Language

Learning
G.O Deák and A. Holt (2008)

University of California at San Diego

Paper Review by Edward Sharick

Introduction

• What is language? What makes is a unique learning Problem

• Why is language hard to learn?
• Nature versus nurture:
• The context of language-learning ‘in the wild’
• How to think about genetic factors in language learning

• Are there critical periods for learning?

Language as a Learning problem
• Focuses specifically on first-language learning in infants and children
• A challenging problem:
• Adult language processing is generally efficient; makes it challenging to
look at the cognitive or neural processes that occur
• Adults don’t have a good memory of the learning process

• Many different processes in learning a language:

• Attention-modulation, classification, retrieval, inference, and cognitive
control.
Why is language hard to learn?
• 3000 to 8000 different languages (depending on definitions) all differ in
many regards, yet have some similarities
• Some universal features of languages:
• Hierarchical structures – phonemes, morphemes, words, and phrases
• Arbitrariness of form
• Modifiability of forms – assimilation, simplification, or metaphor
• Combinatorial complexity
• Idea of self versus other
• Cultural context of language
Why is language hard to learn?
• Differences: • Limited verb morphology (English)
• Tone variations(Mandarin) • Extensive verb inflections and
• Percussive “click” or air-ingestive vowel harmony (Hungarian)
noises (Zulu) • And many more...
• Limited phonology (Hawaiian)

• Thousands of differences like these exists between different languages, but

even within the same language there are many differences:
• Different dialects and accents

• Different meanings for words based on culture (soda versus pop versus cola)
The context of language-learning ‘in the wild’
• The motivation for an infant or child to learn language is to affiliate with
caregivers, predict what others will do, and join in positive social
exchanges.
• They are not consciously reflection on language as a learning problem,
they are not trying to learn the language. It is a means to an end.
• Sometimes learning is tailored to the infant’s ongoing experience (baby
talk), sometimes it is not (two adults talking near baby).
• All cultures do not communicate with their infants in similar ways
• It is not obvious how different cultural practices influence language learning
How to think about genetic factors in language learning

• Some universal language features are not acquired by any other

species besides humans
• What are the genetic factors that allows humans to learn language in
this way that other species cannot?
• Not much is known at this point, but researchers are asking questions
about the FOXP2 and other genes and their role in neurochemical
processes that affect learning language. Current answers are
speculation.
Critical periods for learning?
• Some researchers argue that language fluency can be attained only during a limited
window of age, after which brain plasticity becomes reduced and fluency is difficult or
impossible to achieve
• Study by Johnson and Newport (1989)
• Tested adult Korean and Chinese immigrants on English syntactic distinctions
• Found a linear decline in competence with increasing age of acquisition from ages 8 to 39
years, but no difference in starting ages ranging from 3 to 7 years.
• Suggests a decline from about 7 years to adulthood in the capacity to master syntactic details
of a new language

• Despite evidence in different studies for critical period effects in articulation and syntax,
the exact nature and cause remain controversial. Some researchers (including this
paper) argue there is little or no compelling evidence for these critical periods.
• Relevance to NARS?
First-language learning
• Broken into the following timeframes:
• First year
• Second year
• Third and fourth years
• Later in childhood

• Despite this organization, language learning is ongoing and continuous rather

than stage-like.
• Further, children do not learn separate aspects of language (syntax vs.
morphology vs. pragmatics), but they are interrelated in human language
processing data.
The First Year
• During the first year, the following questions are asked regarding
infants and language learning:
• What sounds do they recognize?
• What sounds do they put together to recognize words?
• What words can they put together to learn phrases and language structure?
Classification problem: speech sounds in the infant’s
soundscape
• Adult speakers cannot even perceive all phonemes of a language
• In English, bilabial consonants from two phonemes [b] and [p]
• In Thai, the same spectrum is three phonemes

• How do infants learn whatever complement of speech-sound distinctions is

relevant in their language?
• Neonates perceive phonetic distinctions
• Within the first few days, infants discriminate native from foreign speech
• During the first year, children learn the basic linguistic units
• Spoken languages – phonemes (consonants and vowels), articulation
• Signed languages – hand shapes, manual motions, and body motions

