Transcription

Communicative efficiency in language production and learning:Optional plural markingChigusa [email protected] of Brain and Cognitive Sciences, Meliora HallUniversity of RochesterAbstractRecent work suggests that language production exhibits a biastowards efficient information transmission. Speakers tend toprovide more linguistic signal for meaning elements that aredifficult to recover while reducing contextually inferable (morefrequent, probable, expected) elements. This trade-off hasbeen hypothesized to shape grammatical systems over generations, contributing to cross-linguistic patterns. We put thisidea to an empirical test using miniature artificial languagelearning over variable input. Two experiments were conductedto demonstrate that the inferability of plurality informationinversely predicts the likelihood of overt plural marking, aswould be expected if learners prefer communicatively efficientsystems. The results were obtained even with input frequencycounts of the plural marker counteract the bias, and thus provide strong support for critical role of inferability of meaningin language learning, production, as well as in typologicallyattested variations.Keywords: language production, artificial language learning, optional morphology, plural marking, communicative efficiencyIntroductionProducing language is a balancing act. On the one hand,the speaker is biased towards minimizing effort by choosing a shorter form and linguistic elements that are readily retrieved and formulated (e.g., Ferreira & Dell, 2000; MacDonald, 2013). On the other hand, the speaker’s choices are, atleast to some extent, optimized under considerations of communicative success. The speaker is more likely to encodea linguistic message that is otherwise less predictable or recoverable (e.g., Aylett & Turk, 2004; Levy & Jaeger, 2007;Buz, Tanenhaus, & Jaeger, 2016). For instance, in English,native speakers are more likely to produce the optional complementizer “that” after verbs that are less likely followed bya complement clause (e.g., I read (that) the president was arrested.) compared to those that are biased towards a complement clause continuation (e.g., I thought (that) the presidentwas arrested) (Jaeger, 2010).It has been argued that such speakers’ preferences reflectsa principle of the computational system underlying languageproduction. That is, linguistic communication necessarily involves transmission of information through a noisy channeland the information is often degraded due to factors suchas production/comprehension mistakes, ambient noise, andnoise in the perceptual and neural mechanisms involved inlanguage processing. The comprehender, therefore, has toinfer a message sent by the speaker rather than simply recognizing and decoding the input (e.g., ? (?), see also Piantadosi,Tily, and Gibson (2011)). A trade-off between the amount ofScott [email protected] of Linguistics, Lattimore HallUniversity of Rochesterinformation and the amount of linguistic signal expended isexpected when the speaker encodes a message in a way sothat the listener has a higher chance of inferring it given thelinguistic input and contextually shared knowledge. Put simply, the speaker should preferentially encode components ofmeanings that are otherwise less likely to be inferred by thelistener given prior expectations.An appealing property of this account is that it providesa potential explanation for typological patterns (although thelink is tentative so far) beyond individual instances of sentence production. It has long been observed that the lexicon and grammar of languages across the world tend to exhibit many properties that would be expected if languagewas shaped by communicative pressures (e.g., Zipf (1949);Plotkin and Nowak (2000), also precisely those predictedby accounts of communicatively efficient language production Piantadosi et al. (2011); Jaeger (2013)). Recent workon learning biases during (miniature artificial) language acquisition has also found the same biases to be active during artificial language learning (e.g., Culbertson, Smolensky,& Legendre, 2012; Fedzechkina, Newport, & Jaeger, 2016).Fedzechkina et al. (2012) found that native speakers of American English, when learning a miniature language with anoptional case marking morphology, restructure the input andcondition the uses of the marker on factors such as Animacy.This is in line with patterns observed in existing optional (ormore categorical) case-marking language, suggesting a tightlink between observations in lab-based studies and typological pattern found in real languages e.g., (Aissen, 2003; Kurumada & Jaeger, 2015).We provide a novel investigation of the possible roleof communicative efficiency in grammatical number marking, in particular the acquisition of Optional Plural Marking (OPM). OPM is not uncommon cross-linguistically (e.g.,Yucatec Maya (Butler, Bohnemeyer, & Jaeger, 2017)) andhas been investigated linguistic work on grammatical systems(see Corbett (2000) and Haspelmath (2013) for general discussion). Yet, the mechanisms that predict when speakerswould use (or would not use) the marker are not well understood.Two classes of accounts have been put forward. One relies on form-based frequency of the input. That is, learnersare more likely to hear the optional marker with a particular class of nouns and reproduce the distributional patternsin the their production (e.g., Tiersma, 1982; Haspelmath &Karjus, 2017). This is largely consistent with a general view

