Chapter 13 Language Production

Transcription

1 Introduction Chapter 13 Language Production Speech Errors & What They Tell Us Syntactic Planning Lexicalization Phonological Encoding The Analysis of Hesitations The Nueropsychology of Speech Production Writing & Agraphia

2 Introduction The processes of speech production fall into 3 broad areas: Conceptualization: determining what to say. Formulation: translating conceptual representation into a linguistic form. Execution: detailed phonetic & articulatory planning.

3 The Psychology of Language, Trevor Harley Figure 13.1 Speech production processes (Levelt, 1989).

4 Speech Errors Spoonerisms: speech error where the initial sounds of 2 words get swapped. Freudian slip: slips of the tongue, Freud proposed they revealed repressed thoughts. We can categorize speech errors by considering the linguistic units involved in the error (see next slide).

5 The Psychology of Language, Trevor Harley Table 13.1 Examples of speech errors classified by unit & mechanism.

6 Garrett s Model of Speech Production We produce speech through a series of discrete levels of processing. Processing is serial, in that at any 1 stage of processing only 1 thing is happening. Levels of processing do not interact. 2 stages of syntactic planning: Functional level: word order is not yet explicitly represented. Semantic content is specified & assigned to syntactic roles (i.e. subject & object). Positional level: words are explicitly ordered.

7 The Psychology of Language, Trevor Harley Figure 13.2 Garrett s model of speech production, showing how the stages correspond to the processes of conceptualization, formulation, & articulation. Based on Garrett (1975,1976).

8 The Psychology of Language, Trevor Harley Table 13.2 An example of how we produce an utterance based on Garrett s (1975, 1976) model of speech production.

9 Garrett s Model of Speech Production Affixes: bound morpheme that cannot exist on its own, must be attached to a stem. Comes before the main word, when it is a prefix, or after, when it is a suffix. Prefixes: an affix that comes before the stem. Bound morphemes: a morpheme that cannot exist on its own. Dissociation: a process is dissociable from other processes if brain damage can disrupt it, while leaving the others intact.

10 Syntactic Planning Word order in speech is determined by a number of factors that interact. Speakers construct sentences so they provide given before new information. Gaze is a reliable indicator of what & when people are thinking & planning.

11 Syntactic Priming Processing of a particular syntactic structure influences processing of subsequently presented sentences. Structural priming/syntactic persistence: repetition of syntactic structure. Common method used to study syntactic priming: participants repeat a prime sentence that contains the syntactic structure of interest, & then describe a picture.

12 Studies have shown: Syntactic Priming Speakers use a particular word order if the prime sentence used that order. Syntactic priming does not depend on superficial similarities between the prime & utterance. Does not depend on reusing words (lexical priming) or on repeating thematic roles. Reflects the more general construction of syntactic constituent structures.

13 Syntactic Priming Does not depend on: Tense: in the past, present, or future. Number: whether 1 or more subjects are doing the action. Aspect: shows whether something is finished, continuing, or repeated. Progressive & Non-Progressive

14 Syntactic Priming Lexical boost: repeating the verb (regardless of tense, aspect, or number) does enhance priming. Priming is enhanced by the repetition of word order between prime & target. Syntactic persistence can last a long time. Suggests priming is due to more than short-term memory, & may have some long-term learning component.

15 Syntactic Priming Speakers tend to reuse the syntactic constructions of other speakers. Potter & Lombardi: participants silently read words presented 1 a time at fast rate on a computer screen, then performed a distracter task before being asked to repeat the sentence out aloud. People tend to reuse previous syntactic structures: they recalled the sentence just presented with the syntactic structure of a previous item. Syntactic priming influences memory.

16 Syntactic Priming Connectionist model of sentence production: sequencing in production makes use of two types of information. A sequencing system uses a recurrent connectionist model that uses statistical information to predict what is coming next.

