Speakers are sometimes overinformative / overspecify their referring expressions. Why? When? Generally: hearer- vs speaker-oriented reasons for overinforming/overspecifying REs.

• incrementality (buy time, not enough time to figure out what's discriminating and what isn't, Pechmann 1989)

• perceptual salience (especially color is overspecified -- but what's the reference for there being variation within color? variation within color suggests it's not something monolithic about color, but something about the function that different colors in different contexts serve)

• overspecification can be detrimental for listeners (Engelhardt et al 2010; Sedivy et al 1999)

• facilitation of hearer's reference resolution (subset identification, hearer incrementality considerations, Deutsch 1976, Nadig & Sedivy 2002, Paraboni et al 2006, Engelhardt et al 2006, Arts et al 2011, Davies & Katsos 2009) -- think about the fact that in general, visual scenes are incredibly complex, it may be that on average, color is more helpful in carving up world perceptually than size is. More generally: things that are easier to distinguish perceptually should be more involved in overspecification?

• levels of informativeness interact with discourse goals (unclear whether it's speaker- or hearer-oriented, Maes, Arts & Noordman 2004): more overspecification when teaching hearer a long-term skill than on one-shot tasks; more overspecification when perceptual common ground is high; generally: Davies rightly says overspecification is expected when precision is important and risk of misunderstanding high

• Pechmann 1989 brings up information theory: redundant information is good because it can help compensate for partial loss of information. He then immediately dismisses this by saying that the redundant information isn't truly redundant because it's not distinguishing and at best can reduce the size of the domain. --> but when in dire straits, that may be a good start!!

• most studies only take into account current context while ignoring past experience

• issue of conceptual pacts / lexical entrainment -- sometimes you don't want to let go of the information that was at one point helpful, simply because it's now easy to retrieve from memory

• contrastive inference (presence of extra modifiers suggests contrast, ie gives additional information about background, via reasoning about why the speaker went to the trouble fo producing extra modifiers when they wouldn't otherwise be necessary; but we want to retain assumption of speaker cooperativity, so they must be necessary for disambiguation, so there must be a contrasting object present) --> ** but then what does this mean for all the cases that are truly overinformative?**

• overspecification may result from incomplete consideration of common ground how is this different from the speaker-oriented reasons above?

• visual processing and optimal linguistic encoding may compromise each other through sharing of cognitive resources

• lower overspecification rates in simpler visual arrays (Davies & Katsos, 2009, 2013)

• more overspecification as number of distractors increases (Pechmann 1989, Koolen et al 2011)

• color vs size: people overspecify color 21% of the time (Pechmann 1989)

• 50% overspecification of size adjectives with a privileged contrast (Nadig & Sedivy 2002)

• 33% location overspecification in "put the apple on the towel on the other towel" type sentences in one-referent contexts (Engelhardt et al 2006)

• "type and colour omnipresent" (Pechmann and others) -- conclusion: these help create a perceptual gestalt there are syntactic reasons to mention the type because they're usually nouns -- what about languages where they're not nouns? what about other languages more generally?

• more overspecification when talking to a real speaker (van der Wege 2009)

• Mitchell, Reiter, and Van Deemter (2013) investigate effects of atypicality and show that speakers prefer to mention the shape or material of objects when it is atypical (e.g., "octagonal mug", "wooden key") --> but they didn't look for a gradient effect, only binary typical/atypical

• For objects that have one or more typical colors associated with them (i.e., color-diagnostic objects, for example tomatoes which are typically red), this knowledge contains color information (e.g., Tanaka, Weiskopf, & Williams, 2001).

• Recognition is slower when (color-diagnostic) objects are presented in atypical colors than when they are typically colored (e.g., Naor-Raz, Tarr, & Kersten, 2003; Tanaka et al., 2001; Therriault, Yaxley, & Zwaan, 2009).

• Furthermore, the contribution of color in object recognition is stronger for objects with a simple and unchar- acteristic shape (e.g., oranges) than for objects with more complex shapes (e.g., fire trucks; Price & Humphreys, 1989).

