Language Production

Language production is the process of producing discourse, from conceptualization to articulation, in the form of speech production or writing.


1   Research

Presently, most research focuses on how speakers accomplish seemingly simple goals; producing utterances in a particular language.

At the outset, this calls for being able to think in ways that readily converted into language or "thinking for speaking".

It means bringing to mind words or morphemes that are suitable for conveying the thought, arranging the words or morphemes in the way that is grammatical and semantically meaningful, and producing the sound with one's throat.

For example, consider describing where one lives. One could use a coordinate system to precisely say so. However, this is unlikely to successfully communicate the idea. A better tactic is to take account of the context of the conversation and what the listener knows to consider "common ground". (An important aspect is the ability of the speaker to retrieve information about the listener from memory.)

Having prototyped some thought with the intention of communicating it, the speaker has the beginnings of a "message". To the transform the message into language and then into speech means retrieving from memory words from the lexicon, arranging them into a sentence, and then forming the syllables at a fast rate.

[1]"Spoken Language Production: Psycholinguistic Approach (Bock, Konopka, Middleton)"
[2]"Speaking and Misspeaking (Dell)"


## Conclusion


## The Freudian Approach

Many people think slips result from a conflict of what one plans to say and some unconscious intention. That slips tend to create words over nonwords (Baars 1975) seems to support the claim.

However it is a big jump to claim that revealed unintended thoughts were repressed and that the slip actually has the function of giving through expression.

## A Cognitive Science Approach

## Word Errors and the Building of Sentences

One of the most striking facts about slips is that they obey a syntactic category rule. When one word erroneously replaces another, most of the time the target and substituting words are of the same syntactic category.

One way to account for these facts is to assume that these error occurring during the construction of syntactic representation of the utterance. Superficially, it has been suggested the processing goes like this: Based on the intended meaning of the utterance (the semantic representation), words are retrieved from the mental lexicon, the store of words we already know. Furthermore, a syntactic frame is constructed that indicates the potential structure of the sentence, which can be thought of as tree that indicates the grammatical relations among the words in the sentence. The frame by itself however does not initially contain any words; it has empty slots for the words to go into, slots that are assumed to be labeled for a syntactic category. The labels on the slots are assumed to guide insertion.

Word errors that obey the syntactic category rule provide good evidence for the idea that sentences are built by placing words units in labeled slots in syntactic frames. The suggests the error ultimate arise because of the need for the syntactic level to be creative. This is possible because the system separates syntactic patterns (the frame) from the words, with the result that the system implies the existence of sentences other than those may have already been produced.

## Semantic and Phonological Relations in Word Errors

## Phonological Errors

## Word and Sound Error and Linguistic Creativity

## Conclusions - Slips and Cognitive Science

We began by promising to take a cognitive science approach rather than Freudian one. We sought explanation for the properties of slips by looking at the nature of language and how it is produced, rather than by looking at repressed memory.

Computational models. By translating one theory's into a computer program, one can see whether the theory in facts behaves the way one expects it to. This translation has been particularly important for account for speech error data because there are so many factors at working producing errors.

## Further Reading


How people express their words into thoughts


2   Word-finding

Because messages are not represented linguistically, appropriate words must be retrieved from the mental dictionary in a speaker's memory.

Lexical entries may be located on the basis of:

  1. Meaning
  2. Syntactic category
  3. Morphological or phonological forms.

Entries are accessible (1) from messages, (2) from the structural procedures that play out words into connected speech, and (3) from sound.

These multiple avenues make the lexical entryway (1) a busy intersection in the process of production, as well as in comprehension and (2) a target of research aimed at uncovering how the intersection works.

## Forces in Word Finding - Context, codability, and word frequency

The semantic and phonological properties of words are linked respectively to two powerful forces in word finding:

  1. Conceptual accessibility 2. Word-form frequency

The meaning of words vary in how specific they are. Because retrieve ability depends in part on (1) the goodness of the match between the retrieval context and (2) the information associated with a word's meaning in memory, contexts with many relevant cues can faciliate naming in comparison to contexts with fewer cues.

Complementary, words with sparser semantic representations (e.g. proper names and abstract words) are harder to retrieve than words with richer representations (e.g. concrete words) other things being equal.