in language acquisition research: higher input frequency inthe input often predicts earlier acquisition and higher usagefrequency of the element.A second set of accounts, relying on conceptual “markedness”, make a distinct prediction. These accounts posit thatsingular (plural) values are conceptually consonant with someentity types more than others. For instance, entities that aretypically conceptualized as individuals (e.g., large animals)tend to be referenced in language as singletons, rather thanmultiples. For these entities, their occurrence in multiples islimited, thus resulting in lower plural inferability, and therefore, plural coding is the unexpected or “marked” value. Conversely, entities that are often conceptualized as collectives(e.g., small insects) have high plural inferability.We argue that the account based on conceptual markedness, or meaning-based predictability, accords with the communicative efficiency hypothesis. Put simply, learning andproduction of OPM is guided by a consideration to communicate the plural meaning most efficiently. That is, learnersshould prefer systems in which markedness of plural meaningis inversely correlated with the production of plural marking.Accounts based in communicative efficiency thus predict that,when learners of an OPM language refer to multiples of individualized items (e.g., large animals), they should be morelikely to produce plural marking, compared to when referringto multiples of collective items (e.g., small insects).Preliminary support for the conceptual markedness account comes from repeated observations across a number ofstudies on typologically-diverse languages which possess asingulative/collective morphology (e.g., Arensen (1998) onMurle, Grimm (2012) on Dagaare, Mifsud (1996) on Maltese, Stolz (2001) on Welsh). In these languages, referentsthat are likely to be conceptualized and manipulated as collectives (e.g., fruits, grains, vegetables) or a group/mass ofindividuals tend to be expressed with lexical items that havea plural meaning by default (e.g., psy “peas” in Welsh) andonly through an additional singulative suffix can singletonsbe designated (e.g., psy-en “pea”).While effects of markedness on number-marking morphology have been hypothesized and widely discussed in linguistics, it is particularly difficult to differentiate predictability of forms and predictability (markedness/inferability) ofmeanings in a corpus based method. Haspelmath and Karjus (2017), for instance, collected token counts of singularvs. plural forms of a word (e.g., psy-en and psy) to argue thatfrequency asymmetries can predict the asymmetrical pluralmarking system such that the more frequent meaning (singular/plural) is often encoded in a simpler form. However, as inmost of existing corpus-based approaches, one cannot easilydissociate the frequency of forms and the frequency of meanings. In other words, there is no simple way of measuring theinferability of meanings apart from the frequency of forms.Here we aim to tease apart these two possible accounts using a miniature language learning paradigm. We present twoproduction experiments on optional number-marking. Learn-Figure 1: Sample images of visual stimuli in Experiments 1and 2.ers acquire 12 novel nouns and one novel verb to producesimple intransitive sentences with the Subject-Verb word order. As we describe below, the novel lexicon consists of twoclasses of referents: six Individuals and six Collectives thatdepict fictitious animals and insects, respectively. In the input, they were visually presented as either singletons or multiples at varying rates: Individuals are more likely to be singletons whereas Collectives are more likely to be multiples. Referents are optionally (stochastically) plural-marked and theprobability of occurrence of the marker was constant acrossIndividuals and Collectives. Notice that, given the fact thatCollectives are more likely to appear as multiples, a largerproportion of token counts of Collectives appeared with theplural marker than Individuals.Frequency-based accounts therefore predict that learnersof this miniature language should be more likely to use theplural marker with the Collectives rather than Individuals.On the other hand, conceptual-markedness based accountswould predict the opposite. Individuals are less likely to appear as multiples, which makes the plural meaning less inferable without the overt marking. Therefore, language shouldbe more likely to use the plural marker with the Individualsrather than Collectives.Experiment 1 in the current study directly pits the frequency vs. communicative-efficiency accounts against eachother. Animals (visually and conceptually more individuated)are more likely to be presented as a singular referent. In contrast, insects (visually and conceptually more collective) aremore likely to be presented as multiples. We test whether theinferability of plurality information affects the likelihood ofovert marking (producing a plural marker), as would be expected if learners prefer communicatively efficient systems.Experiment 2 investigates if the inferability of plural meaningis learned through exposure within the current experiment or

Training phase(1) wordexposure24 items(12 characters*2)(2) 4AFCword learning game48 items(12 characters*4)Test phase(3) word production(4) sentencecomprehension(5) sentence production12 items(12 characters*1)48 items(12 characters*4)24 items(12 characters*2)Altogether participants heard the names of "individuals" (larger animals) and "collectives" 60 times each.singular (75%)individualsplural (25%)ka (10 items)collectivessingular (25%)plural (75%)ka (30 items)Figure 2: Schematic illustration of the flow of the experiment and proportions of singular and plural visual prompts.influenced by participants’ prior experiences.Experiment 1We employ a miniature artificial language learning paradigmmodifying Fedzechkina, Jaeger, and Newport (2012). Learners first learn 12 nouns and then learn to produce intransitivesentences in response to prompt video clips. We manipulatedvisual features of the referents (e.g., size, group size, movements) as well as the probability with which Individuals (animals) and Collectives (insects) appear as singletons and