• Share
  • Email
  • Embed
  • Like
  • Save
  • Private Content
A Computational Model of Staged Language Acquisition
 

A Computational Model of Staged Language Acquisition

on

  • 772 views

This presentation was given at the Neurospin Center, CEA, Paris, France in 2009....

This presentation was given at the Neurospin Center, CEA, Paris, France in 2009.

It describes the main threads from my PhD thesis on the Computational Modelling of Staged Language Acquisition. Results from conducting symbolic simulations of language acquisition suggest that several modular, interconnected language acquisition devices may be at work in children's brains.

Statistics

Views

Total Views
772
Views on SlideShare
772
Embed Views
0

Actions

Likes
0
Downloads
11
Comments
0

0 Embeds 0

No embeds

Accessibility

Categories

Upload Details

Uploaded via as Adobe PDF

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Processing…
Post Comment
Edit your comment

    A Computational Model of Staged Language Acquisition A Computational Model of Staged Language Acquisition Presentation Transcript

    • A Computational Model of StagedLanguage Acquisition JACK, KrisDRT/FAR/LIST/DTSI/SRCI/LIC2M MrKrisJack@gmail.com
    • A Computational Model of StagedLanguage Acquisition Introduction Introduction ● Children appear to acquire language effortlessly Child Language ● They do not, however, do so overnight Models ● Typically, they progress through stages of LAT Testing linguistic development Results ● Computational modelling can help us to better Discussion understand the language acquisition process by Conclusions estimating the problem and developing possible solutions ●A computational model that tackles such staged linguistic development is absent from current literature ● Overview of presentation Kris Jack
    • A Computational Model of StagedLanguage Acquisition Stages in Language Acquisition Introduction ● How does child language typically develop? Child Language ● Language acquisition is consistently described in ➢ Pre-linguistic ➢ Holophrastic stages (e.g. Brown, Pinker, Tomasello) ➢ Early Multi-word ➢ Late Multi-word ● Five stages from birth to 48 months: ➢ Abstract ● The Pre-linguistic Stage Models LAT ● The Holophrastic Stage Testing ● The Early Multi-word Stage Results ● The Late Multi-word Stage Discussion Conclusions ● The Abstract Stage Holophrastic Late Multi-word Pre-linguistic Early Multi-word Abstract 0 6 12 18 24 30 36 42 48 Kris Jack Time (months)
    • A Computational Model of StagedLanguage Acquisition The Pre-linguistic Stage (1/2) Introduction Little activity typically characterised as linguistic Child Language Mini-stages including reflexive vocalisations, cooing, ➢ Pre-linguistic vocal play and babbling ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract Models LAT Testing Results Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Pre-linguistic Stage (2/2) Introduction Can differentiate between languages across rhythmic Child Language families of stress-timed, syllable-timed or mora-timed ➢ Pre-linguistic (Mehler et al., 1996) ➢ Holophrastic ➢ Early Multi-word Sensitive to transitional probabilities within syllable ➢ Late Multi-word sequences (Saffran et al., 1996) ➢ Abstract Models LAT Testing Results Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Holophrastic Stage (1/2) Introduction Begin when children achieve joint attention (Tomasello, Child Language 1995) ➢ Pre-linguistic First utterances are typically ➢ Holophrastic holophrastic ➢ Early Multi-word ➢ Late Multi-word  Holistic or atomic units ➢ Abstract (e.g. “mummy”, “doggy”) Models  Even seemingly multi- LAT word utterances are Testing holistic (e.g. “I-wanna- Results do-it” (Pine and Lieven, Discussion 1993)) Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Holophrastic Stage (2/2) Introduction Child and adult meanings for a holophrase often differ Child Language resulting in: ➢ Pre-linguistic  Underextensions (Reich, 1986) ➢ Holophrastic ➢ Early Multi-word  Overextensions (Barrett, 1978) ➢ Late Multi-word  Mismatches (Rodgon, 1976) ➢ Abstract Models LAT Testing Results Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Early Multi-word Stage Introduction Children regularly combine words to produce multi-word Child Language utterances ➢ Pre-linguistic Novel combinations are made (e.g. “allgone sticky” ➢ Holophrastic ➢ Early Multi-word (Braine, 1971)) ➢ Late Multi-word Many utterances can be described using a pivot grammar ➢ Abstract (P)ivot (O)pen, O P, O O (Braine, 1963) Models  E.g. S = P O, where O = “mummy”, “sticky”, “duck”, “red” LAT and P = “allgone” Testing Results Children are not sensitive Discussion to word-order (Clark, Conclusions 1975; MacWhinney, 1980; de Villiers and de Villiers, 1978) Function words and morphological markings tend to be omitted (Hyams, 1986) Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Late Multi-word Stage Introduction Child language becomes increasingly complex Child Language The emergence of syntactic awareness ➢ Pre-linguistic ➢ Holophrastic In English, word-order can define participant roles e.g. ➢ Early Multi-word “Make the doggie bite the cat” (de Villiers and de ➢ Late Multi-word Villiers, 1973) ➢ Abstract Models Children are found to have an irregular, item-based, knowledge of language: LAT Testing  Verb islands (Tomasello, 1992) Results  Inconsistent use of determiners (Pine and Lieven, 1997) Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Abstract Stage Introduction Evidence of all of the grammatical machinery found in Child Language adults (Pinker, 1994) ➢ Pre-linguistic The item-specific quality of child language gives way to a ➢ Holophrastic ➢ Early Multi-word more abstract quality (Tomasello, 2003) ➢ Late Multi-word Strong generative capacity asserts itself ➢ Abstract Models LAT Testing Results Discussion Conclusions Holophrastic Late Multi-word Pre-linguistic Early Multi-word Abstract 0 6 12 18 24 30 36 42 48 Kris Jack Time (months)
    • A Computational Model of StagedLanguage Acquisition Stages Summary Introduction ● Pre-linguistic Stage Child Language – little real linguistic activity ➢ Pre-linguistic ➢ Holophrastic ● Holophrastic Stage ➢ Early Multi-word ➢ Late Multi-word – first words ➢ Abstract ● Early Multi-word Stage Models LAT – first word combinations Testing ● Late Multi-word Stage Results Discussion – word combinations with syntax Conclusions ● Abstract Stage – strong generative capacity Holophrastic Late Multi-word Pre-linguistic Early Multi-word Abstract 0 6 12 18 24 30 36 42 48 Kris Jack Time (months)
    • A Computational Model of StagedLanguage Acquisition Goals Introduction ● What triggers the emergence of each stage? Child Language ● What accounts for the linguistic shape of each ➢ Pre-linguistic ➢ Holophrastic stage? ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract Models LAT Testing Results Discussion ● We can produce computational models that Conclusions estimate learning tasks faced by children to help us better understand the problem Kris Jack