The familiarity of words (whether due to frequency with which they are produced or the age at which they were learned) has been known to affect how quickly or easily they are uttered. Higher-frequency words are produced faster than lower-frequency words. The phenomenon of "Frequency Inheritance" suggest that frequency affects the retrieval of word forms more than access to lexical entries proper. Frequency Inheritance refers to a finding the producing of low-frequency words benefit from the presence in the lexicon of high-frequency homophones that are unrelated in meaning..


# Connectionist Models of Aphasia and Other Language Impairments

## Introduction

Theorists have used models in order to make sense of the variety of symptoms of aphasia.

Example: In 1885, Lichtheim diagrammed the relation between aphasic syndromes and brain regions.

Example: Flow-chart Example: Box-and-arrows

A computational model is a model that is expressed as a computer program.

A computational model is helpful for determining the consequences of damage to components in complex or probabilistic models.

A connectionist model is a computational model. A connectionist model is a network of units that connect to one another through links (connections) that can vary in strength.

Some models have a learning component that determines the connection weights.

A model that has a learning component must be trained by giving it many trials.

An input activation pattern and (typically) a desired output activation pattern compose a trial.

Each unit ("node") has an activation value (usually a real number between 0 and 1).

The activation value of a unit changes over time as activation passes through connections from unit to unit.

Each connection has a strength ("weight").

An excitatory connection is a connection that has positive strength. An inhibitory connection is a connection that has negative strength.

Processing is a mapping from an input to an output. Processing is carried out by spreading activation.

  1. A model is given an input by setting the activation levels of some of its input units to particular values. 2. The activation spreads in parallel throughout the network via the connections. 3. The output of the model is determined by examining the activation levels of a set of output units.

An activation rule is an equation that specifies how the activation value of each unit changes when it receives activation from its neighboring units.

Activation rules govern the spread.

Example: The input to a model of word retrieval during production might consist in setting the activations of units representing semantic features of the sought after word to positive values. The output units might represent the phonemes or phonetic features of the retrieved word.

Note: Although connectionist researchers use computers at tools, they often reject the computational metaphor for cognition - the idea that the brain is much like a standard computer and that cognition is the product of programming.

Connectionist models are inspired by neural systems.

Example: Connectiont model mimic assumption that the brain process information in parallel, whereas Von Neumna architectures carry out instructions sequentially.

Connectionist modelers rarely link model parts with brain regions. Rather, connectionist modelers aim to correctly characterize the cognitive mechanisms of language processing, with the hopes these mechanisms can be identified with brain areas.

Connectionist models have been applied to language disorders for more than 20 years.

## The Interactive Two-Step Model of Lexical Retrieval in Aphasic Speakers

The interactive two-step model is a model of single-word production that is derived from a general theory of production in which linguistic units are retrieved by spreading activation in a layered network.

The interactive two-step model is an interactive lexical retrieval process.

The units in the network create localist representations; the units in the network correspond directly to particular linguistic units.

Note: The alternative is distributed representation, in which linguistic units correspond to a pattern of activation across many network units rather than to a single unit.

Each connection in the interactive two-step model is excitatory. Each connection in the interactive two-step model runs between adjacent layers.

A top-down connection is a connection that goes toward an output node.

An interactive retrieval process is a retrieval process that has both top-down and bottom-up connections.

Word retrieval and phonological retrieval compose lexical retrieval.

### Word Retrieval

The word retrieval step begins with a jolt of activation to the semantic features of the targets word. This activation spread throughout the network, down to word and phoneme units and upward as well. The most active word is selected after a period of time, which completes the word retrieval step.

The activation of all units has random variation.

Words related to a target in phonological form gain activation during the word retrieval steps.

A lexical error is a ...

A formal error is an error that occurs when a form-related word is selected.

A mixed error is an error that occurs when a word that is semantically and formally related to the target is selected.

TODO: Selected by whom?

### Phonological Retrieval

The phonological retrieval step begins with a jolt of activation to the lexical units of the selected word.

Then, activation spreads throughout the network, down to phoneme units and upward to semantics.

Finally, the most active phoneme unit is selected.

