1. Language production Holly Branigan Email: Holly.Branigan@ed.ac.uk Office: US46 Office hour: Mon 10-11 Psychling coffee hour: Wed 11-12 Holly.Branigan@ed.ac.uk

2. Course overview Overview of the production system Methodological issues Lexical access Syntactic encoding Beyond the sentence…

3. Overview What does production involve? Methodological challenges: – Studying comprehension vs production Approaches: – Observational approaches Advantages/disadvantages – Experimental approaches Challenges: controlling input and output Two classes of methods: –Manipulating pathways (altering processor’s state) –Manipulating message – Neurophysiological approaches

4. Background Production forms half of language ability: – Input to comprehension – More difficult problem than comprehension? e.g. Evidence from 1st & 2nd language acqn The problem: – Expressing non-ordered conceptual message via ordered array of sounds. – But: under several constraints, in real time.

5. What we don’t do H: How much money is there in my current account and in my deposit account? H: Hello? C: Colourless green ideas sleep furiously. H: How much money is there in my current account and in my deposit account? C: Your current a-ccount encompasses two hundred pounds. I cannot access how.... in your deposit account money much is there.

6. Undesirable features Meaningless and irrelevant content. Long silences, strange pausing. Infelicities of vocabulary and structure: – ‘Your current account encompasses £200’ – ‘I cannot access how in your deposit account money much is there.’ Strange intonation and pronunciation: – ‘Your current a-ccount’ – ‘Sleeeeeep’

7. What we do do Speakers must produce utterances with: – Appropriate meaningful content; – Appropriate lexical items; – Appropriate syntax - grammatical and appropriate word order and structure; – Appropriate pronunciation, intonation, and phrasing. And they must do this fluently, in real time.

8. Getting the form right Speakers have to get every aspect of the form right, whether or not germane to message. – cf. Hearers - details of form can sometimes (often?) be ignored (e.g. missing words, not paying attention).

9. Getting the content wrong Paradox: adept at getting form right but content wrong: – Subject-verb agreement errors e.g. The report about the fires are very long Less than 5% errors in expmt designed to elicit them (Bock & Miller 1991). – Serious structural anomalies (unparseable) 0.5% utterances (Deese 1984). – Sound/word errors (Garnham et al 1982): Sound errors 3.2/10,000 words Word errors 5.1/10,000 words – Can you put the desk back on my book when you’ve finished with it? – It’ll get fast a lot hotter if you put the burner on.

10. Doing it in time Strongest constraint may be fluency: – have to get form right under time pressure. Incrementality: – ‘Work with what you’ve got’ – Flexibility: allows speaker to say something quickly, also respond to changing environment. Modularity: – ‘Work only with what you’ve got’ – Regulate flow of information.

11. An outline of sentence production Three broad stages: – Conceptualisation deciding on the message (= meaning to express) – Formulation turning the message into linguistic representations Grammatical encoding (finding words and putting them together) Phonological encoding (finding sounds and putting them together) – Articulation speaking (or writing or signing)

12. Methodology: Background ‘..an intrinsically more difficult subject to study than language comprehension’ – Not susceptible to experimental study? Solutions: – Evidence from other disciplines e.g., social psychology, linguistics, AI… – Cognitive psychology: Historically: observational methods Recently: experimental methods

13. What’s the problem? Comprehension: – Can control input precisely – Moving from language to conceptual representation e.g., understanding anaphora: participants read same texts; measure reading times Production: – How do we control input? – Moving from (unobservable) conceptual representation to language e.g., when participants produce anaphora, do they do so on the same basis? BUT: end product is observable in production but not comprehension

14. Measures What people say: – Under which circumstances do they produce particular words, utterances etc – May be intended, or may be errors – How frequently do they do this Timecourse: – How quickly do people produce language Neurophysiological: – How is language production represented in the brain?

15. Observational methods: Analyses of spontaneous speech – Researchers’ own corpora (e.g., Stemberger, 1985) – Publicly available corpora: Non-experimental (London –Lund - Svartvik & Quirk, 1980; Wall Street Journal; CHILDES – MacWhinney & Snow, 1990) Experimental (controlled features) (Map Task Corpus – Thompson et al., 1993). – Controlled experimental tasks: Berman & Slobin, 1994.

16. Observation: Distributional analyses Fluent speech: – Sentence types, verb forms, prosodic markers etc (Deese, 1984) – Distribution of extraposed structures (Arnold, Wasow, Losongco & Ginstrom, 2000) – Distribution of thuh vs thee (Clark & Fox-Tree, 1997) – Distribution of reduced phonological forms (Bard et al., 2000) Disfluent speech: – Scope of utterance planning (Ford & Holmes, 1978; Beattie, 1983) – Error detection and correction (Levelt, 1983)

17. Observational analyses of disfluencies Speech errors – Pattern of errors (Stemberger, 1985) – Relative frequency of errors Problems: – Paucity of data errors = 3% self-interruptions (Blackmer & Mitton, 1991) – Bias/inaccuracies in corpus transcription: Transcriber bias/inaccuracy (Ferber, 1991) Distributional characteristics of language – Categorisation problems

18. Experimental approaches Not prey to same problems as observational studies: – Reduces observer bias; – isolates phenomenon of interest; – increases potential for systematic observation. Different problems! – How to control input and output? – Input: ecological validity problem (‘controlling thoughts’) – Output: controlling responses: response specification - artificiality ‘exuberant responding’ – loss of data