• objects that have a color different from stored object knowledge are known to attract visual attention. Becker, Pashler, and Lubin (2007) eye-tracked participants who were presented with naturalistic scenes containing an object in an atypical color (a green hand), or in a typical color (a flesh-colored hand). Participants fixated earlier, more often, and longer on the green hand than on the normal hand. This result could not be ascribed to green being more salient than flesh-color, which Becker et al. controlled for by swapping the hand's color with a mug (which is equally typical in green or flesh- color).

• pose a challenge for RSA (Frank & Goodman 2012 style): people overspecify, so RSA is never going to make the right predictions for situations in which you don't artificially restrict the alternatives.

• various REG algorithms:

  • Full Brevity Algorithm (Dale 1989): search expressions in order of length until distinguishing one is found. Problem: intractable in the worst case.
  • Greedy Algorithm (Dale 1989): incremental adding of a single property at a time, at each stage considering which one excludes the greatest number of distractor items
  • Incremental Algorithm (Dale & Reiter 1995): start to de-emphasize discriminatory power, start to emphasize preference/salience. Properties arranged in fixed linear order determined by degree of prefernce; go through list and include property if at least one distractor is excluded
  • RSA (Frank & Goodman 2012): basically the Greedy Algorithm with stochasticity; can't get all the preference-based results --> Judith, make sure you get a good overview of what the preference-based results are!! eg they claim that "overspecification may actually be detrimental to listeners", citing Engelhardt et al -- read this! --, but it's not clear that it always is; find the counter-refs; this is important because it affects how strongly you can sell the `people aren't rational' line
  • PRO (van Gompel et al 2012): motivated by the color-over-size-preference lightbulb experiment; combines discriminatory power & preference, as well as stochasticity; 2 parameters: preference for color, eagernes to overspecify

Two conditions (low vs high variation) in each experiment, using furniture stimuli that could vary in four features:

• type: chair, sofa, fan, television, desk
• color: red, blue, green, brown, gray
• orientation: front, back, left, right
• size: large, small

Each critical trial contained eight items: one target (for which color was not necessary for identification) and seven distractors.

Fillers: four pictures of Greebles. only 4 instead of 8 -- ie clearly distinguishable from critical trials. Problem?

1. Exp. 1: on critical trials (20 out of 60), mentioning type was sufficient to distinguish the target.

   • low variation: furniture items differed only in type
   • high variation: objects differed in all features (but Fig. 2 suggests that color was in fact it would not be overinformative if we assume you don't have to mention a noun, though it's not clear if it was on all trials)
   • result:
     • low variation: 4% (2% sd) color mention
     • high variation: 25% (6% sd) color mention

2. Exp. 2: same as Exp. 1, with following exception:

   • low variation: items differed in type, orientation, and size (ie low variation condition has more variation than in Exp. 1)
   • mentioning type and one other distinguishing feature (not color) was sufficient to distinguish
   • result:
     • low variation: 9% (5% sd) color mention
     • high variation: 18% (5% sd) color mention

3. Exp. 3:

   • low variation: furniture items differed in orientation and size
   • high variation: furniture items differed in type, color, orientation, and size
   • mentioning type and one other distinguishing feature (not color) was sufficient to distinguish
   • result:
     • low variation: 10% (5% sd) color mention
     • high variation: 27% (8% sd) color mention

Their between-experiment analysis didn't show any differences in overspecification rates between experiments, only between conditions. So these minor modifications in degree/type of scene variation have no effect? See whether this is predicted by model.

On all critical trials, mentioning color was redundant. Filler trials consisted of greebles (no color). All critical trials were separated by a filler. Each critical scene contained six objects, positioned randomly in a three by two grid, in three different colors (each on two objects such that the target's color was never unique within a scene). Typicality was equalized in the scenes by selecting two somewhat typical objects, two atypical ones, and two falling in-between typical and atypical, for a mean typicality score over all objects in a scene between 40 and 60. (Typicality scores estimated in pre-test.) They only analyzed cases where type was mentioned (otherwise it's unclear whether the color is typical).

1. Exp. 1:

   • 42 critical trials.
   • Types (14, each one appearing three times with colors that spanned typicality range): apple, banana, carrot, cheese, corn, grapes, lemon, lettuce, orange, pear, pepper, pineapple, pumpkin, tomato
   • Result: Color mention decreased with increasing color typicality, r=.86.