A phonological error is an error that occurs when a selected phoneme does not correspond to a selected word unit.

A phonological error can produce a non-word or a word (which would create a formal error).

Note: A formal error can occur at either the word retrieval step or the phonological retrieval step. Note: A phonological error can be made on top of a lexical error. (Such errors are uncommon in normal speakers, but are not uncommon in aphasic speech.)

The continuity thesis is the assumption of the account of aphasia of the model that states aphasic errors are generated from the same mechanisms that create speech errors in unimpaired errors.

### Simulating Naming in Control Subjects

### Specifying the Nature of Lesions in the Model

To lesion a model is to alter it in some way so that the model performs less accurately.

Example: Aphasia models have two lesionable parameters: the strength of bidirectional connections between semantic and lexical units and the strength of those between phonological and lexical units.

### Fitting the Model to Naming Data

Example: Schwartz 2006.

### Testing Predictions from the Model

A model is valid if it can generate predictions that are verified.

Example: Model's account of picture naming can be used to predict performance on word repetition.

The interactive two-step mode assumes that all the connections of a particular type have the same weight. Clearly this is wrong. The strength of connection should reflect differences in an individual's experience with words.

Example: Lexical experience has strong effect on aphasic errors as these errors are more likely on low-frequency and late-acquired words.

## A Parallel Distributed Processing Model of Naming Errors in Optic Aphasia

The parallel distributed processing (PDP) model is designed to mimic the impairment of optic aphasia.

Optic aphasia causes patients to make errors when naming visually presented objects. However, it does not impair naming objects presented in other modalities nor tests of visual recognition.

To develop the optic aphasia model, Plaut and Shallice modified the deep dyslexia model, lesioned it to simulate brain damage, and compared its response to those of optic aphasics.

### Initial Model Architecture

The model assumes optic aphasia (like deep dyslexia) stems from an impairment in accessing semantic representation from visual representations.

PS constructed a model that accesses semantic information in two stages:

  1. A rough approximation of the object's semantics is generated from its visual reprsentation (direct pathway) 2. This approximation is iteratively refined to the precise semantics of the object (indirect pathway)

The direct pathway of the model consists of:

  1. A visual input layer 2. An intermediate layer 3. A semantic output layer

Top-down excitatory connections link the visual input layer to the intermediate layer Top-down excitatory connections link the intermediate layer to the semantic layer.

Representations in the optic aphasia model are distributed.

Example: The visual representation of "spoon" consists of a pattern of activity over many units in the visual layer rather than the activation of a single node.

The units of the input layer represent the visual information available from object recognition.

Subgroups of these units code different visual features.

Similarly, the semantic output layer has subgroups of units that code different semantic features.

The intermediate units mediate between the input and output layer.

Distributed representation reflect similarity. Related objects will have similar patterns of activation over the units of the model.

### Training: Setting the Model's Long-Term Connection Weights

The architecture of the model specifies the units and the potential connection between them, but the connection weights must be learned.

### Lesioning and Testing the Model

### Comparing Model Behavior with Patient Performance

## Aphasic Lexical Access in Sentence Production

## Summary and Conclusion

We are far from a unified model of aphasia.

Models are as diverse as data.

Some models appeal to connectionist learning principles to explain the data. Some models attribute error effect to interactive spreading activation. Some do both.

Despite differences, connection approaches share a cooperative view of language processing.

Cooperative view of language Processing:

Although the models have separate network levels that correspond to distinct representational types, the levels work together to explain empirical phenomena.

We might describe an error as "semantic" or "phonological" but that does not mean that responsibility for the error lies within a single level.

We see multi-level effects in model with a bidirectional or interactive flow of activation.

Connectionist models allow for multiple levels to affect processing through learning.

The cooperative view of language has much to offer research on communication disorders.


Researchers try to fit models to predict aphasic behavior on some task, then test it on another related task.

3   History

Until relatively recently (this is from 2006), neither linguists or psycholinguists paid much attention to the normal process of adult language production.

Studying how people talk is hard and easy to dismiss as a viable scientific problem. Both of these issues have receded. By 2006, there were valid, reliable methods for getting at the underpinnings of production, burgeoning scientific interest in the language production ability.