    • A Computational Model of StagedLanguage Acquisition Language Models (1/2) Introduction Many computational models have been produced to Child Language study language learning Models The Miniature Language Acquisition Paradigm ➢ Data Filtering ➢ Incremental (Feldman et al., 1990) ➢ Iterated  Place a computational model in an environment with LAT access to visual and acoustic stimuli (simulated or Testing grounded) Results  Train the model by providing descriptions of visually- Discussion based scenes from a miniature language (e.g. “the Conclusions red square is on top of the green circle”)  The model is said to have acquired the language when it can both comprehend and produce all sentences within the miniature language Kris Jack
    • A Computational Model of StagedLanguage Acquisition Language Models (2/2) Introduction A computational model that demonstrates stage Child Language transitions from the Pre-linguistic Stage to the Models Abstract Stage is missing from current literature ➢ Data Filtering ➢ Incremental Some models demonstrate stage-like learning by: ➢ Iterated  Externally modifying the training data LAT Testing  Modifying the functionality of the model Results  Internally modifying the training data Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition Temporal Data Filtering Introduction Training data are modified during training (e.g. Child Language Elman, 1993; Roy, 2002) Models  Elman (1993) trained a neural network to acquire ➢ Data Filtering sentences with both short and long distance ➢ Incremental ➢ Iterated dependencies LAT  Success only when the training data were biased Testing towards including more examples of short distance dependencies in early learning and long distance Results dependencies in late learning Discussion Conclusions Data 0 < t <= x Data x < t <= y Data Data Model Model Data y < t <= z Data Unrealistic assumption  Although infant direct speech contains different characteristics to adult directed speech, children are still exposed to complex sentences from birth Kris Jack
    • A Computational Model of StagedLanguage Acquisition Incremental Learning Introduction Functionality is modified during training (Elman, Child Language 1993; Dominey and Boucher, 2005) Models  Elman (1993) trained a neural network to acquire ➢ Data Filtering sentences with both short and long distance ➢ Incremental ➢ Iterated dependencies LAT  The neural network’s ‘short-term memory’ was Testing incrementally increased during learning, allowing the network to acquire both types of dependencies Results Discussion Incremental learning has problems: Conclusions  When should increments be made?  What unit of data should be restricted, syllables  Transitions are not clean and clear t > 0 Module Module Data t > x Module Model Data Module Model t > y Module Module Kris Jack
    • A Computational Model of StagedLanguage Acquisition Iterated Learning Introduction Training data are modified by model Child Language  In modelling language evolution, Kirby (2002) shows Models that learning through cultural transmission can modify ➢ Data Filtering the structure of a language ➢ Incremental ➢ Iterated  One generation of agents learn a language and then LAT produce a progressively more structured language for Testing teaching to the next generation (Iterated Learning) Results However, this is language evolution, not language Discussion acquisition Conclusions  Training data are constant for children Data Data Data Data Data Data Model Model Model Model Model Model Model Model Kris Jack
    • A Computational Model of StagedLanguage Acquisition Problem Introduction Can we model the stages of language acquisition Child Language when: Models  The functionality of the model is kept constant AND ➢ Data Filtering ➢ Incremental  The training data provided to the model are constant? ➢ Iterated LAT Testing Results Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition Language Acquisition Toolkit Introduction LAT provides a framework for investigating staged Child Language language acquisition Models Aim: LAT ➢ Architecture  Develop a computational model that demonstrates ➢ Learning Env. realistic stages of linguistic development ➢ World Perception ➢ Cross-situational Investigating the problem within a Miniature ➢ Holophrastic Language Acquisition Framework where: ➢ Early Multi-word ➢ Late Multi-word  The language contains enough complexity to allow ➢ Abstract the model to demonstrate stages of development Testing  The language is not so complex that it cannot be Results learned in entirety Discussion  Concentrating on comprehension Conclusions Holophrastic Late Multi-word Pre-linguistic Early Multi-word Abstract 0 6 12 18 24 30 36 42 48 Kris Jack Time (months)
    • A Computational Model of StagedLanguage Acquisition LAT General Architecture Introduction Child Language Environment (real or Models simulated) LAT ➢ Architecture                            Sensory stimuli ➢ Learning Env. World ➢ World Perception L e a rn in g Perception Module C o m p re h e n s io n M o d u le s M o d u le s ➢ Cross-situational           renv ➢ Holophrastic renv ➢ Early Multi-word r1 l1 c1 ➢ Late Multi-word R e s o u rc e s r1 ➢ Abstract r1 r2 l2 c2 Testing r2 renv Results r2 r3 l3 r1 c3 Discussion r3 r2 . . Conclusions . . . . . . rx­1 . rx rx lx cx rx Kris Jack
    • A Computational Model of StagedLanguage Acquisition LAT Architecture Instantiated Introduction Child Language Environment (real or Models simulated) LAT ➢ Architecture                            Sensory stimuli ➢ Learning Env. World ➢ World Perception L e a rn in g Perception Module C o m p re h e n s io n M o d u le s M o d u le s ➢ Cross-situational           renv ➢ Holophrastic renv ➢ Early Multi-word lcross­sit R e s o u rc e s ➢ Late Multi-word rcross­sit ➢ Abstract rcross­sit rholophrastic lholophrastic cholophrastic Testing rholophrastic Results rholophrastic renv rearly learly cearly Discussion rearly rcross­sit Conclusions llate rholophrastic rholophrastic rearly rlate clate rlate rlate rlate rabstract rabstract labstract cabstract rabstract Kris Jack
    • A Computational Model of StagedLanguage Acquisition Learning Environment Introduction The model plays the Scene Building Game (Jack, Child Language 2005) Models  Algorithm: LAT ➢ Architecture 1.The model watches a scene containing a single ➢ Learning Env. geometric object ➢ World Perception 2.Another geometric object is added to the scene and ➢ Cross-situational the event is described ➢ Holophrastic ➢ Early Multi-word 3.Return to 1. ➢ Late Multi-word  Notice that the landmark object is described using the ➢ Abstract definite article “the” and the new object is described Testing using the indefinite article “a” Results Discussion Conclusions a blue circle below the red square Kris Jack