2. Exp. 2:

   • 16 critical trials and 32 greeble trials.
   • Types (16, half simple from Exp. 1 and half complex). Color typical or atypical. Each participant saw each item in only one color condition.
     • Simple: basketball, lemon, lettuce, orange, strawberry, tennis ball, tomato, watermelon
     • Complex: broccoli, carrot, cheese, chick, crocodile, gold- fish, lobster, phone booth
   • Result: replication of typicality effect from Exp. 1 (categorical rather than continuous result). In addition: more redundant color use for complex than simple objects (THIS FITS WITH THE KOOLEN ET AL 2013 SCENE VARIATION RESULT.)

   Color typicality	Shape complexity	Proportion of color use
   typical	simple	12%
   typical	complex	16%
   atypical	simple	65%
   atypical	complex	85%

Has anyone compiled a list of overspecification rates for different modifiers (color, size, location, age, orientation, etc)?

General focus on definite descriptions: definite determiner, head noun, one or more (pre- or post-nominal) modifiers. What happens if we move away from that?

Adjectives have typical ordering preferences (Pechmann 1989 and references therein): eg in English size before color, but violations are OK when there's contrastive stress on the color adjective or when color is particularly salient. Different languages, different preferences, which might give some insight into incrementality vs other explanations by looking at how people operate within the bounds of grammer (or are willing to violate rules). Tom Wasow mentioned Richard Futtrell's undergrad thesis on German vs English, where the finding is that there's a lot less overspecification in German, which has grammatical gender, than in English. Explanation: prenominal overspecification (eg "cute small puppy") serves as probabilistic cue to referent, just as grammatical gender is in German.

Koolen et al 2011 (like so many others) frames the issue as one of how to decide which attributes to include in a target description, rather than talking about which words to use, even though one property can be referred to in many different ways ("red", "reddish", "bright", "colored"). Does anyone look at the way that the number of an object's features (if we even want to commit to such a number existing...) interacts with the number of utterance alternatives for referring to that feature (if we want to commit to such a number existing...)?

So we know that people like to overspecify color but not size/shape. What if there's a contextual QUD that really favors information about size/shape but doesn't care about color? Is there any research on this?

Gatt et al's notion of 'utility'. There's not sth inherently evil to utility; in fact, you can build preferences into the utility function. All you can say is that the utility function is not yet complex enough to capture people's actual preferences yet.

Can we get preferences with costs of utterances alone? Or with prior salience of properties? Generally, those guys collapse utterance and property concept. This doesn't seem right, because you could in preinciple use different utterances to refer to the same property, so there must be at least two separate things going on that in these simple cases maybe collapse into one.

There's something unsatisfying about just saying there are preferences. Where do the preferences come from? And can you shift the preferences around contextually, either with the QUD or with visual properties of the scene (eg by making the differences in color much smaller and the differences in size much larger)? That seems to me to be the more interesting thing.

Here's an interesting issue:

• on the one hand, people generate inferences from modificational material in order to avoid thinking the speaker overspecifies (contrastive inferences used in a lot of Tanenhaus offspring work, eg Sedivy, Heller), especially with size adjectives

• on the other hand, there doesn't seem to be a cost to "actual" overspecification (in general....)

• Can we think about this from a noisy channel perspective? In essence, "actual" overspecification is redundant signal. Consider two different scenarios.

  1. We know that either for the speaker or for us, it's plausibly hard to produce/interpret the referential expression right (ie when there's either difficulty, incrementality-related; or there's eg a complex visual scene or maybe other objects that I might misinterprets as being reasonable referents for the expression produced).

     --> in this case, we should be happy with redundant signal, because we know the signal serves no "meaning" purpose but just helps us buy time/ensure communicative success in one way or another. This predicts that in this case we shouldn't get any inference (or cognitive cost if inference not warranted) from overspecification.

  2. We know that there's an easy way to refer to things; properties that one would mention are very salient in common ground, the space of alternatives is small.

     --> in this case, extra signal would be surprising and we should try to draw inferences from it, and if there are no plausible inferences, we should be confused.

==> This is basically like the way Flo and I (and Roger and his student Mark Myslin) have been thinking about things: extra signal can be produced either because it helps with processing, or because you want to convey some meaning. Ie plausibility of processing difficulty as explaining away redundancy.

