Neural Structure of Language - 1.1

Why do we have langugage

• to reason about things and make arguments 

• Structure - has a built in, formal, logical structure

  • mathematics/logic built upon langauge

definition of language

a discrete combinatorial system with an infinitive generative capacity 

Computational representation theory - collapsing computational & algorithmic levels to one level 

• There is a utility to combining 2 levels because it gets fuzzy in linguistics

• Algorithmic level - a lot of freedom 

• Highly constrained representational type - numbers

How Markov Model works

• Add up probabilities involved 0 sum gives probability of entire sentence

Score  If its below certain value/threshold = bad  If its above certain value/threshold = good

Fodor - what is language

• mentalese - logical form + means something

• innate, non spoken internal code called mentalese

•  has a syntactic structure that separates it from meaning - think in a systematic way & build things 

What you know when you know a language

What sequences of sounds can be strung together to make a valid word

Ex: sequence like “blick” is a valid word in English, but “bnick” is not.

• What words and sentences mean 

  • Basically what sequences of words make a valid sentence

Ex: The cat slept on the mat *The cat mat slept on *Cat the mat the on slept *Cat mat slept on the the

Theory of Language

 Claude Shannon - 1st time you could translate language into something that can be implemented by a computer 

• Whatever model you come up w for language should be able to be implemented by a computer 

Chomsky - “Colorless green ideas sleep furiously”

lets treat sentences as strings

• probability (statistic percentage of word coming next) fails to make a logical, grammatical sentence - no meaning

• Chomsky, Pinkert - basically same argument AGAINST probability

• support internal structures - universal grammar

Language as strings

Sentence = string of words - sequences

Markov Model of Language

knowing the probability of one word following another word is linguistic knowledge

• model seems like it could describe language,

  • produces sequences but not language like humans

What is the goal of linguistics

make a model that describes how language works

Local probability

 (prob of a word following after another)

Global probability

 (prob of whole sentence occurring)

Local vs global probability - The house to ask for is to earn our living.

• this sentence has a low probability - its not even grammatical

• every transition between words is moderate- but the probability of the whole sentence is very low

Ex:

• the → house - kinda high prob 

• house → to - ok amount of prob 

Probability does not determine gramaticality

The things are separated

Generative Grammar 

human language is shaped by a set of basic principles that are part of the human brain

• Chomsky’s alternative to language as strings

• Linguistic knowledge - knowing the abstract rules that generate the grammatical sentences in your language 

• Use rules and united to have structure that you gain - to model human language 

Mental grammar made up of

Phonology

Lexicon

Syntax

Phonology

- store of sounds and rules to combine them 

Lexicon

 Mental store of words and meanings 

Syntax

- Rules that combine words into structures

Parts of a language

• All languages rely on notion of structure 

• Make infinite use of finite meanings 

• Can generate unbounded amount/length of sentence

Language as structure - meaning chases snytax up the tree

basically meaning is in every part of the tree from the bottom - defines the syntax of it in a way - how the sentence is perceived and understood

How are LLM (large language models, AI) counter evidence to Chomsy’s generative grammar

It doesn’t have hierarchical structure but its still grammatically correct

• Natural Language Processing (NLP) has largely abandoned generative linguistics.

• Modern advances in language understanding systems look more like probabilistic models than Chomsky’s generative grammar.

Levels of analysis

• The mind is a complex information processing system 

• The mind performs computations 

• Language is a complex system computed in the mind

Computational level

Algorithmic level

Implementational level

Computations

• operations performed on representations

Computational level

• goals of system

• what will be accomplished after system has run its course

• output of computation

Algorithmic level

• steps taken to accomplish goal/solve problem

• 2nd step

  • Types of Representations created

  • Operations performed

Ex: brain recognizing speech sounds by keeping a table with sound waves matched to appropriate sounds - can match/look up when it hears sound

Implementational level

• How neurons/physical system carry out algorithm that solves a problem

• the 3rd and final step

  • What brain regions are active

  • What are the physical processes/system that executes things

example of computational level

 Abstract description of computation

• Includes formal properties of language 

• Rules and constraints 

• Ex: Past Tense = Verb + “-ed”

Example of algorithmic level

 - Sequence of the steps to perform the computation

• Form of representations 

• How language is processed

• Ex: 

  • 1. kick

  • 2. kick + ed

  • 3. Kicked

Vs 

• 1. kicked

Example of implementation level

 - Physical system that executes steps 

• Where language is processed in the brain 

• Which brain region does what? 