    • A Computational Model of StagedLanguage Acquisition The World Perception Module Introduction The World Perception Module encodes events (renv) Child Language Models LAT ➢ Architecture ➢ Learning Env. ➢ World Perception ➢ Cross-situational Simulated Visual input – detects colour, shape and ➢ Holophrastic relative positions ➢ Early Multi-word {below(rel), blue(2), circle(2), horizontal_even(rel), red(1), square(1)} ➢ Late Multi-word ➢ Abstract Simulated Acoustic input – event description is Testing perceived as a sequence of syllables Results a blue cir cle be low the red square Discussion Conclusions Joint attention is assumed from the outset Holophrastic Late Multi-word Pre-linguistic Early Multi-word Abstract 0 6 12 18 24 30 36 42 48 Kris Jack Time (months)
    • A Computational Model of StagedLanguage Acquisition Cross-situational Learning Module Introduction Aims Child Language  Find similarities between observed events Models LAT  Derive possible form-meaning pairs ➢ Architecture  Create new resource rcross-sit ➢ Learning Env. ➢ World Perception ➢ Cross-situational ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract Testing Results Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition Cross-situational Learning Introduction Method Child Language  Form of Cross-situational Analysis (Siskind, 1996) Models LAT  Words co-occur more often with their intended ➢ Architecture meanings than with other meanings ➢ Learning Env. Example ➢ World Perception ➢ Cross-situational  Equal string parts are found ➢ Holophrastic  Equal feature value parts are found ➢ Early Multi-word  New extensions are derived ➢ Late Multi-word ➢ Abstract 1) a blue cir cle be low the red square {below(rel), blue(2), circle(2), horizontal_even(rel), red(1), square(1)} Testing Results 2) a green star to the low er right of the blue tri ang gle Discussion {below(rel), blue(1), green(2), right(rel), star(2), triangle(1)} Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition Cross-situational Learning Introduction Method Child Language  Form of Cross-situational Analysis (Siskind, 1996) Models LAT  Words co-occur more often with their intended ➢ Architecture meanings than with other meanings ➢ Learning Env. Example ➢ World Perception ➢ Cross-situational  Equal string parts are found ➢ Holophrastic  Equal feature value parts are found ➢ Early Multi-word  New extensions are derived ➢ Late Multi-word ➢ Abstract 1) a blue cir cle be low the red square {below(rel), blue(2), circle(2), horizontal_even(rel), red(1), square(1)} Testing Results 2) a green star to the low er right of the blue tri ang gle Discussion {below(rel), blue(1), green(2), right(rel), star(2), triangle(1)} Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition Cross-situational Learning Introduction Method Child Language  Form of Cross-situational Analysis (Siskind, 1996) Models LAT  Words co-occur more often with their intended ➢ Architecture meanings than with other meanings ➢ Learning Env. Example ➢ World Perception ➢ Cross-situational  Equal string parts are found ➢ Holophrastic  Equal feature value parts are found ➢ Early Multi-word  New extensions are derived ➢ Late Multi-word ➢ Abstract 1) a blue cir cle be low the red square {below(rel), blue(2), circle(2), horizontal_even(rel), red(1), square(1)} Testing Results 2) a green star to the low er right of the blue tri ang gle Discussion {below(rel), blue(1), green(2), right(rel), star(2), triangle(1)} Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition Cross-situational Learning Introduction Method Child Language  Form of Cross-situational Analysis (Siskind, 1996) Models LAT  Words co-occur more often with their intended ➢ Architecture meanings than with other meanings ➢ Learning Env. Example ➢ World Perception ➢ Cross-situational  Equal string parts are found ➢ Holophrastic  Equal feature value parts are found ➢ Early Multi-word  New extensions are derived ➢ Late Multi-word ➢ Abstract 1) a blue cir cle be low the red square {below(rel), blue(2), circle(2), horizontal_even(rel), red(1), square(1)} Testing Results 2) a green star to the low er right of the blue tri ang gle Discussion {below(rel), blue(1), green(2), right(rel), star(2), triangle(1)} Conclusions a {below(rel), blue(1)} a {below(rel), blue(2)} the {below(rel), blue(1)} the {below(rel), blue(2)} rcross-sit blue {below(rel), blue(1)} Kris Jack blue {below(rel), blue(2)}
    • A Computational Model of StagedLanguage Acquisition Holophrastic Learning Module Introduction Aims Child Language  Reduce ambiguity by removing homonyms Models LAT  Reduce ambiguity by removing synonyms ➢ Architecture  Create new resource rholophrastic ➢ Learning Env. ➢ World Perception ➢ Cross-situational ➢ Holophrastic blue {blue(1,2)} ➢ Early Multi-word red {red(1,2)} ➢ Late Multi-word green {green(1,2)} ➢ Abstract square {square(1,2)} cir cle {circle(1,2)} Testing tri ang gle {triangle(1,2)} Results be low {below(rel), vertical_even(rel)} a bove {above(rel), vertical_even(rel)} Discussion blue square {blue(1,2), square(1,2)} Conclusions blue cir cle {blue(1,2), circle(1,2)} . . . Kris Jack
    • A Computational Model of StagedLanguage Acquisition Holophrastic Learning Module Introduction In order to remove homonyms and synonyms Child Language 1. create abstract extensions Models LAT (1) red {red(1)} ➢ Architecture is merged with ➢ Learning Env. ➢ World Perception (2) red {red(2)} ➢ Cross-situational to produce ➢ Holophrastic ➢ Early Multi-word (3) red {red(1, 2)} ➢ Late Multi-word ➢ Abstract 3. keep only the most similar meaning for each form Testing using Frequency  F , M  j i Results Similarity  M i , F j = Frequency  F j  Discussion Conclusions 3. erase meanings of all extensions that have similarities lower than other extensions with the same meaning, where similarity is Frequency  M j , Fi  Similarity  F i , M j = Frequency  M j  rholophrastic Kris Jack
    • A Computational Model of Staged Holophrastic Comprehension ModuleLanguage Acquisition Introduction Comprehension: Child Language Given a string to comprehend, the model searches Models rholophrastic for extensions that contain the string LAT ➢ Architecture From those found, the meaning of the extension that is ➢ Learning Env. most similar is returned ➢ World Perception ➢ Cross-situational ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract Testing Results Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition Early Multi-word Learning Module Introduction Aims Child Language  To find compositional relationships between form- Models meaning pairs in rholophrastic LAT ➢ Architecture show no sensitivity to word order nor object roles ➢ Learning Env. ➢ World Perception  Create new resource rearly ➢ Cross-situational ➢ Holophrastic ➢ Early Multi-word 1) blue cir cle 2) green cir cle ➢ Late Multi-word {blue(1,2), circle(1,2)} {circle(1,2), green(1,2)} ➢ Abstract Testing Results blue cir cle green cir cle Discussion {blue(1,2)} {circle(1,2)} {green(1,2)} {circle(1,2)} Conclusions 3) the red cir cle {circle(1), red(1)} the red cir cle {} {circle(1,2), red(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Early Multi-word Learning Module Introduction Finding compositionality Child Language an extension is a function of two other extensions when Models its LAT ➢ Architecture  form is equal to the concatenation of the forms of the parts ➢ Learning Env. (ignoring word order) ➢ World Perception  meaning is equal to the feature set union of the parts ➢ Cross-situational (ignoring object roles) ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract Testing Results Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition Early Multi-word Learning Module Introduction Finding compositionality Child Language an extension is a function of two other extensions when Models its LAT ➢ Architecture  form is equal to the concatenation of the forms of the parts ➢ Learning Env. (ignoring word order) ➢ World Perception  meaning is equal to the feature set union of the parts ➢ Cross-situational (ignoring object roles) ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word do extensions 1) 2) and 3) express a compositional grammar ➢ Abstract fragment? Testing Results 1) blue cir cle 2) blue 3) cir cle Discussion {blue(1,2), circle(1,2)} {blue(1,2)} {circle(1,2)} Conclusions blue cir cle = blue + cir cle ? blue cir cle = cir cle + blue ? Kris Jack
    • A Computational Model of StagedLanguage Acquisition Early Multi-word Learning Module Introduction Finding compositionality Child Language an extension is a function of two other extensions when Models its LAT ➢ Architecture  form is equal to the concatenation of the forms of the parts ➢ Learning Env. (ignoring word order) ➢ World Perception  meaning is equal to the feature set union of the parts ➢ Cross-situational (ignoring object roles) ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word do extensions 1) 2) and 3) express a compositional grammar ➢ Abstract fragment? Testing Results 1) blue cir cle 2) blue 3) cir cle Discussion {blue(1,2), circle(1,2)} {blue(1,2)} {circle(1,2)} Conclusions {blue(1,2), circle(1,2)} = {blue(1,2)} U {circle(1,2)} ? Kris Jack
    • A Computational Model of StagedLanguage Acquisition Early Multi-word Learning Module Introduction Finding compositionality Child Language an extension is a function of two other extensions when Models its LAT ➢ Architecture  form is equal to the concatenation of the forms of the parts ➢ Learning Env. (ignoring word order) ➢ World Perception  meaning is equal to the feature set union of the parts ➢ Cross-situational (ignoring object roles) ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word do extensions 1) 2) and 3) express a compositional grammar ➢ Abstract fragment? Testing Results 1) blue cir cle 1) blue cir cle Discussion {blue(1,2), circle(1,2)} {blue(1,2), circle(1,2)} Conclusions OR 2) blue 3) cir cle 3) cir cle 2) blue {blue(1,2)} {circle(1,2)} {circle(1,2)} {blue(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Early Multi-word Learning Module Introduction Finding compositionality Child Language To reflect child sensitivity during this period, each Models grammar fragment must have a part that appears in LAT another fragment ➢ Architecture ➢ Learning Env. This produces a form of pivot grammar where pivot parts ➢ World Perception can appear with many open parts (Braine, 1963) ➢ Cross-situational ➢ Holophrastic ➢ Early Multi-word 1) blue cir cle 2) green cir cle ➢ Late Multi-word {blue(1,2), circle(1,2)} {circle(1,2), green(1,2)} ➢ Abstract Testing blue cir cle green cir cle Results {blue(1,2)} {circle(1,2)} {green(1,2)} {circle(1,2)} Discussion Conclusions 3) the red cir cle {circle(1), red(1)} the red cir cle {} {circle(1,2), red(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Early Multi-word Learning Module Introduction Finding compositionality Child Language To reflect child sensitivity during this period, each Models grammar fragment must have a part that appears in LAT another fragment ➢ Architecture ➢ Learning Env. This produces a form of pivot grammar where pivot parts ➢ World Perception can appear with many open parts (Braine, 1963) ➢ Cross-situational ➢ Holophrastic ➢ Early Multi-word 1) blue cir cle 2) green cir cle ➢ Late Multi-word {blue(1,2), circle(1,2)} {circle(1,2), green(1,2)} ➢ Abstract Testing blue cir cle green cir cle Results {blue(1,2)} {circle(1,2)} {green(1,2)} {circle(1,2)} Discussion Conclusions 3) the red cir cle 3) {circle(1), red(1)} the red cir cle {} {circle(1,2), red(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Early Multi-word Learning Module Introduction Finding compositionality Child Language To reflect child sensitivity during this period, each Models grammar fragment must have a part that appears in LAT another fragment ➢ Architecture ➢ Learning Env. This produces a form of pivot grammar where pivot parts ➢ World Perception can appear with many open parts (Braine, 1963) ➢ Cross-situational ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract 1) blue cir cle 2) green cir cle Testing {blue(1,2), circle(1,2)} {circle(1,2), green(1,2)} Results Discussion Conclusions blue cir cle green cir cle {blue(1,2)} {circle(1,2)} {green(1,2)} {circle(1,2)} rearly Kris Jack
    • A Computational Model of Staged Early Multi-word Comprehension ModuleLanguage Acquisition Introduction Comprehension: Child Language Given a string to comprehend, the model searches rearly Models for grammar fragments whose parts can be combined LAT to make the string ➢ Architecture ➢ Learning Env. For each grammar fragment found, its meanings are ➢ World Perception combined through union and each result is returned ➢ Cross-situational ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word comprehend blue cir cle ➢ Abstract Testing blue cir cle Results {blue(1,2), circle(1,2)} Discussion Conclusions blue cir cle {blue(1,2)} {circle(1,2)} {blue(1,2)} U {circle(1,2)} = {blue(1,2), circle(1,2)} Kris Jack
    • A Computational Model of Staged Early Multi-word Comprehension ModuleLanguage Acquisition Introduction Comprehension: Child Language Given a string to comprehend, the model searches rearly Models for grammar fragments whose parts can be combined LAT to make the string ➢ Architecture ➢ Learning Env. For each grammar fragment found, its meanings are ➢ World Perception combined through union and each result is returned ➢ Cross-situational ➢ Holophrastic however... ➢ Early Multi-word ➢ Late Multi-word comprehend cir cle blue ➢ Abstract Testing blue cir cle Results {blue(1,2), circle(1,2)} Discussion Conclusions blue cir cle {blue(1,2)} {circle(1,2)} {blue(1,2)} U {circle(1,2)} = {blue(1,2), circle(1,2)} Kris Jack
