Language production


In psycholinguistics, language production is the production of spoken or written language. It describes all of the stages between having a concept, and translating that concept into linguistic form. In computational linguistics/natural language processing and artificial intelligence, the term natural language generation is more common, and those models may or may not be psychologically motivated.

Stages of production

Language production consists of several interdependent processes which transform a nonlinguistic message into a spoken, signed, or written linguistic signal. Though the following steps proceed in this approximate order, there is plenty of interaction and communication between them. The process of message planning is an active area of psycholinguistic research, but researchers have found that it is an ongoing process throughout language production. Research suggests that messages are planned in roughly the same order that they are in an utterance. After identifying a message, or part of a message, to be linguistically encoded, a speaker must select the individual words—also known as lexical items—to represent that message. This process is called lexical selection. The words are selected based on their meaning, which in linguistics is called semantic information. Lexical selection activates the word's lemma, which contains both semantic and grammatical information about the word.
This grammatical information is then used in the next step of language production, grammatical encoding. Critical grammatical information includes characteristics such as the word's syntactic category, what objects it takes, and grammatical gender if it is present in the language. Using some of these characteristics as well as information about the thematic roles of each word in the intended message, each word is then assigned the grammatical and thematic role it will have in the sentence. Function morphemes, like the plural /s/ or the past tense /ɪd/, are added in this stage as well. After an utterance, or part of one, has been formed, it then goes through phonological encoding. In this stage of language production, the mental representation of the words to be spoken is transformed into a sequence of speech sounds to be pronounced. The speech sounds are assembled in the order they are to be produced.
The basic loop occurring in the creation of language consists of the following stages:
According to the lexical access model, in terms of lexical access, two different stages of cognition are employed; thus, this concept is known as the two-stage theory of lexical access. The first stage, lexical selection provides information about lexical items required to construct the functional level representation. These items are retrieved according to their specific semantic and syntactic properties, but phonological forms are not yet made available at this stage. The second stage, retrieval of wordforms, provides information required for building the positional level representation.

Models of production

Serial model

A serial model of language production divides the process into several stages. For example, there may be one stage for determining pronunciation and a stage for determining lexical content. The serial model does not allow overlap of these stages, so they may only be completed one at a time.

Connectionist model

Several researchers have proposed a connectionist model, one notable example being Dell. According to his connectionist model, there are four layers of processing and understanding: semantic, syntactic, morphological, and phonological. These work in parallel and in series, with activation at each level. Interference and misactivation can occur at any of these stages. Production begins with concepts, and continues down from there. One might start with the concept of a cat: a four-legged, furry, domesticated mammal with whiskers, etc. This conceptual set would attempt to find the corresponding word. This selected word would then select morphological and phonological data /k / at/. The distinction of this model is that, during this process, other elements would also be primed, as they are physically similar, and so can cause conceptual interference. Errors might also occur at the phoneme level, as many words are phonetically similar, e.g. mat. Substitutions of similar consonant sounds are more likely to occur, e.g. between plosive stop consonants such as d, p and b. Lower primed words are less likely to be chosen, but interference is thought to occur in cases of early selection, where the level of activation of the target and interference words is at the same level.

Lexical access model

This model states that the sentence is made by a sequence of processes generating differing levels of representations. For instance, the functional level representation is made on the a preverbal representation, which is essentially what the speaker seeks to express. This level is responsible for encoding the meanings of lexical items and the way that grammar forms relationships between them. Next, the positional level representation is built, which functions to encode the phonological forms of words and the order they are found in sentence structures. Lexical access, according to this model, is a process that encompasses two serially ordered and independent stages.

Additional aspects of production

Fluency

Fluency can be defined in part by prosody, which is shown graphically by a smooth intonation contour, and by a number of other elements: control of speech rate, relative timing of stressed and unstressed syllables, changes in amplitude, and changes in fundamental frequency.

Multilingualism

Whether or not a speaker is fluent in only one language, the process for producing language remains the same. However, bilinguals speaking two languages within a conversation access both languages at the same time. The three most commonly discussed models for multilingual language access are the Bilingual Interactive Activation model, the Revised Hierarchical Model, and the Language Mode model:
Speakers fluent in multiple languages can inhibit access to one of their languages, but this suppression can only be done once the speaker is at a certain level of proficiency in that language. A speaker can decide to inhibit a language based on non-linguistic cues in their conversation, such as a speaker of both English and French conversing with people who only speak English. When especially proficient multilingual speakers communicate, they can participate in code-switching. Code-switching has been shown to indicate bilingual proficiency in a speaker, though it had previously been seen as a sign of poor language ability.

Research into production

There are two main types of research into speech production. One type focuses on using the analysis of speech errors. The other looks at reaction-time data from picture-naming latencies. Although originally disparate, these two methodologies are generally looking at the same underlying processes of speech production.

Speech errors

Analysis of speech errors has found that not all are random, but rather systematic and fall into several categories. Although speech production is very fast, the error rate of the utterances are relatively rare and those errors are categorized as follows: