Upload
julian-heath
View
219
Download
1
Embed Size (px)
Citation preview
Common Ground
• Linguistic referents are established w/in a “domain of interpretation”, which includes context
– One component of context = Common Ground• Mutual knowledge, beliefs, & assumptions among
participants in conversation• Comes from community co-membership, physical co-
presence, linguistic co-presence, …
Conversation & Common Ground
• Clark (1996) – Language = Joint action by people cooperating
to achieve particular goals– Optimal communication requires keeping track
of what’s in Common Ground and using that in both producing and understanding language
• As usual, what’s at issue is timing– How quickly do/can language users make use of
knowledge about Common Ground when speaking & listening?
Keysar & Colleagues (1996, 2000)
• In several studies, find that listener’s immediate interpretations are “egocentric”– i.e., Don’t take into account which knowledge they
share with speaker vs which knowledge only they have
– Monitor eye movements to visual display containing referents for items mentioned in speech
• Speaker (confederate) gives listener spoken instructions to move objects in display
• Some objects visible to both (Common Ground)• Some objects visible only to listener (Privileged Ground)• Listeners just as likely to look at Privileged objects as at
Common objects, if they match instructions best
Problem with Keysar’s Studies
• Very best physical match for instruction = object in Privileged Ground– “Put the bottom block below the apple.”
• When bottom-most block in listener’s display is visible only to listener
• Listeners just as likely to look at the bottom-most block in their display as to look at the block that they know is bottom-most in the speaker’s display
• Better to make the object in Common Ground just as good a match for the instructions as the object in Privileged Ground is
Another Potential Problem
• Common Ground is typically accumulated and adjusted incrementally over the course of conversation– In Keysar’s studies, listeners simply instructed
about what was visible to whom
• Better to use an explicit grounding process with common ground established on the basis of linguistic co-presence
Hanna et al. (2003) – Experiment 1
• Participants– 12 Listeners (L) +Confederate Speaker (S) pairs– Listener told that Speaker was lab assistant but
naïve about purpose of study (false)
• Apparatus = Head-mounted eyetracker– Eye camera tracks pupil using infrared light– Scene camera shows line of sight, so head can
move & still know where eyes are looking– Spatial resolution = 1 degree– Temporal resolution = 33.33 msec
• Sampling rate = 30 frames/sec
Experiment 1 Procedure
• Materials (more on next slide)– Boxes containing 3 x 3 grid of locations– For each trial, 7 objects each for S & L
• 5 are to be placed in the grid by L according to S’s instructions
– S and L don’t see each other’s grids or objects
• Instructions– On each trial, get envelope with 7 objects + printed
instructions– S’s instructions show layout of objects for trial + script of
instructions to give Listener– L’s instructions include which of the objects is “secret” (= Privileged Ground) & where to put it in grid
• L knows S doesn’t know what shape is secret or where it is (true)
Visual Display Stimuli
Target = Red triangle
Competitor = Other triangle- Same or diff color from Target
S = Secret Shape, seen by Listener only= Competitor or some other shape
When Competitor = Secret, it’s in Privileged Ground- Will they look at it as much as they do to a Competitor that’s in Common Ground???
Competitor in Competitor in
Critical (= last) instruction from Speaker
Competitor same color as Target
“Competitor” diff color from Target
Same color competitor in Common Ground- L has to ask for clarification in this cond
Same color competitor in Privileged Ground
Fewer looks to Competitor in Privileged Ground by 400 msec after onset of Adj - Very similar to diff color conds
Hanna et al. (2002) – Expt 2
• Possible criticism of Expt 1 (& other expts w/ similar design)
– By design, Confederate Speaker doesn’t know about Privileged Ground objects & so never mentions them
– Maybe Listeners move their eyes to objects that are more likely to be mentioned by Speaker
• Rather than taking Speaker’s perspective into account while interpreting referring expressions?
• Solution = Give Speaker & Listener different information about the state of some objects
• Listener given 2 pairs of objects of same type– 2 jars, 2 martini glasses, …– Objects described aloud by
Exptr, left to right
• Sometimes objects described inaccurately– “2 empty jars”, when there is
only 1 empty jar & 2 empty martini glasses
– Listener told will be mistakes sometimes but not to talk with Speaker to correct them
• Told there’s another condition in Expt where Listeners get to make corrections
Stimulus Displays & Descriptions
Instructions & Displays
• Display properties disambiguate instruction Early or Late– “Pick up the empty martini glass” – Late: 1 empty version of
both types of objects
– Early: 1 empty version of 1 type of object & 2 empty versions of other type of object
• Disambiguates early because definite “the empty” can have a unique ref only if only 1 of objects of a type is empty
Definite vs Indefinite Displays & Instructions
“Pick up the emptymartini glass.” vs“Pick up one of theempty martini glasses.”
Late
Early
Late
EarlyMismatch
Speaker toldthis
But Listener seesthis
Design Logic
• Listeners told to remember how objects were described to Speaker & to do what they think the Speaker intends them to do
• In Mismatch conditions, which type of object is the referent is disambiguated Early– But it’s a different object type for Speaker & Listener– Instruction: “Pick up the empty martini glass”
• Speaker believes: Only 1 empty martini glass & 2 empty jars• Listener sees: 2 empty martini glasses & only 1 empty jar
• Questions– Will Listener look at glass she knows Speaker thinks empty?– If yes, how soon, compared to Early & Late Match conditions?
Results
Late
Early
Mismatch
Conclusions
• Listeners look at target faster when instructions pick out a unique referent earlier– “the empty” when only 1 pair of objects has only 1 empty
version
• Listeners can quickly take Speaker’s perspective into account – Look at target faster in Mismatch than in Late conditions
• From Speaker’s perspective, instruction picks out a unique referent early
• From Listener’s perspective, picks out other object type
– But not as fast as in Early conditions• So there’s some effort in taking Speaker’s perspective
• No evidence here for an initial egocentric stage