    • A Computational Model of Staged Early Multi-word Comprehension ModuleLanguage Acquisition Introduction Comprehension: Child Language Given a string to comprehend, the model searches rearly Models for grammar fragments whose parts can be combined LAT to make the string ➢ Architecture ➢ Learning Env. For each grammar fragment found, its meanings are ➢ World Perception combined through union and each result is returned ➢ Cross-situational ➢ Holophrastic however... ➢ Early Multi-word ➢ Late Multi-word comprehend the red cir cle ➢ Abstract Testing the red cir cle Results {circle(1), red(1)} Discussion a blue cir cle be low the red Conclusions the red cir cle square {} {circle(1,2), red(1,2)} {} U {circle(1,2), red(1,2)} = {circle(1,2), red(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Late Multi-word Learning Module Introduction Aims Child Language  To find compositional relationships between form- Models meaning pairs in rholophrastic LAT ➢ Architecture show sensitivity to word order and object roles ➢ Learning Env. ➢ World Perception  Create new resource rlate ➢ Cross-situational ➢ Holophrastic 1) blue cir cle ➢ Early Multi-word {blue(1,2), circle(1,2)} ➢ Late Multi-word ((1,2)­>(1,2)) ((1,2)­>(1,2)) ➢ Abstract Testing blue cir cle {blue(1,2)} {circle(1,2)} Results Discussion Conclusions 2) the blue cir cle 3) a blue cir cle {blue(1), circle(1)} {blue(2), circle(2)} () ((1,2)­>(1)) () ((1,2)­>(2)) the blue cir cle a blue cir cle {} {blue(1,2), circle(1,2)} {} {circle(1,2), blue(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Late Multi-word Learning Module Introduction Finding compositionality Child Language an extension is a function of two other extensions when Models its LAT ➢ Architecture  form is equal to the concatenation of the forms of the parts ➢ Learning Env. (consider word order) ➢ World Perception  meaning is equal to the feature set union of the parts, after ➢ Cross-situational transfomation (consider object roles) ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract Testing Results Discussion Conclusions Kris Jack
    • A Computational Model of StagedLanguage Acquisition Late Multi-word Learning Module Introduction Finding compositionality Child Language an extension is a function of two other extensions when Models its LAT ➢ Architecture  form is equal to the concatenation of the forms of the parts ➢ Learning Env. (consider word order) ➢ World Perception  meaning is equal to the feature set union of the parts, after ➢ Cross-situational transfomation (consider object roles) ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word do extensions 1) 2) and 3) express a compositional grammar ➢ Abstract fragment? Testing Results 1) the blue cir cle 2) the 3) blue cir cle Discussion {blue(1), circle(1)} {} {blue(1,2), circle(1,2)} Conclusions blue cir cle = blue + cir cle ? Kris Jack
    • A Computational Model of StagedLanguage Acquisition Late Multi-word Learning Module Introduction Finding compositionality Child Language an extension is a function of two other extensions when Models its LAT ➢ Architecture  form is equal to the concatenation of the forms of the parts ➢ Learning Env. (consider word order) ➢ World Perception  meaning is equal to the feature set union of the parts, after ➢ Cross-situational transfomation (consider object roles) ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word do extensions 1) 2) and 3) express a compositional grammar ➢ Abstract fragment? Testing Results 1) the blue cir cle 2) the 3) blue cir cle Discussion {blue(1), circle(1)} {} {blue(1,2), circle(1,2)} Conclusions {blue(1), circle(1)} = T({}, ()) U T({blue(1,2), circle(1,2)}, ((1,2)­>(1))) ? i.e. {blue(1), circle(1)} = {} U {blue(1), circle(1)} ? Kris Jack
    • A Computational Model of StagedLanguage Acquisition Late Multi-word Learning Module Introduction Finding compositionality Child Language an extension is a function of two other extensions when Models its LAT ➢ Architecture  form is equal to the concatenation of the forms of the parts ➢ Learning Env. (consider word order) ➢ World Perception  meaning is equal to the feature set union of the parts, after ➢ Cross-situational transfomation (consider object roles) ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word do extensions 1) 2) and 3) express a compositional grammar ➢ Abstract fragment? Testing Results Discussion 1) the blue cir cle {blue(1), circle(1)} Conclusions () ((1,2)­>(1)) 2) the 3) blue cir cle {} {blue(1,2), circle(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Late Multi-word Learning Module Introduction Finding compositionality Child Language Each grammar fragment must have a part that appears in Models another fragment LAT ➢ Architecture  They appears on the same side (same word order) ➢ Learning Env.  AND the transformations are the same (same object ➢ World Perception ➢ Cross-situational roles) ➢ Holophrastic ➢ Early Multi-word 1) the blue cir cle 2) the blue square ➢ Late Multi-word {blue(1), circle(1)} {blue(1), square(1)} ➢ Abstract () ((1,2)­>(1)) () ((1,2)­>(1)) Testing Results the blue cir cle the blue square {} {blue(1,2), circle(1,2)} {} {blue(1,2), square(1,2)} Discussion Conclusions 3) a red cir cle {circle(2), red(2)} () ((1,2)­>(2)) a red cir cle {} {circle(1,2), red(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Late Multi-word Learning Module Introduction Finding compositionality Child Language Each grammar fragment must have a part that appears in Models another fragment LAT ➢ Architecture  They appears on the same side (same word order) ➢ Learning Env.  AND the transformations are the same (same object ➢ World Perception ➢ Cross-situational roles) ➢ Holophrastic ➢ Early Multi-word 1) the blue cir cle 2) the blue square ➢ Late Multi-word {blue(1), circle(1)} {blue(1), square(1)} ➢ Abstract () ((1,2)­>(1)) () ((1,2)­>(1)) Testing Results the blue cir cle the blue square {} {blue(1,2), circle(1,2)} {} {blue(1,2), square(1,2)} Discussion Conclusions 3) a red cir cle {circle(2), red(2)} () ((1,2)­>(2)) a red cir cle {} {circle(1,2), red(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Late Multi-word Learning Module Introduction Finding compositionality Child Language Each grammar fragment must have a part that appears in Models another fragment LAT ➢ Architecture  They appears on the same side (same word order) ➢ Learning Env.  AND the transformations are the same (same object ➢ World Perception ➢ Cross-situational roles) ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract 1) the blue cir cle 2) the blue square Testing {blue(1), circle(1)} {blue(1), square(1)} Results () ((1,2)­>(1)) () ((1,2)­>(1)) Discussion the blue cir cle the blue square Conclusions {} {blue(1,2), circle(1,2)} {} {blue(1,2), square(1,2)} rlate Kris Jack
    • A Computational Model of Staged Late Multi-word Comprehension ModuleLanguage Acquisition Introduction Comprehension: Child Language Given a string to comprehend, the model searches rlate for Models grammar fragments whose parts can be combined to LAT make the string ➢ Architecture ➢ Learning Env. For each grammar fragment found, its meanings are ➢ World Perception mapped and then combined through union and each ➢ Cross-situational result is returned ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word comprehend blue cir cle ➢ Abstract Testing blue cir cle {blue(1,2), circle(1,2)} Results Discussion ((1,2)­>(1,2)) ((1,2)­>(1,2)) Conclusions blue cir cle {blue(1,2)} {circle(1,2)} T({blue(1,2)},((1,2)­>(1,2))) U T({circle(1,2)},((1,2)­>(1,2)))  = {blue(1,2), circle(1,2)} Kris Jack
    • A Computational Model of Staged Late Multi-word Comprehension ModuleLanguage Acquisition Introduction Comprehension: Child Language Given a string to comprehend, the model searches rlate for Models grammar fragments whose parts can be combined to LAT make the string ➢ Architecture ➢ Learning Env. For each grammar fragment found, its meanings are ➢ World Perception mapped and then combined through union and each ➢ Cross-situational result is returned ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word comprehend cir cle blue ➢ Abstract Testing blue cir cle {blue(1,2), circle(1,2)} Results Discussion ((1,2)­>(1,2)) ((1,2)­>(1,2)) Conclusions blue cir cle {blue(1,2)} {circle(1,2)} Meaning not found Kris Jack