What findings does this go well with?

• Julie Sedivy's 2003 findings:

  1. In production, people tend to be more over-informative in terms of producing a color term when the color is less predictable from the category denoted by the noun than if it is more predictable. So, people say "yellow cup" a lot even when there is only one cup, but they don't say "yellow banana" when there is only one banana.

  2. In comprehension, color modifiers lead to contrastive inferences when the color is predictable, but not when it's not. Ie there's early convergence on the yellow banana but not on the yellow cup.

• more overspecification when talking to a real speaker (van der Wege 2009)

• lower overspecification rates in simpler visual arrays (Davies & Katsos, 2009, 2013)

• more overspecification as number of distractors increases (Pechmann 1989, Koolen et al 2011)

• Does it fit in with Engelhardt et al 2006 findings? Their design: 4 sentence types*2 referent confexts (either one or two referents): (1) Put the apple on the towel. (2) Put the apple in the box. (3) Put the apple on the towel on the other towel. (4) Put the apple on the towel in the box.

  1. Production -- yes: People produce modifier in two-referent conditions all the time (good), but even in one-referent conditions 1/3 of the time (and more for towel than box, but this is probably because of the non-uniqueness of "towel")

  2. Listener ratings of utterances in different contexts -- neutral: modifier better than no-modifier in two-referent condition, but maybe slightly worse in one-referent condition, but barely

  3. Eye-tracking in only one-referent contexts -- these results are pretty opaque: they're a little slower to look to the target location with the modifier than without -- it would have been interesting to see the breakup by modifier, seeing as how they preferred the modifier when both origin and target location were the same (eg both towels). Our prediction in this case: cost goes away.

• generally:

  1. SIZE > COLOR > PROVENANCE
  2. QUALITY > SHAPE
  3. SIZE > SHAPE
  4. POSSESSIVE > SPEAKER-ORIENTED > SUBJECT-ORIENTED > MANNER/THEMATIC (Cinque 1994)
  5. VALUE > DIMENSION > PHYSICAL PROPERTY > SPEED > HUMAN PROPENSITY > AGE > COLOR (Dixon, 1982)

• Mandarin allows both SIZE COLOR and COLOR SIZE, but only if both modifiers are "-de"-marked, otherwise only SIZE COLOR

• propose two types of modification:

  1. direct (subtypes: hierarchical & parallel)
  2. indirect (via coindexation), based on manner of theta-role assignment (this bit is a little dubious)

• according to them, indirect modification plays out in Mandarin as relative clauses, in Arabic as adjunction

• their primary claim: "Restrictions on the ordering of multiple adjectival modifiers --- henceforth AOR --- obtain iff the adjectives involved are hierarchical direct modifiers."

• what matters: linear ordering or head proximity? they say head proximity.

• semantic factors: absoluteness as major determinant of ordering pref (color, shape as examples --> but: these too need not be absolute; is there already evidence out there or can we show this?)

• missing piece of the puzzle for them: why should AOR be confined to cases of direct hierarchical modification cross-linguistically?

• Claim: violation of preferences needs special comma intonation, then can be interpreted as parallel (ie independent modification of head noun). (1a) She loves all those Oriental, orange, wonderful ivories. (1b) *She loves all those Oriental orange wonderful ivories. (2a) She loves all those wonderful, orange, Oriental, ivories. (2b) She loves all those wonderful orange Oriental ivories.

• they dismiss the possibility that determiners are modifiers, but it's not clear that one should!! eg German. And that the possessives and other non-canonical definite determiners in the partitive "some" database follow the modified rather than the unmodified head pattern

• Cross-linguistic generalizations:

  1. Direct modification & associated cluster of properties (ie AOR following universal pattern): Dutch, Greek, Kannada, Mokilese (in Mokilese, pattern in reverse because it has only post-nominal modification)
  2. Indirect modification & associated cluster of properties (ie no AOR): Japanese, Thai, Arabic
  3. Parallel modiifcation: French -- no AOR postnominally, but only few adjectives allowed to occur pre-nominally

• in English: reordering adjectives that differ in absoluteness is a lot worse than reordering adjectives that don't differ in absoluteness