17 Syntactic Priming 2 main functions: 1. Enables speakers in a conversation to coordinate or align information. 2. Results from implicit learning of how people use syntax to convey meaning people unconsciously adjust how they convey information on the basis of experience.

18 Dependencies Number-agreement errors: number agreement between subjects & verbs is incorrect. Bock & Eberhard: used a sentence-completion task designed to elicit agreement errors. Agreement cannot be determined by the sound of surrounding words but by something more fundamental.

19 Dependencies Seems likely that whether we find semantic effects on verb agreement depends on subtle factors (precise materials used in experiments). Haskell & MacDonald: showed number agreement can be accounted for in terms of constraint satisfaction.

20 Dependencies Bock, Eberhard, & Cutting: we need 2 processes to ensure that number agreement proceeds smoothly. 1. A specification that takes into account the number of things we are talking about in the message, processing marking. 2. To over-ride the syntactic process of marking with a process that takes account of the morphology of the subject, morphing.

21 Lexicalization The process in speech production whereby we turn the thoughts underlying words into sounds. We translate a semantic representation (the meaning) of a content word into its phonological representation of form (its sound).

22 Lexicalization Widespread agreement that it is a 2 stage process: First stage being meaning based. Second phonologically based. Lexical selection: choosing the word. Phonological encoding: retrieving the phonological forms of these words.

23 Lexicalization Lemma: level of representation of a word between semantic & phonological representations; syntactically specified, does not contain sound-level information; the intermediate stage of 2 stage models of lexicalization. Lemma selection: stage of specifying in a prephonological, abstract way the word that we are about to say. Lexeme/phonological form selection: stage of specifying actual concrete phonological form of the word.

24 The Psychology of Language, Trevor Harley Figure 13.3 Two-stage model of lexicalization.

25 The Psychology of Language, Trevor Harley Figure 13.4 An example of a search-based single lexicon model. Based on Fay and Cutler (1977).

26 Evidence for 2 Stages Speech errors: 2 distinct types of whole word substitution speech error: semantic substitutions, & form-based substitutions, or malapropisms. Experiments: Simple scenes. Repetition priming. Semantic-interference studies. Neuroscience: Different regions of the brain become activated in sequence as we produce words. Tip-of-the-tongue phenomenon: extreme form of a pause, where the word takes a noticeable time to come out.

27 Is Lexicalization Interactive? Interaction: the influence of 1 level of processing on the operation of another. Temporal discreteness. Reverse flow of information. When processing levels overlap, in that one level can pass on information to the next before it has completed its processing, is known as processing in cascade.

28 Time Course of Lexicalization Speech error evidence of the existence of mixed whole word substitutions indicates overlap or interaction between the 2 stages. Mediated priming effect: when you say sheep, it facilitates the recognition of the word goat.

29 The Psychology of Language, Trevor Harley Figure 13.6 Processes involved in naming an object in a picture, according to the two-stage model of lexicalization. Reproduced with permission from Dr. Willem Levelt, Max Planck Institute for Psycholinguistics.

30 Feedback in Lexicalization Familiarity bias: tendency for errors to produce familiar sequences of phonemes. Lexical bias: tendency for sound level speech errors such as spoonerisms to result in a word rather than a nonword (e.g., barn door being produced as darn bore ) more often than chance would predict. Similarity effects: when the error is more similar to the target according to some criterion than would be expected by chance. Mixed substitutions: intrusion is both semantically & phonologically related to the target.

31 The Interactive Activation Model Emerging consensus. of Lexicalization Harley s model lexicalization proceeds in 2 stages. 1. The meaning of a word is represented as a set of semantic features. 2. These feed into a level of representation where abstract lexical representations equivalent to the lemmas are stored, and these in turn activate the phonological representations equivalent to lexemes. See next slide.

32 The Psychology of Language, Trevor Harley Figure 13.7 Architecture of an interactive activation model of lexicalization. Arrows show excitatory connections; filled circles show inhibitory connections. The semantic within-level connections are more complex, with partial connectivity, as indicated by the unfilled circle.

33 Dell s Interactive Model Based on spreading activation. Items are slotted into frames at each level of processing. Processing units specify syntactic, morphological, & phonological properties of words. Activation spreads down from the sentence level, where items are coded for syntactic properties, through a morphological level, to a phonological level.

34 The Psychology of Language, Trevor Harley Figure 13.8 Dell s (1986) connectionist model of speech production.

35 Evaluation of Work on Lexicalization There is consensus that lexicalization in speech production occurs in 2 stages. Plenty of evidence that information cascades between levels. Less consensus on whether the stages are discrete, & on whether they interact.

36 Phonological Encoding Main problem: ensuring that the sounds of words come out in the appropriate order, with the appropriate prosody.

37 Phonological Encoding Frame-based models: Linguistic structures create frames with slots, & we retrieve linguistic content to fill these slots. A frame is stored for each word we know. Sound segments are retrieved separately and inserted into appropriate slots in a syllabic frame. When we speak, we produce an abstract frame for the upcoming phrase.

38 Phonological Encoding Frame-based models: The frame specifies the syllabic structure of the phrase. A scan-copier device works through a syllabic frame in left-to-right serial order selecting phonemes to insert into each position of the frame. As a phoneme is selected, it is checked off. Disruption of this mechanism leads to difficulty in sequencing sounds in words

39 Phonological Encoding Competitive queuing, connectionist model: Provides an explicit mechanism for inserting segments into slots. Segments to be inserted form an ordered queue controlled by processes of activation & inhibition. 2 control units, an initiation & end unit. Sounds at the start of a word have strong connections to a unit that controls the initiation of speech, sounds at the ends of words have strong connections to a unit that controls the end of the sequence.

40 Phonological Encoding Competitive queuing, connectionist model: Strength of connections of other sounds in a word to these control units varies as a function of their position in a word. After a sound is selected, it is temporarily suppressed. Failure to do this properly leads to perseveration errors.

41 Phonological Encoding Recurrent networks: Associate words with phonological representations in sequence, without any explicit representation of the structure-content distinction. External feedback: copied the output of the most recent segment, & provided the model with memory of the past phonological states of the model. Internal feedback: copied the past state of the hidden units of the network, & provided the model with memory of its past internal structure.

42 Phonological Encoding Recurrent networks: When model made errors, it exhibited 4 properties observed in human sound speech errors. 1. Obeyed the phonotactic constraint. 2. Consonants exchanged with other consonants, & vowels exchanged with other vowels. 3. The syllabic constituent effect. 4. Initial consonants are more likely to slip than non-initial.

43 Phonological Encoding Lemma Model: WEAVER++: a discrete two-stage model without interaction between levels. See next slide.

44 The Psychology of Language, Trevor Harley Figure 13.9 The Weaver++ computational model. Adapted from Levelt et al. (1999).

45 Hesitations Unfilled pause: is simply a moment of silence. Filled hesitation: can be a filled pause (where a gap in the flow of words is filled with a sound such as uh or um ), a repetition, a false start, or a parenthetical remark (such as well or I mean ).

46 The Psychology of Language, Trevor Harley Figure Speech Dysfluencies.

47 Pausing Before Individual Words There are many pauses within a single conversational turn. They tend to occur every 5 to 8 words. Goldman-Eisler: they are more likely to occur, & are longer in duration, before less predictable words. Beattie & Butterworth: main component of predictability that determined hesitations was difficulty in semantic planning.

48 Pauses & Sentence Planning Goldman-Eisler: in some pauses we plan the content of what we are about to say. Ford & Holmes used dual-task performance: Participant had to speak while monitoring for a tone over headphones. Planning does not span sentences because reaction times to the tone were no longer at the ends of sentences, suggesting that people are not planning the next sentence at the end of the previous one.

49 Pauses & Sentence Planning Holmes asked participants to read several sentences that began a story, & then produce a 1-sentence continuation. Some speakers produced more than 1 sentence, & when they did so a pause was more likely at the start of their speech than when they produced only 1 sentence. Different tasks seem to indicate that different units are the fundamental unit.

50 Aphasia Broca s aphasia: non-fluent speech, characterized by slow, laborious, hesitant speech, with little intonation, obvious articulation difficulties and obvious impairment in the ability to order words. At the most general level, Broca s-type patients have difficulty with sequencing units of the language.

51 Aphasia Wernicke s aphasia: production of fluent but often meaningless speech. Patients speak in well-formed sentences, with copious grammatical elements & normal prosody. Comprehension is noticeably poor, & there are major content word-finding difficulties, with many word substitutions & made-up words.

52 The Psychology of Language, Trevor Harley Figure Comparison between Broca s and Wernicke s aphasias.

53 Agrammatism Literally, without grammar ; a type of aphasia distinguished by an impairment of syntactic processing (e.g., difficulties in sentence formation, inflection formation, & parsing).

54 Jargon Aphasia Extreme type of fluent aphasia. Syntax is primarily intact, but speech is marked by gross word-finding difficulties. Difficulty in recognizing that their speech is aberrant, & may become irritated when people fail to understand them. Word-finding difficulties are marked by contentword substitutions (paraphasias)& made-up words (neologisms).

55 Anomia Impairment of retrieving the names of objects & pictures of objects. Can be found in isolation, or with other disorders. Virtually all types of aphasia are marked by some degree of anomia. Lexical semantic anomia: inability to use the semantic representation to select the correct lemma Phonological anomia: poor at retrieving any phonological information about the target.

56 Aphasia Global aphasia: spontaneous speech, naming, & repetition are all severely affected. Crossed aphasia: disorders of language arise from damage to the right hemisphere, even in right-handed people. Conduction aphasia: repetition is relatively impaired, while production & comprehension are relatively good. Reproduction conduction aphasia: repetition is poor because of poor phonological encoding. Impairments in all language production tasks (speaking, repetition, reading, & writing).

57 Aphasia STM conduction aphasia: repetition is poor because of an impairment of input auditory short-term memory; few errors in spontaneous speech production, but repetition of strings of short familiar words is poor. Transcortical sensory aphasia: comprehension is impaired, output is fluent and may even include jargon, but repetition is relatively good. Transcortical motor aphasia: comprehension & repetition are very good, but there is very little spontaneous speech output.

58 Writing Hayes & Flower: 3 stages of writing. 1. Planning stage: goals are set, ideas are generated, and information is retrieved from long-term memory and organized into a plan for what to write. 2. Translation stage: written language is produced from the representation in memory. The plan has to be turned into sentences. 3. Reviewing: the writer reads & edits what has been written.

59 Neuropsychology of Writing Phonic mediation theory: when we write, we first retrieve the spoken sounds of words & then produce the written word. Neuropsychological data show that this theory is almost certainly wrong! There are patients who can spell words that they cannot speak (inner speech is not necessary for writing).

60 Summary Speech production has been studied less than language comprehension because of the difficulty in controlling the input (our thoughts). Speech production can be divided into conceptualization, formulation, and execution. In the two-stage model of lexicalization, a meaning-based stage is followed by a phonologically based stage. Tip-of-the-tongue (TOT) states are noticeable pauses in retrieving a word; they arise because of insufficient activation of the words in the lexicon.

61 Summary Speech errors show lexical (familiarity) and similarity biases; these findings suggest that lexicalization is interactive. It is not clear why feedback connections exist. The main problem for phonological encoding is ensuring that we produce the sounds in the correct sequence. Hesitations reflect planning by the speaker, although they may also serve social and segmentation functions.

62 Summary Aphasia is an impairment of language processing following brain damage. Agrammatism covers a number of aspects of impaired syntactic processing, including a sentence construction deficit, loss of grammatical elements of speech, & impaired syntactic comprehension.