    • A Computational Model of Staged Late Multi-word Comprehension ModuleLanguage Acquisition Introduction Comprehension: Child Language Given a string to comprehend, the model searches rlate for Models grammar fragments whose parts can be combined to LAT make the string ➢ Architecture ➢ Learning Env. For each grammar fragment found, its meanings are ➢ World Perception mapped and then combined through union and each ➢ Cross-situational result is returned ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word comprehend a blue cir cle ➢ Abstract Testing a blue cir cle a blue cir cle be low the red square {blue(1,2), circle(1,2)} Results Discussion () ((1,2)­>(2)) Conclusions a blue cir cle {} {blue(1,2), circle(1,2)} T({},()) U T({blue(1,2),circle(1,2)},((1,2)­>(2)))  = {blue(2), circle(2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Learning Module Introduction Aims Child Language  Derive phrasal categories from grammar fragments in Models rlate LAT ➢ Architecture  Derive grammar rules that make reference to phrasal ➢ Learning Env. categories ➢ World Perception ➢ Cross-situational  Create new resource rabstract ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word a blue cir cle be low the red square ➢ Abstract {below(rel), blue(2), circle(2), horizontal_even(rel), red(1), square(1)} ((2)­>(2)) ((1)­>(1), (rel)­>(rel)) Testing a blue cir cle be low the red square Results {blue(2), circle(2)} {below(rel), horizontal_even(rel), red(1), square(1)} ((1)­>(1)) Discussion the red square ((1,2)­>(2)) Conclusions {red(1), square(1)} ((1,2)­>(1)) () ((rel)­>(rel)) blue cir cle () red square {blue(1,2), circle(1,2)} {red(1,2), square(1,2)} ((1,2)­>(1,2)) ((1,2)­>(1,2)) ((1,2)­>(1,2)) ((1,2)­>(1,2)) a blue cir cle be low the red square {} {blue(1,2)} {circle(1,2)} {below(rel), horizontal_even(rel)} {} {red(1,2)} {square(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Learning Module Introduction Aims Child Language  Derive phrasal categories from grammar fragments in Models rlate LAT ➢ Architecture  Derive grammar rules that make reference to phrasal ➢ Learning Env. categories ➢ World Perception ➢ Cross-situational  Create new resource rabstract ➢ Holophrastic ➢ Early Multi-word ➢ Late Multi-word S ➢ Abstract ((2)­>(2)) ((1)­>(1), (rel)­>(rel)) Testing NP1 POS Results ((1)­>(1)) Discussion ((1,2)­>(2)) NP2 Conclusions ((1,2)­>(1)) () ((rel)­>(rel)) NP () NP ((1,2)­>(1,2)) ((1,2)­>(1,2)) ((1,2)­>(1,2)) ((1,2)­>(1,2)) DET1 ADJ N REL DET2 ADJ N Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Learning Module Introduction Creating phrasal categories: Child Language Phrasal categories can be derived from the grammmar Models fragments in rlate by assuming that their members share LAT distributional information ➢ Architecture ➢ Learning Env. 1) blue cir cle ➢ World Perception {blue(1,2), circle(1,2)} ➢ Cross-situational ((1,2)­>(1,2)) ((1,2)­>(1,2)) ➢ Holophrastic ➢ Early Multi-word blue cir cle ➢ Late Multi-word {blue(1,2)} {circle(1,2)} ➢ Abstract Testing 2) blue square {blue(1,2), square(1,2)} Results Discussion ((1,2)­>(1,2)) ((1,2)­>(1,2)) Conclusions blue square {blue(1,2)} {square(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Learning Module Introduction Creating phrasal categories: Child Language Phrasal categories can be derived from the grammmar Models fragments in rlate by assuming that their members share LAT distributional information ➢ Architecture ➢ Learning Env. 1) blue cir cle ➢ World Perception {blue(1,2), circle(1,2)} ➢ Cross-situational ((1,2)­>(1,2)) ((1,2)­>(1,2)) ➢ Holophrastic ➢ Early Multi-word blue cir cle ➢ Late Multi-word {blue(1,2)} {circle(1,2)} ➢ Abstract Testing 2) blue square {blue(1,2), square(1,2)} Results Discussion ((1,2)­>(1,2)) ((1,2)­>(1,2)) Phrasal category 1: Conclusions blue square cir cle {blue(1,2)} {square(1,2)} {circle(1,2)} square {square(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Learning Module Introduction Creating phrasal categories: Child Language Phrasal categories can be derived from the grammmar Models fragments in rlate by assuming that their members share LAT distributional information ➢ Architecture ➢ Learning Env. 1) blue cir cle ➢ World Perception {blue(1,2), circle(1,2)} ➢ Cross-situational ((1,2)­>(1,2)) ((1,2)­>(1,2)) ➢ Holophrastic ➢ Early Multi-word blue ➢ Late Multi-word Phrasal category 1 {blue(1,2)} ➢ Abstract Testing 2) blue square {blue(1,2), square(1,2)} Results Discussion ((1,2)­>(1,2)) ((1,2)­>(1,2)) Phrasal category 1: Conclusions blue square cir cle {blue(1,2)} {square(1,2)} {circle(1,2)} square {square(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Learning Module Introduction Creating phrasal categories: Child Language Phrasal categories often share similar members Models LAT Subset categories are replaced by their superset ➢ Architecture categories ➢ Learning Env. ➢ World Perception ➢ Cross-situational ➢ Holophrastic Phrasal category 1: Phrasal category 2: ➢ Early Multi-word cir cle cir cle ➢ Late Multi-word {circle(1,2)} {circle(1,2)} ➢ Abstract Testing square square Results {square(1,2)} {square(1,2)} is replaced by Discussion star Conclusions {star(1,2)} tri ang gle {triangle(1,2)} Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Learning Module Introduction Creating grammar rules: Child Language Grammar rules are created by linking the grammar Models fragments that make reference to phrasal categories LAT ➢ Architecture ➢ Learning Env. ➢ World Perception NP1 ➢ Cross-situational ➢ Holophrastic () ((1,2)­>(2)) ➢ Early Multi-word ➢ Late Multi-word DET1 NP ➢ Abstract Testing Results NP Discussion ((1,2)­>(1,2)) ((1,2)­>(1,2)) Conclusions ADJ N Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Learning Module Introduction Creating grammar rules: Child Language Grammar rules are created by linking the grammar Models fragments that make reference to phrasal categories LAT ➢ Architecture ➢ Learning Env. ➢ World Perception NP1 ➢ Cross-situational ➢ Holophrastic () ((1,2)­>(2)) ➢ Early Multi-word ➢ Late Multi-word DET1 NP ➢ Abstract Testing Results NP Discussion ((1,2)­>(1,2)) ((1,2)­>(1,2)) Conclusions ADJ N Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Learning Module Introduction Creating grammar rules: Child Language Grammar rules are created by linking the grammar Models fragments that make reference to phrasal categories LAT ➢ Architecture ➢ Learning Env. ➢ World Perception NP1 ➢ Cross-situational ➢ Holophrastic () ((1,2)­>(2)) ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract DET1 NP Testing ((1,2)­>(1,2)) ((1,2)­>(1,2)) Results ADJ N Discussion Conclusions rabstract Kris Jack
