more than once
The Agents have none of
these Syntactic Abilities
Can only associate words and meanings
Can’t associate part of a meaning with
part of an utterance
When speaking just choose the word(s)
that seem to be most appropriate given
the whole proposition to be expressed
The model is therefore of protolanguage
only
Not the transition to syntax
The Iterated Learning Model
• One agent per generation
• Each agent sees meanings and hears
words produced by the agent in the
previous generation
• Agents’ knowledge of language is just a
list of meanings they’ve heard each
word used to express
Nature of Utterances
• Each utterance consisted of only a
single word
• Agents could use only a small number
of words (limited communicative
capacity)
• Available words fixed throughout each
simulation
Meaning Representations
• Small number of semantic primitives
(limited conceptual capacity)
• Each utterance would try to express a
complex meaning represented by a
group of three different primitives
hunt, pig, forest means ‘Hunt pigs in the
forest’
dog, pig, sleep means ‘Dogs and pigs are
sleeping’
Agents’ Choice of Words
• Agents use the word whose past uses
have been most similar to the current
meaning
• If current meaning is: eat, house, pig
• Previously heard used to express:
eat, house, dog
hunt, forest, pig
Match = 4/9
hunt, forest, pig
Agents’ Choice of Words
• If degree of match is one, agent will
always use that word
• Otherwise will use any available unused
word (and the word-meaning pair
remembered)
• Otherwise pick highest degree of match
(choosing at random in the event of a
tie)
Simulations
• Each meaning contained 3 elements from a
set of 10 (so 120 distinct complex meanings)
• Each agent produced 1000 utterances for the
agent in the next generation (each expressing
a randomly chosen meaning)
• Simulations were run for ten generations
• The number of available words was varied
Emergence of Words with
Atomic Meanings
• In the first simulation the agents could only
use 10 distinct words
• All the agents made use of all 10 available
words
• Most words were used only when one
particular semantic element was present
Their meanings appear to correspond to
those elements
Example Atomic Type Words
Meanings Expressed by
Word (frequency in
brackets)
forest, pig, eat (12)
eat, dog, gather (10)
forest, gather, eat (11)
hunt, eat, pig (11)
fight, pig, eat (6)
gather, eat, pig (7)
house, eat, gather (9)
dog, fight, eat (7)
eat, hunt, fight (11)
gather, eat, fight (10)
gather, eat, sleep (13)
dog, eat, pig (4)
gather, hunt, eat (4)
Description of Word
Meaning
Word denotes eat
Meanings Expressed by Word
(frequency in brackets)
dog, river, forest (7)
pig, forest, dog (10)
river, dog, pig (17)
eat, dog, forest (10)
dog, forest, gather (5)
dog, pig,fight(7)
forest, dog, fight (6)
fight, river, dog (10)
forest, sleep, dog (5)
sleep, dog, river (5)
Description of Word Meaning
Word denotes dog
The words are like
modern nouns or
verbs
Emergence of Holophrastic
Words
• In the next simulation the agents could
use 150 distinct words
• All agents used 120 of these words
• Each word expresses a single complex
meaning
These words are all holophrastic
Example Holophrastic Words
Meanings Expressed by Word
Description of Word Meaning
(frequency in brackets)
sleep, pig, hunt (5)
Word denotes events or objects
containing all the elements sleep,
pig and hunt
pig, house, fight (9)
Word denotes events or objects
containing all the elements pig,
houseand fight
forest, fight, hunt(10)
Word denotes events or objects
containing all the elements forest,
fight and hunt
Emergence of Intermediate
Types of Language
• What happens when the number of available
words is in between the number of semantic
elements and the number of complex
meanings?
• Do we get a mixture of atomic and
holophrastic words?
A new simulation with 50 available words
tested this
All agents used all the available words
Meanings Expressed by Word
Description of Word Meaning
(frequency in brackets)
sleep, forest, gather (7)
Word denotes events or objects
containing all the elements sleep,
forest and gather
forest, river, house (5)
Word denotes river, but only in
house,river, fight (8)
relation to forest, house or fight
forest, river, fight(6)
sleep, hunt, eat (13)
Word denotes events or objects
sleep, eat, river (11)
containingboth of the elements
pig, eat, sleep (8)
sleep and eat
eat, gather, sleep (11)
sleep, fight, eat (6)
sleep, dog, eat (9)
forest, sleep, eat (6)
gather, sleep, pig (10)
Word denotes events or objects
pig, dog,sleep (10)
containingboth of the elements
pig, sleep, house (5)
sleep and pig, but only in relation
to gather, dog or house
dog, hunt, hous
e (5)
Denotes meanings containing
hunt, eat, dog (6)
three of the elements dog, hunt,
eat, dog, house (12)
houseand eat
house, hunt, eat (5)
Language contains
words with
varying degrees
of holophrasticity
It’s intermediate
between a
protolanguage
with atomic
words and a
holophrastic one
Varying Degrees of
Holophrasticity
• The frequency of each type of word depended on
the number of distinct words available
• With 50 words there were:
10 holophrastic words, 35 words containing two
fixed semantic elements, 4 words containing one
fixed element, and 1word containing no fixed
elements at all
With a smaller number of distinct words, the
languages become less holophrastic
With more available words they became more
holophrastic
Fewer Words than Semantic
Elements
• What if the agents could not even
produce one distinct word for each
semantic element?
• (Or alternatively what if they knew so
many semantic elements they could use
more than there were words?)
• A new simulation was conducted with
only 5 distinct words available
Emergent Words
• Three words had one fixed semantic
element
• One word used any three of the other seven
semantic elements
• The final word contains at least two of the
elements eat, gather and pig
This is a new type of word, and is partly
holophrastic
Co-evolution of Agents and
Protolanguages
• What happens as agents
communicative and conceptual
capacities evolve phylogenetically?
• We would expect the agents’
protolanguages to rapidly adapt to the
agents new capacities
• How will the degree of holophrasticity
change over time?
Increasing Communicative
Capacity
• The first agents could use only a single
word
• After every 10 generations the number
of words they could use was increased
by 1
• Number of semantic elements fixed at
10
• 1300 generations simulated
Increasing Conceptual
Complexity
• In this simulation there were always 120
words available
• Initially there were 10 semantic elements
• After every 10 generations the number of
available semantic elements was increased
by 1
• This simulation was also run over 1300
generations (so there were 447,580 different
complex meanings at the end of the
simulation)
Transitions between Atomic
and Holophrastic Words
• The first simulation showed a progression
from atomic words to holophrasis
• The second from holophrasis to atomic words
So depending on the relative rates of monotonic
evolution of communicative and conceptual
capacity we could see multiple swings
between each type of protolanguage
Multiple Word Utterances
Did the first protolanguage utterances contain single or
multiple words?
protolanguag
e
gorillas
humans
chimps
multi-word proto-languages
Multiple Word Utterances
• Maximum length of utterances was fixed at 3 words
• If agents knew a word whose past uses exactly matched
the current meaning they used it (and nothing else)
• Otherwise if a new word was available they used that
(and remembered it)
• Finally, if none of the above apply, the agents chose the
3 closest matches to the current meaning.
Each of these words was then paired with the current
meaning by the hearing agent
Common
Elements
Number of Words
10
50
90
150
0
3
3
6
1
7
9
8
38
71
2
3
5
120
• With 10 words, some express all or most meaning
elements equally
• Many fewer fully holophrastic words
• More words with 2 common meanings than distinct 2
element combinations (45)
E-Language and I-Language
• If we look at agents’ internal linguistic
representations (I-Language) we only see an
unanalyzed list of the past uses of a word
• However, if we could only observe the agents’
speech (E-Language) and the meanings they tried
to express we would (in some cases) be able to
identify the words with specific semantic elements
So a categorization ability is not necessary for a
language with categorical words to emerge
Are there Atomic
Pre-Linguistic Concepts?
• Isn’t boy a holophrase for young male
human?
• And eat a holophrase for ingest solid
stuff?
Isn’t it arbitrary what we consider to be
atomic?
Also some proposed holophrases contain
far more concepts than others
• So is there a clear idea of what makes a
holophrase either?
3 Elements Per Meaning is
Arbitrary
New simulations:
• 4 elements per meaning
• out of a total of 20 meaning elements
• 3 words per utterance
• 10,000 word bottleneck
(4845 possible distinct meanings)
Types of Words Emerging
Common
Elements
10
Number of Words
20
200
1000
0
8
4
77
11
1
2
16
23
32
100
164
2
3
4
5000
793
4229
Conclusions
• Atomic words or holophrases are not the only
possibilities for protolanguages.
• If agents try to express a limited number of meanings
(relative to their communicative abilities) holophrasis
will results.
• With more meanings or fewer words forms, words
become less holophrastic and more like atomic words.
 The atomic word vs. holophrasis debate is not as
clear cut as it might at first seem