• Ex: parts of brain where language are at

Example of “operations” in computational models

Wh- questions

• In wh- questions, the wh- word appears in a fronted position, but is interpreted in a different position 

• This is described as a movement process 

• Does this mean that the mind executes a “move” operation? 

How Wh- questions show that there is less diversity in language than it seems

• Wh- questions 

  • Goes in front of sentences 

  • Can go at end of a sentence - to replace what you are talking about 

    • You ate what (for lunch)? 

    • What did you eat (for lunch)?

CR theories

(Computational representation theories) - theories about the computations and representations used in language.

• models focus on the abstract rules, symbols, mental grammar, and algorithmic steps the mind uses to process language (linguistic competence)

  
  

NB

(Neurobiological models) - the study of the structure and function of the brain.

• : models focus on the actual physical structures, neurons, pathways, and anatomy of the brain (e.g., the auditory cortex or Broca's area) used during language tasks

Epistemological priority

 needing to know A before you can know B.

Granularity

- how detailed a model is.

• Pinpointing differences in a model/language are different because of mismatch modularity 

• Mismatch modularity - they think each model works different 

GMP

 (Granularity Mismatch Problem)

CR language models are more fine-grained than NB language models.

Competence

 - Linguistic knowledge; mental grammar.

• We all have perfect knowledge of the language that we know 

• Making speech errors doesn’t mean that you know any less - types of speech errors ppl make reveal allot about their language competence on a basic level 

Performance

The execution of linguistic knowledge via cognitive/biological systems.

Neural Circuit

 A small population of connected neurons that execute a single function

What is the mapping problem

•  difference between linguistics and neuroscience in mapping brain processes

• mapping does not directly translate between the 2 systems - because they have different operations and object its hard to directly shift it

Ontology

• core essence of something - cant be changed 

• What is true of the nature of what it is 

• things that ontologically true can’t be separated from object

Ex: Morpheme

• Ontological status - theoretical concepts - diff from the ontology of something being a neuron 

Linking theories

Computational - define model you want 

• You need a theory 

• Ex: Language - speech sounds 

Processing model - how do you process these things 

• What are you representing 

• What info is contained in representations 

• How do you do it? 

Nuerobiological model - biological stuff happening in brain - where is the task 

• Might want to test brain directly 

Behavioral model 

• Once the process happens - how do you make a decision - make decision to push button 

What’s the purpose of a linking theory?

• Link a computational level of description → algorithm → behavioral point 

• So that you can make testable prediction about thing you can actually measure 

• So that you can evaluate it 

What was the Karen Stromswold study

2 sentences given

1 - The juice that the child spilled__ stained the rug

2 - The child spilled juice that ___stained the rug

• can actually be broken down into 2 sentence

  • The child spilled juice. The juice stained the rug.

Sentences were presented to people to see which ones they processed better

Results of Karen Stromswold study

Results: 

On average ppl find sentence 1 harder to process than sentence 2 

• center embedded sentences - slower reaction time/ less understandable -

• right embedded sentences - quicker reaction time/more understandable -

What changed between the 2 conditions

• distance

• internal structure

PET study 

• PET - positron emission tomography - mostly been replaced by FMRI 

• 1st cortical study on Language 

• Used functional micro imaging to show how ppl process lang 

• Brain activation at different levels of complexity

Center embedded sentences show greater activation in pars opercularis than right branching sentences - What does this tell us?

• pars opercularis - rear section of the brain's inferior frontal gyrus - understands hierarchy in brain

• More time 

• More discrete operations

What does Broca’s area do 

• speech production - what studies have amounted to showing

• We’re not entirely sure - depends on algorithmic level theory and linking theory 

• Understand what process is it so you can understand functional operation