    • A Computational Model of StagedLanguage Acquisition Abstract Comprehension Module Introduction Comprehension: Child Language Given a string to comprehend, the model searches rabstract Models that can be instantiated to make the string LAT ➢ Architecture The accompanying meanings is returned ➢ Learning Env. ➢ World Perception NP1 ➢ Cross-situational ➢ Holophrastic () ((1,2)­>(2)) ➢ Early Multi-word ➢ Late Multi-word ➢ Abstract DET1 NP Testing ((1,2)­>(1,2)) ((1,2)­>(1,2)) Results ADJ N Discussion Conclusions If DET1 = a, ADJ = red, ye low and N = cir cle, heart, could comprehend a red cir cle, a red heart, a ye low cir cle and a ye low heart Kris Jack
    • A Computational Model of StagedLanguage Acquisition Summary Introduction Child Language Environment (real or Models simulated) LAT ➢ Architecture                            Sensory stimuli ➢ Learning Env. World ➢ World Perception L e a rn in g Perception Module C o m p re h e n s io n M o d u le s M o d u le s ➢ Cross-situational           renv ➢ Holophrastic renv ➢ Early Multi-word lcross­sit R e s o u rc e s ➢ Late Multi-word rcross­sit ➢ Abstract rcross­sit rholophrastic lholophrastic cholophrastic Testing rholophrastic Results rholophrastic renv rearly learly cearly Discussion rearly rcross­sit Conclusions llate rholophrastic rholophrastic rearly rlate clate rlate rlate rlate rabstract rabstract labstract cabstract rabstract Kris Jack
    • A Computational Model of StagedLanguage Acquisition Testing Introduction Training Child Language  The model is trained to learn a miniature language by Models observing event-description pairs LAT  100 sets of 125 event-description pairs were Testing randomly generated ➢Miniature Lang. ➢Templates  After each pair is entered, the model is tested for ➢Requirements comprehension of a set of strings Results  The results are used to determine the models stage Discussion of linguistic development Conclusions Holophrastic Late Multi-word Pre-linguistic Early Multi-word Abstract 0 6 12 18 24 30 36 42 48 Kris Jack Time (months)
    • A Computational Model of StagedLanguage Acquisition Miniature Language Introduction S = NP1 REL NP2 Child Language REL = REL1 | REL2 Models REL1 = a bove | be low | to the REL4 LAT REL2 = REL3 REL4 Testing REL3 = to the low er | to the u pper ➢Miniature Lang. REL4 = left of | right of ➢Templates ➢Requirements NP1 = DET1 NP Results NP2 = DET2 NP Discussion NP = SHAPE COLOUR COLOUR = black | blue | grey | green | pink | black | red Conclusions | white SHAPE = cir cle | cross | dia mond | heart | rec tang gle | star | square | tri ang gle Can create 32,768 unique sentences such as: a blue cir cle a bove the green square a red dia mond to the left of the white star a pink rec tang gle to the low er right of the black square . . Kris Jack .
    • A Computational Model of StagedLanguage Acquisition String Templates for Testing Introduction String Templates Child Language To observe the performance of the model, it is tested for Models comprehension of a set of strings, shown below by LAT template Testing ➢Miniature Lang. String Template Example String Total ➢Templates Shape cir cle 8 ➢Requirements Colour red 8 Results Position a bove 6 Discussion Half Relative Position to the u pper 4 Conclusions Relative Position a bove the 8 Object red cir cle 82 = 64 Indefinite Object a red cir cle 82 = 64 Definite Object the red cir cle 82 = 64 Object Relative Position a red cir cle above the 83 = 512 Relative Position Object a bove the red cir cle 83 = 512 Event a red cir cle a bove the red square 85 = 32,768 Total No: 34,018 Kris Jack
    • A Computational Model of StagedLanguage Acquisition Judging Linguistic Stage Introduction When the model comprehends strings, its behaviour can Child Language be described in terms of stages: Models  Pre-linguistic – no comprehension LAT  Holophrastic – comprehension of any string Testing ➢Miniature Lang.  Early – string is comprehended as a composite of its ➢Templates parts ➢Requirements  Late – string is comprehended as a composite of its Results parts that require use of syntactic markings Discussion Conclusions  Abstract – a set of NPs are comprehended, where the set includes all known ADJs and Ns  End point – all sentences are successfully comprehended Holophrastic Late Multi-word Pre-linguistic Early Multi-word Abstract 0 6 12 18 24 30 36 42 48 Kris Jack Time (months)
    • A Computational Model of StagedLanguage Acquisition Results Stages of Language Acquisition Introduction 100.00% Child Language Models 90.00% LAT 80.00% Testing Results 70.00% ➢Holophrastic % of Requirements Met holophrastic early multi­word late multi­word post­abstract abstract ➢Early Multi-word 60.00% Holo ➢Late Multi-word 50.00% Early ➢Abstract Late Discussion 40.00% Abstract Conclusions End 30.00% 20.00% 10.00% 0.00% 0 20 40 60 80 100 120 No. Events Observed Onsets: Holo (1); Early (11.9); Late (23.88); Abstract (49.83); End (88.04) Kris Jack Lengths: Holo (10.9); Early (11.98); Late (25.95); Abstract (38.21)
    • A Computational Model of StagedLanguage Acquisition The Holophrastic Stage Introduction ● The majority of strings comprehended (94%) are atomic Child Language units in the language Models ● Word segmentation and association with appropriate LAT meanings Testing ● Discovery of atomic units in the miniature language e.g. Results cir cle and to the u pper ➢Holophrastic ➢Early Multi-word ➢Late Multi-word ➢Abstract The Holophrastic Stage Discussion Definite-Object 1% Complete-Event Conclusions Object 3% 2% Relative-Position Shape 13% 31% Half-Relative- Position 21% Colour 29% Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Holophrastic Stage Introduction ● Under-extensions such as cross means {cross(1, 2), Child Language above(rel)} Models ● Over-extensions such as blue cross means {blue(1, 2)} LAT ● Mismatches such as low means {pink(1, 2), below(rel)} Testing Results ➢Holophrastic ➢Early Multi-word ➢Late Multi-word ➢Abstract The Holophrastic Stage Discussion Definite-Object 1% Complete-Event Conclusions Object 3% 2% Relative-Position Shape 13% 31% Half-Relative- Position 21% Colour 29% Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Early Multi-word Stage Introduction ● There is a rise in the comprehension of composite Child Language strings Models ● Strings are comprehended as a composite of their parts LAT e.g. red cir cle is comprehended from the meanings of Testing red and cir cle Results ➢Holophrastic ➢Early Multi-word ➢Late Multi-word ➢Abstract The Early Multi-word Stage Discussion Object-Relative- Position Conclusions Definite-Object 4% 1% Complete-Event 2% Indefinite-Object 2% Shape 24% Object 22% Colour Relative-Position 24% 9% Half-Relative- Position 12% Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Early Multi-word Stage Introduction No sensitivity to syntactic markings Child Language  a red square means the same as the red square and Models red square LAT  a red square a bove the green cir cle means the Testing same as a green cir cle a bove the red square Results ➢Holophrastic ➢Early Multi-word ➢Late Multi-word ➢Abstract The Early Multi-word Stage Discussion Object-Relative- Position Conclusions Definite-Object 4% 1% Complete-Event 2% Indefinite-Object 