• In a sequence of adjectives, those denoting non-inherent qualities precede inherent adjectives (cf. Whorf 1945; Biber et al. 1999:599)

• Adjectives which are less dependent on comparison are put nearer to the head noun (cf. Martin 1969b; Posner 1986)

• Quirk et al. (1985: 1341) argue for a subjective-objective gradience determining AO such that “modifiers relating to properties which are (relatively) inherent in the head of the noun phrase . . . will tend to be placed nearer to the head and be preceded by modifiers concerned with what is relatively a matter of opinion.” --> see also Hetzron’s (1978) subjectivity-objectivity gradience and hypothesized order of adjectives, 13 classes

• semantically congruent adjectives ("poor wretched child" vs "poor happy child") have (i) significantly more stable ordering preferences in attributive as well as predicative contexts, (ii) significantly subjectively stronger prenominal ordering preferences, and (iii) are judged significantly more acceptable (cf. Richards 1977:494–499). With semantically incongruent adjectives, the factor of affective loading comes into play: Richards argues the ordering of incongruent adjectives is governed by a ‘first the good news, then the bad news’ principle

• Lockhart and Martin (1969) were able to demonstrate that those adjectives that tend to stand closest to a noun are the ones which are remembered most easily upon the occurrence of the noun. --> semantic association! "Local proximity to the head noun indicates greater noun-specific frequency." (Posner 1986)

• Finally, the general frequency of the adjective has been claimed to influence its position in a string (cf., e.g., Bock 1982) such that more frequent adjectives precede less frequent ones.

• goal: push an information-theoretic account of the function-form relationship in German noun class (gender)

• usefulness of gender: it gives you an early cue to noun form, semantics, and NP referent (if applicable); that is, it reduces the noun's entropy along multiple dimensions

• But how does that relate to our finding that things typically get less noisy as you approach the noun?

• Reflexes of efficient noun class system in German:

  • frequently co-occurring (and thus a priori more confusable) nouns are more likely than chance to be in different genders

Note: this is now outdated, but might be re-integrated at some later point.

Played around with different shapes of priors and utterance costs and values to communicate. For all combinations (3675, see overinformativeness.md) ran model and created some visualizations of the most interesting effects.

Assumptions

• four utterances: silence, color, size, color-and-size. The semantics of color and size is: 'true' if actual color/size above some threshold (to be inferred). The semantics of color-and-size is the conjunction of the individual meanings. Semantics of silence is 'true'.

• utterance costs:

  • uniform: same cost for each utterance

  • prefer-fewer-words: silence 10 times as likely as color/size, which in turn are ten times as likely as color-and-size

  • prefer-fewer-words-and-color: silence 2.5 times as likely as color, which is 4 times as likely as size, which is ten times as likely as color-and-size

• both size and color are expressed on the same discretized scale with 5 values

• each object to be communicated has two values, one for color and one for size (gives 25 potential objects)

• there are seven possible priors on each dimension (color, size), giving 49 different combinations of priors: flat, super-left-peak, left-peak, mid-peak, super-mid-peak, right-peak, super-right-peak

• The model is an S2 speaker that chooses an utterance from among the above 4, given an object (ie two values from the scale) and knowledge about that object type's color and size prior. The speaker reasons (mh-query) about a pragmatic L1 listener who infers the object's color and size. The pragmatic listener reasons (enumeration-query) about what an S1 speaker would have said who knows about the object's color and size value, the object types' color and size priors, and the thresholds for using 'color' and 'size'. The S1 tries to be informative (enumeration-query) to a literal listener who knows the thresholds, object type priors, and utterance and infers (enumeration-query) object values.

Speaker results --- utterance probabilities for objects that match in size and color value as well as in object type size and color prior.

• Most basic result: as object value increases, silence becomes less likely -- this is just in the nature of the threshold semantics.

• What is said instead of silence dpends on both costs and priors.

  • when costs are uniform, color-and-size is almost always the preferred alternative -- this is of course heavily dependent on the fact that we're looking here only at objects that are identical in value on both dimensions

  • when there's a penalty for each added word, color-and-size is only really good for communicating high-value objects when the priors are very left-peaked, otherwise color-and-size is never used. Color and size individually are sometimes used, but mostly the word penalty gives silence a huge boost.