• Some evidence suggests a loss of sensitivity to distinguish these basic

linguistic blocks after 9-12 months
Beyond phonology: finding the words
• Infants need to:
• Learn combinations of speech sounds (words and phrases)
• Parse a speech stream (what separates words)using transitional
probabilities:
• Ex: Pretty baby (2 words) versus Pritt ebay bee (3 words)
• In English, more likely to have the sequence ‘tee’ or ‘bay’ than ‘eeb’
• Not limited to language learning (used in musical motifs or sequence of visual
shapes)

• NARS application: suggests a probabilistic approach to parsing input

streams
First Words: Content and conditions of
learning
• Besides deciphering the speech stream, children must associate certain
sequences of phonemes (words) with contexts of use.
• By 4 months, attend more to their own name than another name with same
stress pattern
• By 7 months, preferences extend to high-frequency words (ex: cup)
• By 11 months, represent the phonological details of familiar words.

• It is still unclear when and how infants learn words as abstract symbols
(associating object types, people, events or properties with the word),
but it is input-driven.
Beyond words: learning phrase structure and
lexical-syntactic categories
• In the first year, infants are sensitive to melodic contours of infant-
directed utterances (approval vs. prohibition, for example)
• These cues can be consistent across languages

• 6-month-olds found to discriminate between content words (‘cup’,

‘chair’) from grammatical words (‘the’, ‘you’) and prefer the former.
• 12-month-olds found to be sensitive to ‘grammatical’ versus
‘agrammatical’ sentences and similar transitional probabilities between
words (as was shown with phonemes)
Beyond words: learning phrase structure and
lexical-syntactic categories
• Current understanding is that infants discriminate:
• Familiar from less familiar orderings of syllables
• Acoustic and prosodic (melodic) cues that correlate with phrase and clause
boundaries
• Phonological cues that differentiate broad syntactic categories (content vs.
grammatical words).
First uses: reasons to learn language
• It should be noted that although language researchers describe infants
as trying to solve this complicated mapping problem, the infant’s goal
is to stay regulated, reduce uncertainty and maximize hedonic states.
They must pick and choose what information to pay attention to in rich
environments. Faces, voices and hands, tend to attract their attention.
• Do we have the same problem in NARS’ language processing? What
should we pay attention to and what should we ignore?
The Second Year
• Research shows that language advance in several critical ways:
• Lexical knowledge – words and phrases
• Pragmatics – context of the words and phrases
• Syntax - rules for putting them together
New math: populating the lexicon
• First words are typically generic object labels, proper names, action
words and social routine words (contextual)
• By 13-18 months, extend a novel word to a new object and can
generalize generic words from first referents to abstract classes.
• They can however overextend a words use (‘ball’ for any sphere) or
underextend it (excluding penguins from ‘bird’).
• They then use input that will modify the boundaries of the use of these
words
New math: populating the lexicon
• How does the child’s mind work in learning new words? What mechanism do
they use to accelerate in word learning?
• One hypothesized explanation: as neural representation patterns evoked by
particular words become more stable and better defined, this stability can
make it easier to learn new word-referent associations.
• They can use certain words types to guide inferences about new word
meanings (ex. Shapes and materials).
• NARS application: what input is needed and how (mechanism) do we adjust the
boundaries of words that are already known? How do we use prior knowledge
to affect acquisition of new knowledge?
Pragmatics: Inferring the meaning behind the words

• Toddlers can use nonlinguistic social cues to reduce uncertainty of a

speaker’s referential meaning
• “gazzer” study

• Implication for NARS? Infer meaning from any possible context (even
other streams)
• Ex: From presentation last week, inferring the meaning of “he” in a
question by referring to a previous question or statement.
Combinatorial explosion: putting words
together
• Early combinations express relations:
• Action – ‘Kathryn jumps’
• Locative action – ‘Tape on there’
• Locative state – ‘I sitting’
• Static state – ‘Caroline sick’
• Recurrence – ‘more milk’
• Possession - ‘Mommy sock’
• And others...
Combinatorial explosion: putting words
together
• Transformations vs. rules-with exceptions
• Ex: English past tense – ‘make’/’made’, ‘come’/’came’, ‘go’/’went’, ‘cut’/’cut’

• Toddlers will sometimes overregularize, producing form like ‘goed’,

‘runned’ or ‘breaked’ instead of ‘went’, ‘ran’ or ‘broke’.
• Marcus et al. argued that exceptions must become strong enough as
memory traces to be retrieved before the rule is applied.
• 2-year-olds are starting to learn the syntactic properties of different
words and phrases, but their specific knowledge is variable, ephemeral,
and unorganized by abstract distinctions.
The Third and Fourth Years

• Acquiring semantic relations

• New uses of language
Acquiring Semantic Relations
• As the vocabulary grows beyond a certain size, semantic relations
become important:
• Inclusion, overlap and exclusion
• Ex: are all pets animals?