2% Shape 24% Object 22% Colour Relative-Position 24% 9% Half-Relative- Position 12% Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Late Multi-word Stage Introduction ● There is a rise in the comprehension of composite Child Language strings that require sensitivity to syntactic markings Models ● Strings are comprehended as a composite of their parts LAT e.g. “red cir cle” is comprehended from the meanings of Testing “red” and “cir cle” Results ➢Holophrastic ➢Early Multi-word ➢Late Multi-word ➢Abstract The Late Multi-word Stage Discussion Complete-Event 8% Colour Conclusions Relative-Position- Object Shape 6% 6% 2% Half-Relative- Object-Relative- Position Position 3% 9% Relative-Position Definite-Object 3% 17% Object Indefinite-Object 30% 16% Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Late Multi-word Stage Introduction Sensitivity to syntactic markings Child Language  a red square means a novel red square Models  the red square means the existing red square LAT Testing  red square means a novel or the existing red square Results  a red square a bove the green cir cle is differentiated ➢Holophrastic from a green cir cle a bove the red square ➢Early Multi-word ➢Late Multi-word ➢Abstract The Late Multi-word Stage Discussion Complete-Event 8% Colour Conclusions Relative-Position- Object Shape 6% 6% 2% Half-Relative- Object-Relative- Position Position 3% 9% Relative-Position Definite-Object 3% 17% Object Indefinite-Object 30% 16% Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Abstract Stage Introduction ● The majority of strings comprehended are Complete Child Language Events Models ● Complete Events are comprehended as the LAT composition of multiple atomic units Testing Results ➢Holophrastic ➢Early Multi-word ➢Late Multi-word ➢Abstract The Abstract Stage Discussion Indefinite-Object Definite-Object Conclusions 1% 1% Object-Relative- Object Position 1% 3% Relative-Position- Object 3% Complete-Event 91% Kris Jack
    • A Computational Model of StagedLanguage Acquisition The Abstract Stage Introduction ● A form of rote learning is displaced by generative Child Language comprehension Models ● Grammars are derived that allow any string in the LAT miniature language to be comprehended from a Testing relatively small exposure to examples Results ➢Holophrastic ➢Early Multi-word ➢Late Multi-word ➢Abstract The Abstract Stage Discussion Indefinite-Object Definite-Object Conclusions 1% 1% Object-Relative- Object Position 1% 3% Relative-Position- Object 3% Complete-Event 91% Kris Jack
    • A Computational Model of StagedLanguage Acquisition Discussion Introduction Behavioural stages emerge: Child Language ● In the same order as found in child language Models LAT ● At similar time intervals as found in child language Testing ● With similar developmental characteristics as found Results in child language Discussion What accounts for this similar developmental trajectory Conclusions bearing in mind that: ● Training data are kept constant? ● The model’s functionality is kept constant? Kris Jack
    • A Computational Model of StagedLanguage Acquisition Explaining Development Introduction The Modular Architecture Child Language  Each module concentrates on performing a different Models task LAT  Each task requires a different amount of training to Testing produce results Results  A new behaviour emerges when a learning Discussion mechanism solves a task for the first time Conclusions  Modules depend upon training data which can be internally filtered by other modules Kris Jack
    • A Computational Model of StagedLanguage Acquisition Explaining Development Introduction ● The Cross-situational Learning Module receives Child Language unaltered training data Models ● The Holophrastic Modules breaks the language down LAT into atomic units producing holophrastic behaviour Testing ● The Early Multi-word Modules begins to reconstruct the Results language by discovering compositional relationships Discussion Conclusions ● Both the Holophrastic and Early Multi-word Modules work simultaneously, allowing the model to continue learning words while discovering compositions ● The Late Multi-word Module begins to reconstruct the language by discovering compositionality WITH sensitivity to word order and syntactic markings ● Why is there such a gap between the results produced by the Early and Late Multi-word Comprehension Modules when they perform similar tasks? Kris Jack
    • A Computational Model of StagedLanguage Acquisition Explaining Development Introduction ● Why is there such a gap between the results produced Child Language by the Early and Late Multi-word Comprehension Modules when they perform similar tasks? Models LAT  The Late Multi-word Learning Module is performing a Testing more complex task than the Early Multi-word Learning Results Module Discussion ● The Late Multi-word Learning Module has tougher Conclusions constraints (word-order and transformations must match in constructions). Given the fragments; a blue blue cir cle {blue(2)} {blue(1,2), circle(1,2)} () ((1,2)­>(2)) ((1,2)­>(1,2)) ((1,2)­>(1,2)) a blue blue cir cle {} {blue(1,2)} {blue(1,2)} {circle(1,2)} ● The Early Multi-word Learning Module can keep the fragments but the Late Multi-word Learning Module cannot Kris Jack
    • A Computational Model of StagedLanguage Acquisition Explaining Development Introduction ● The Abstract Modules produce results much earlier Child Language than the abstract stage begins Models ● Much of the generative capacity in the model comes LAT from the Abstract Modules Testing ● The Abstract Comprehension Module accounts for Results comprehension of novel strings even during the early Discussion multi-word stage Conclusions ● It is inappropriate to think of each module’s contribution to comprehension as being limited to a particular stage ● It is better to think of each stage as being the result of all modules producing the best results that they can given their experience Kris Jack
    • A Computational Model of StagedLanguage Acquisition Conclusions Introduction A computational model that demonstrates a similar Child Language developmental trajectory as found in child language has been produced Models LAT  There is linguistic maturation without physical Testing maturation Results  The model is given a realistic exposure to training Discussion data Conclusions A Modular Structure accounts for much of the developmental shape  The stages emerge on a reasonable timescale  The stages emerge in the same order  Different modules focus upon different problems Different linguistic behaviours may be the best indicators of underlying learning mechanisms in children Children may also have a modular framework for learning and comprehending Kris Jack
    • A Computational Model of StagedLanguage Acquisition Coming soon... Introduction ● Language Acquisition Toolkit (LAT) Child Language Models online LAT Testing – Freely available for research Results – GNU Licence Discussion Conclusions – Run language acquisition simulations with your own modules – Compare results within a common framework Kris Jack