  • when there's a penalty for added words, but less for color than for size, color-and-size is used even less (and again only for communicating high-value objects with very left peaked priors). Silence is now a somewhat less viable option than when every word counts. Color is increasingly used to communicate high value objects, especially when the prior is left or mid peaked.

Get rid of the threshold semantics and instead model features like size and color as binary things. Eg assume four utterances: "blue", "yellow", "big", "small", and combinations thereof. The semantics of each is just true if property holds, false otherwise. Create contexts as worlds consisting of different numbers of objects with different properties. (Make blue/yellow and big/small mutually exclusive.) Then do just standard RSA with varying costs (uniform, penalty for each additional word, penalty for each additional word unless that word is color, smaller penalty for adding color). Vary priors on features.

We use the model in 1_basic_overinformativeness/overinformativeness.church to generate predictions for the situation in which there is a big red object, a small red object, and a small yellow object. This is formally equivalent to the situations described in Gatt, van Gompel, Krahmer, and van Deemter (2011, CogSci) -- see the figure. Model predictions for optimal parameters (eye-balling) are shown alongside the empirical data for both Dutch and English from the paper. Graphs with a fuller exploration of the parameter space are in /graphs.

The results look really nice and fall out purely from an asymmetry in how noisy the predicates are.

The model has the following free parameters: color cost, size cost, color fidelity, size fidelity, speaker optimality.

Important things of note:

1. NO COST ASYMMETRY NECESSARY (in fact, cost asymmetries make results worse, except in certain cases when size costs .2 and color .1)

2. very high speaker optimality necessary -- do we care about this?

3. asymmetry in fidelity necessary -- ie, color fidelity needs to be higher than size fidelity. This is awesome and as we hoped.

We use the model in 2_number_of_distractors/overinformativeness.church to generate predictions for the situation in which there is a big red object, a small red object, and a varying number of additional small (either red or yellow) objects (where objects don't differ in type). The model predicts an increase in redundant color use with increasing number of distractors when the color of the distractors is different from the color of the target object, but not when it's the same. A similar effect is not predicted for redundant size use. Utterance probabilities are shown for speaker optimalities of 5, 10, and 15. Beyond 4 distractors, there is no additional effect of number of distractors (because overmodification is already at ceiling).

Note that when distractortype is "varied", that means that there is one distractor that shares the non-distinguishing feature with the target (eg, when the target is big and red, and size is sufficient to distinguish, this distractor is small and red) and all other distractors share neither feature with the target (ie in our example, all other distractors are small and yellow).

Beta stage. So far the model has been extended to include type (eg "chair"), and type is a necessary part of the utterance. There are now two new parameters: type fidelity and type cost. (Type cost is actually irrelevant since mentioning type is mandatory.) The model gets the low vs high variation asymmetry -- ie more (redundant) color mention in high variation than low variation contexts for the first two of the following contexts taken from Koolen et al 2013. These correspond to the target conditions in their Exps. 1 and 2. I haven't yet tried modelng the third context, i.e. the one that requires inclusion of an orientation feature.

Exp. 1. Left: low variation. Right: high variation.

Exp. 2. Left: low variation. Right: high variation.

Exp. 3. Left: low variation. Right: high variation.

Free parameters of the model are set as follows (based on earlier results reported above): size cost .1, color cost .1, size fidelity .8, color fidelity .999.

Important: type fidelity has to be lower than color fidelity to get the effect. If it isn't, the high variation condition behaves like the low variation condition (see the plot below, in particular the qualitative jump in the high variation conditions between when type fidelity is lower than color fidelity vs the last row where it's equal). The asymmetry is too great (empirically the overmodification values are around .25 in the high variation and .05 in the low variation condition), but it's there. Note that orientation is not yet included in the model, so the context from Exp. 3 above can't yet be captured, and who knows what leaving out orientation does to the results of the contexts from Exps. 1 and 2. In addition, the alternative "one" is not included.

This plot shows utterance probabilities for the low and high variation conditions in Exps. 1 and 2 above (only for those utterances that have a probability greater than .05 under some parameter setting). For the remaining utterances, values are only plotted if utterance probability is greater than .0001. Rows indicate different type fidelities. Columns indicate the different experimental conditions, which map directly onto the pictures above. Colors indicate different speaker optimalities.

