Transcript Addendum: How to study issues in spoken language comprehension

SPOKEN LANGUAGE
COMPREHENSION
Anne Cutler
Addendum:
How to study issues in
spoken language comprehension
The psycholinguist’s problem
• We want to know HOW spoken language is comprehended
• But the process of comprehension is a mental operation,
invisible to direct inspection
• So we have to devise ways of looking at the process
indirectly (in the laboratory, mostly)
• These laboratory methods often involve measuring RT - the
speed with which a decision is made, or a target detected
• It is important to know how to relate the laboratory task to
the processes one wants to study!!! (a “linking hypothesis”)
• And also to make sure that the task is reflecting what we
want it to reflect i.e. there are no uncontrolled artifacts,
which might provide alternative interpretations….
The lectures so far
• Lecture 1: speech is fast and continuous, but
somehow listeners have to identify the words
in it, because words are known entities, and
identifying those known entities and putting
them together is the only way to reach the
(unknown) goal, i.e. the speaker’s message.
But words themselves are not unique – they
resemble one another and can be embedded
within one another…. So speech can contain
many spurious words, not part of the message
• So do listeners process only the words which
the speaker uttered, or do other words
become activated and have to be eliminated?
The lectures, contd.
• Lecture 2: all the words which are simultaneously
activated compete with one another
• Lecture 3: although competition alone could explain how
segmentation occurs, there are also processes of
segmentation, and they differ across languages
• Lecture 4: the process of activating words involves
continuous evaluation of information in the speech signal;
there is no necessary intermediate stage in which such
representations as syllable, mora are extracted
• Lecture 5: the lawful phonological processes which result
in, say, lean being spoken as leam, or petit being spoken
sometimes with a final vowel and sometimes with a final
consonant, neither disrupt nor facilitate processing
LEXICAL DECISION
(hear words and decide: is that a real word?)
Lexical decision is the
simplest word processing
task. For instance, it can be
tell us whether words can
be recognised, or nonwords
rejected, as soon as enough
of them is heard that no
other words are possible
(the “uniqueness point”).
Auditory lexical decision is
dependent on word length –
no response possible until
end because it could always
become a nonword after all!
CROSS-MODAL PRIMING
(Hear prime, see target; decide: is it a real word?)
Prime and
target may
be identical
(e.g. giveGIVE), or
related by
association
(give-TAKE)
CMP is a way of looking at what is activated when a listener
hears speech – we measure the RT to decide whether the
visual target is a real word, and if that RT varies when the
spoken prime varies, then we have observed an effect of the
spoken prime. Hearing “give” makes recognising TAKE easier.
CROSS-MODAL FRAGMENT PRIMING
(decision: is that a real word?)
Whether the prime
and target are the
same or related
by association is
one factor;
whether the prime
is presented as a
whole or just as a
fragment is a
separate factor
Likewise, hearing “octo-” makes recognising OCTOPUS easier
Eye-tracking experiment
Eye-tracking also looks at activation – listeners who hear “ha-”
look at the ham or the hamster – both are potential candidates
GATING
(hear a word in fragments of increasing size, and at
each fragment guess what the word is)
• E.g. p- pr- pra- prak- pract- practi- practik• The “gated” fragments can be of constant
size (e.g. 50 ms, 100 ms, 150 ms etc.); or
they can systematically add more phonetic
information (e.g. each fragment adds another
phoneme transition – Fragment 1: to middle
of 1st phoneme; 2: to middle of 2nd phoneme;
3 to middle of 3rd phoneme; etc.)
• Gating tells us what information can be used
at a particular point – but it is a problem that
listeners sometimes stick with bad guesses…
PHONEME DETECTION AND FRAGMENT DETECTION
(hear speech, listen for target phoneme or fragment)
These are also very simple tasks. They probably don’t directly
reflect prelexical processing, but they can reflect how easy it is
to extract information (below the word level) at a given point.
Thus they might reflect segmentation, or how easy a preceding
word or phoneme was to process, etc.
WORD SPOTTING
(hear nonsense item – is there a real word in it?)
Word spotting is especially good for looking at word recognition
in context – by minimising the context, we can look at the local
effect of a context on how hard or easy it is to find a word
WORD RECONSTRUCTION
Change a nonword into a real word
by altering a single sound
This task was used to
look at the processing of
vowels and consonants –
which type of phoneme
constrains word identity
more strongly? Then, it
was also used to look at
whether rhythmic
categories like the mora
are used in recognising
words
PHONETIC CATEGORIZATION
(hear artificial sounds, decide what they are)
Listeners normally hear speech sounds which are
more or less good exemplars of their categories.
It is possible to make an artificial continuum from one
sound to another, and present these sounds to listeners.
What they report in the middle is not new non-sounds,
but a sudden switch from tokens of one category to
tokens of the other - “categorical perception”.
The phonetic categorization task (developed for phonetic
research) has also been useful in psycholinguistics. E.g.
categorical functions can shift if one decision would make
a word but the other would make a nonword (Ganong)…
Other tasks
Blending task: Construct a blend of two pseudo-names, using
the 1st part of the 1st name and the 2nd part of the 2nd name.
Has been used to look at what information in the signal is more
vs. less important (e.g. is place of articulation unspecified?)
Reversal task: A sort of language game – reverse the parts of
a word (e.g. syllables, phonemes….). Has been used to look
at people’s internal representations of words (e.g. are syllable
boundaries clear? Are intervocalic consonants ambisyllabic?).
Artificial language learning: Hear nonsense input, try to learn
the “words” (and other structures) it consists of. Has been
used to look at how easily different sequences can be
segmented, and whether listeners have expectations about
what words will be like.
What the tasks told us
• For instance: there are lots of types of converging
evidence for multiple concurrently active words – after
capt- BOTH captain and captive are facilitated etc.
• (N.B. Different tasks look at the same aspect of
processing)
• For instance: segmentation relies on language-specific
information
• Different tasks look at the same aspect of processing,
again – e.g. with word-spotting we discover languagespecific segmentation procedures, then we can predict
that listeners will use these procedures also in learning
new languages, and test this with an artificial
vocabulary learning experiment
Summary
• Studying spoken language comprehension can’t
be done directly (we can’t look into the brain), only
indirectly, with the help of laboratory methods
• This means that we have to translate the bigger
questions we are interested in into questions which
can be answered using our laboratory methods
• For instance: can Finnish listeners use vowel
harmony to help them find word boundaries? We
turn that into a smaller question: is a boundary
easier to find if the vowels on either side of it are
disharmonious rather than harmonious? And then
we can use word-spotting – a real word, abutted to
a nonsense context (harmonious or disharmonious)