19. Controlling input Philosophical problems – Does language production require ‘freedom of thought’? Practical issues: – Problem: how to characterise non-linguistic message? – Solution: hold message constant, and manipulate ‘pathway’ of processing instead (state of processor) –Priming paradigms (effects of prior processing) –Creating conflicts (cf. ambiguity resolution in comprehension)

20. Controlling output: Specified elicitation Specified elicitation: tell participants what to say. – Usually used when semantic/syntactic structure not of interest. – Responses specified in advance for given stimulus: Picture naming Implicit priming (Roelofs & Meyer, 1998) –DOG – BONE –SAIL – BOAT Array description (Smith & Wheeldon, 2001) Repeating sentences (Ferreira, 1993)

21. Controlling output: Normative elicitation Normative elicitation: use stimuli designed to induce desired response. – Pictures of events/objects – Descriptions of objects ‘A very large mammal that swims in the sea and was widely hunted’ – Questions/fragments ‘The junior surgeon handed the senior surgeon….’

22. Manipulating pathways: Error elicitation Basic idea: – set up situations which lead to errors in ‘natural’ speech Agreement errors: – Participants repeat and complete sentence fragments: The key to the cabinets…were heavy – Cause of errors: Conflict between number (or gender) of head and local noun – Used to examine e.g. contribution of conceptual info and morphology to agreement (Bock & Eberhard, 1993; Vigliocco, Butterworth & Semenza, 1995)

23. Manipulating pathways: priming Priming: – change probability/ease of producing particular utterance. – Cooperating (rather than competing) plan. Concurrent presentation: – Distractor and target presented at same time Consecutive presentation: – Distractor presented and processed before target

24. Manipulating pathways: Concurrent presentation Picture-word interference: – Target stimulus: –presented visually –must be named – Distractor stimulus: –presented auditorily or visually –must be ignored – Stimulus onsets may be simultaneous or staggered

25. Manipulating pathways: Concurrent presentation Used for exploring timecourse of lexical access (Schriefers, Meyer & Levelt, 1990) – Targets were objects such as sheep – Distractors: Different relations to target: –Phonologically-related (sheet) –Semantically-related (goat) –Unrelated (bed) Different presentation onsets: –150 ms before target –Simultaneous with target –150 ms after target

26. Manipulating pathways: Concurrent presentation Results: – early in timecourse: semantic distractors slow naming more than unrelated or phonological distractors; – later in timecourse: phonological distractors speed naming more than unrelated or semantic distractors. Control experiment used non-production (recognition) task – Excluded comprehension-based explanation?

27. Manipulating pathways: Consecutive presentation ‘Prime’ stimulus processed – Cf concurrent presentation, where distractor stimulus is ignored Target then processed – How does prior processing of prime affect processing of target?

28. Manipulating pathways: Consecutive presentation Word-priming (Wheeldon & Monsell, 1994) – Participants read dictionary definitions and generate response ‘A very large mammal that swims in the sea and was widely hunted’ – Then picture of related object (shark) presented – Here, slower responses when prime is related than unrelated – Attributed to competitive activation

29. Manipulating message Oldest method of studying production? – Create minimal contrasts in intended message; – study differences in realisation of message. ‘Simply describe’ (Osgood, 1971) –Enact (or show film) of minimally distinct events e.g., ball rolling across table vs man holding ball before ball rolls across table –Use of indefinite article a in first case vs definite article the in second case Very simple method – but many problems: – How do we characterise ‘minimal semantic contrast’?

30. Neurophysiological Measures Recent technological developments allow research on neurophysiological aspects of production. – Which areas of the brain are involved? – What is the timecourse of processing? – Are different areas/processes/timecourses associated with different aspects of production?

31. Some Methods Event-related potentials (ERPS): – brain responses time-locked to some "event“ – sensory stimulus (visual flash or auditory sound), mental event (recognition of a specified target stimulus), or omission of stimulus (increased time gap between stimuli). functional magnetic resonance imaging (fMRI): – form of magnetic resonance imaging of brain registering blood flow to functioning areas of the brain Positron emission tomography (PET): – uses detection of subatomic particles to identify how different areas of brain function. NB: other methods coming into use

32. Summary Language production requires assembling multiple levels of linguistic structure accurately and fluently, in real time. Language production in some ways harder to study than comprehension: – How to control input? Many methods: – keep propositional content constant – create and study variations in processing mechanisms, rather than effects of variations in message itself. – Problem remains: what is relationship between conceptual and linguistic processing? New technologies offer new possibilities for tracing timecourse and neurophysiological underpinnings of language production

33. Assessing models of production Semantic interference effect – Objects are harder to name in presence of semantically-related word. – Effect may be related to conceptual processing or feedback from phonological processing. fMRI study: – present same stimuli, and see which areas of brain activated. – results: differential activation of various areas in semantic-interference condition relative to control condition; consistent with phonological feedback. (de Zubicaray,Wilson, McMahon & Muthiah, 2001)

34. Identifying neural bases of production Grammatical gender: – central aspect of lexical representation in many languages. fMRI study: which areas of brain activated in gender production? – compared producing gender-marked determiner with naming object itself. – results: pronounced activation of single region in Broca’s area when producing determiner. (Heim, Opitz & Friederici, 2002).