Including "one" as an alternative to mentioning type (and which is in the extension of every object) retains the overinformativeness effect, though the probability of producing "one" may well be too high (see plot below). Interestingly, "one" is only produced in the high, but not in the low variation conditions -- presumably, because in the low variation conditions the ratio of type to "one" informativeness is much higher than in the high variation conditions.

Rows indicate different fidelities for "one". Clearly, fidelity doesn't matter. (Why not?) The plot is the result of setting free parameters of the model as follows: size cost .1, color cost .1, type cost .1, size fidelity .8, color fidelity .999, type fidelity .9.

Exp. 1.

Target color typicality	Example context
Typical
Intermediate
Atypical

Empirical results:

Modeling problems:

Where to put the effect?

1. Conditional prior: we at some point put this out there as a potential explanation. But I don't think this can work: the adjectives get their semantics--which is what determines their informativeness relative to other adjectives--on a context-by-context basis. So any notion of prior experience with the adjectives would have to come in somewhere else. Where? confused

2. Brute force: have a noun-conditional adjective cost; i.e., for each adjective, hard-wire its cost (but vary it by noun). ad hoc, unelegant, unwieldy, bad at scaling up, boring. But implemented and works.

3. QUD mechanism: assume that the speaker assumes the listener wants to know not just the identity of the object, but also its color. In that case, the more predictable the color is, the less use there is in mentioning it. Conversely, the less predictable, the more it should be mentioned. how to justify the assumption that the listener wants to know about color? why not other features? or should we assume this is instead a speaker-internal can't-do-otherwise pressure having to do with the salience of the color when it's unpredictable? in which case it should probably go directly into the cost after all...

The effects we want to get are:

• basic asymmetry in redundant use of color vs. size adjectives done
• increasing number of objects in world leads to more redundant color use done
• increasing scene variation in world leads to more redundant color use done
• increasing predictability of color from noun category leads to decrease in redundant color use

The next steps on the road to this are:

Notes from lab meeting, 11/04/2015:

• Modeling: update color predictability model to get it closer to the "noisy truth conditions" one, by having fidelity depend on object type, such that type is less likely to be true for weirdly colored things. I.e., this is like a noisy semantics plus feature statistics model. Practical: update the get-fidelity function such that the flip probability depends on (is?) P(color|type).

• Experiments: Gatt et al 2011 replication (Caroline)

• Experiments/model exploration: there's a lot one can do in the "number of distractors" domain. Probe model predictions for following IV variations:

  1. number of distractors

  2. ratio of same- vs different-color distractors

  3. vary fidelity of color; if it's not at .999 (ie not super-duper-high), are there gradient predictions for overmodification even when all distractors are of the same color, as opposed to the flat lines that are predicted for .999?

• Experiments/model exploration:

  1. Explore more systematically what the color predictability model predicts for a particular color typicality value, if you vary the tpicality of the other contextually available objects of the same color

  2. Can you do typicality experiments with other properties? Texture? Material? What's the relation of the fidelity parameter to subjectivity? To typicality/predictability? To perceptual uncertainty (in listener, in speaker)?

Previous notes:

• Modeling: write to Rud to see if he'll share the full dataset?

• Modeling: think about how to get the color predictability effect (global QUD type of mechanism? conditional prior?)

• Modeling: wait for Albert's dataset

• Get familiar with Robert's paradigm to do production study a la Gatt (need to validate first, ie do we get similar results with written production as the Dutch have been getting in spoken production?)

• Run experiment with adjectives that are between SIZE and COLOR on the ordering preference scale: empirically, are overmodification rates correlated with position on the scale? (eg age, texture, shape, material)

1. Theory paper that focuses on demonstrating that we can get the four qualitative effects on the list above with a basic version of the model. This paper will include no experiments of our own, except maybe one using faultless disagreement to show that for the stimuli in question color is indeed less noisy than size (and less noisy than type, too?).

2. Experimental paper for which we first need to go about piloting a bunch of things in order to find a paradigm within which we both replicate the important qualitative effects that have been found previously, and which we can use as an experimental basis to model the crap out of in one fell swoop.