• As early as children know enough words to begin filling in semantic

frameworks, they can constrain inferences and naming decisions
• Ex: “not the red one, the chromium one”. They assume ‘chromium’ is a
color
Acquiring Semantic Relations
• Toddler’s ways of learning appropriate semantic relations:
• Speakers couch words in meaningful information (such as statements of
contrast)
• Syntactic context
• Analogize from familiar semantic relations – “A pug is a dog”
New uses of language
• Preschool children’s language skills develop in service of social
knowledge and interaction
• Different purposes (narrative, conversation or teasing)
• Different contexts (home, school, mealtime)
• Different interlocutors (peers, siblings, parents)

• Navigating these different contexts requires very flexible linguistic

skills, and being immersed in these different situations bring great
advances in ability to use language
• Application to NARS? Experience is the best teacher
Later Childhood
• By age 4-5, children show basic fluency and mastery of basic
morphological and syntactic structures.
• What remains? The ability to apply basic linguistic knowledge in:
• contexts that are more challenging or complex
• require integration of linguistic and paralinguistic information
Learning the nuances
• Continue to learn relational semantics, such as inclusion statements
• “A pug is a dog” can infer that “A pug is an animal”
• “A pug is an animal” does not infer that “A pug is a dog”

• Complex online syntactic judgements

• “Mrs. Brown working at the library called home to say she would be late”
• “Put the frog on the napkin in the bowl”
• Adults can withhold judgement until the end of the sentence
• Children try to resolve syntactic ambiguities before parsing is complete
Learning the nuances: Reference, pragmatics, and implicature

• Good discourse cohesion

• Topic-introducing-and-shifting (“There was this guy. He…”)
• Topic-continuing (“yeah, and…”)
• Perspective-shifting (“No, he didn’t do it, she did.”)

• Rely heavily on the use of pronouns and generic descriptions

• Ex: “Chris and Heidi went to a new restaurant. The waitress asked Chris
and Heidi if Chris and Heidi wanted drinks. ‘‘No,’’ said Heidi. Heidi had
already had some wine.”
From fluency to flexibility and meta-language
• To progress in linguistic achievements involves greater precision of
interpretation or production (representing different perspectives),
which require representational flexibility.
• Shifting attention
• Generating/selecting new representations
• Suppressing prior cues and associations
Cognitive flexibility
• Cognitive flexibility encompasses the ability to formulate and select
appropriate but nonobvious representations of a referent or sentence in
light of contextual information
• Ex: /har/ as a synonym for rabbit, as opposed to hair

• How to develop cognitive flexibility has been linked to working memory

and inhibitory processes or coordinating multiple response-
contingencies in response selection
• But research in this area is still unclear
Becoming an expert language user
• As language skills become consolidated, children become faster and
more accurate and processing and producing more complex and novel
utterances.
• However, there is no neurobiological accounts of later language
development (surprising since training interventions can measurably
change children’s neural activity during language processing).
• Also, expertise acquisition is disconnected from later language
development (phonological similarity neighborhoods are fewer in
children versus adults).
• The authors speculate that language expertise depends on input-
expectant learning (similar to face processing).
Knowledge about language

• Order children develop metalinguistics, the ability to reflect on

language
• Focuses on dissociation of representations, as between a word as an action
and as a symbol of whatever it represents.
• Ex: Which is longer? ‘Mosquito’ or ‘Cow’

• Metalinguistic knowledge develops in conjunction with other meta-

representational skills.
• Development seems to improve based on experience; more experience
means a better metalinguistic knowledge.
Conclusions and Implications
• As advances in neuroscience change our understanding of human
cognition, they challenge the persistent myths and assumptions about
language
• Ex: Chomsky’s language acquisition device
• Growing sophistication of computational simulations of language learning
support neurally plausible accounts of language development

• Despite extensive use of terms like ‘syntax’, ‘semantic’, ‘morphology’, and

‘discourse’, these are conveniences based on historical convention. More
typical are complex interrelations among aspects of linguistic knowledge.
• Research on different populations, including infants and children with
various developmental disorders, and adults with neurological and sensory
deficits are necessary to understand typical language development.