multiples, respectively. If optional number-marking is affected bya preference for communicative efficiency, speakers shouldbe more likely to produce responses with a plural-marker forIndividual (animal) compared to Collective (insect) referents.MethodsParticipants 40 native speakers of American English atUniversity of Rochester participated in this study. They received 10 for their participation.The languageLexicon We constructed twelve nonce nouns. Six of themdenote large animal characters and the other six denote smallinsect characters (e.g., Fig.1). To ensure that results did notinclude spurious phonological effects, we created two versions of of character-noun combinations. All of the nounswere 1-2 syllables following the English phonotactics (e.g,norg, velmick, zamper). When characters were presented asmultiples, the noun was optionally suffixed with the pluralmarker (-ka) that optionally marked 2/3 of the time.We included only one verb – glim – meaning “moving upand down”. In a sentence, the verb followed a noun, constituting a SV (intransitive) word order (e.g., Velmick-ka glim).ProcedureThere were five phases in this experiment (Fig. 2). Participants went through (1) - (3) for six of the twelve noun types(three animals and three insects) and then repeated the sameprocedure to learn the other six words.(1) Word exposure (12 characters * 2 24 trials total):During word exposure participants were presented with pictures of each of the characters. Participants were instructedto repeat the names of the characters aloud. In this phase, allthe characters were presented as singletons. An animal wasdepicted approximately three times as large as an insect.(2) Word learning game (12 characters * 4 48 trials total): The initial word presentation was followed by aword learning phase where participants were presented withfour pictures (4 Alternative-Forced-Choice task) and asked tochoose the correct match for the noun provided (48 trials total). Feedback was provided after each trial. In this phase,Individuals and Collectives were presented as singletons andmultiples at different rates. Individuals occurred 75% of thetime as a singleton (i.e., one animal, Fig. 1a), and 25% asmultiples (Fig. 1b). Collectives had the inverse distribution(25% singleton, 75% multiples). Both Individual (animal)nouns and Collective (insect) nouns were followed by theplural-marker (ka) 2/3 of the time when occurring as multiples.(3) Word production (12 characters * 1 12 trials total): Participants were shown 12 characters (singleton) oneby one and asked to name each of them.(4) Sentence comprehension (12 characters * 4 48 trials total): During the sentence comprehension phase, participants viewed short clips and heard their descriptions in thenovel language. Participants were asked to repeat the sentences out loud. As in the word learning phase, Individuals and Collectives occurred as singletons 75% and 25% ofthe time, respectively, and they were followed by the pluralmarker (ka) 2/3 of the time when occurring as multiples. Consequently, participants heard the animal and insect nouns withka 10 times and 30 times, respectively by the end of this phase(Fig. 2). Critically, this means that input frequency biases

against the prediction of communicative-efficiency: the inputin our experiment(s) provides more instances of training forplural-marked Collectives than Individuals.(5) Sentence production (12 characters * 2 24 trialstotal): In the final test (sentence production) phase, participants saw silent videos of singletons and multiples and hadto produce intransitive descriptions. In this phase, visual images for the multiples had three instances of the charactersboth for animals and insects. This was done to ensure thatparticipants use -ka to signal plurality rather than the particular number of instances (two for animals and ten for insects)seen in the exposure input.ScoringIn the 4AFC comprehension test, participants’ responseswere scored as ‘correct’ if they matched the intended referent.Following the standard in similar studies (e.g., Fedzechkinaet al. (2012)), we a priori decided to exclude participants whofailed to achieve mean accuracy of 65% from all analyses.We transcribed the production obtained in (5) and annotated if participants produced a given noun correctly and if anoun was produced with ka or not. In the comprehension test,participants responses were scored as “correct” if it matchedthe provided input while subtle phonological variations (e.g.,velmick pronounced as belmick) were ignored.Results and DiscussionComprehension Accuracy To ensure that participants haveachieved a sufficient level of accuracy in identifying referents, we first measured their performance in the 4AFC wordlearning game. The average rate of correct response was 74%and all the subject means were above the pre-determined cutoff rate of 65%. The mean accuracy of the word productionphase (3) was above 85%. This suggests that the task wasfeasible and the lexicon was acquired reasonably well beforeparticipants performed the production task.Plural marker use in Production We excluded five(12.5%) of the participants who failed to produce 50% ofthe sentences in the final sentence production phase. Thiswas done to ensure that the data analyzed are produced bythose who have mastered the language at a more or less sufficient level. All the results we report below remain unchanged,however, when we include all the participants. We then further removed 105 (14.5%) sentences that included wrongnouns such as a different character’s name or a noun that didnot belong to the learned lexicon. The final dataset included35 subjects and 619 sentences.Proportions of participants’ plural marker use in Experiment 1 are illustrated in Figure 3. To analyzed the data, weused a mixed effect logit model in R, predicting the use ofthe optional plural marker. We included the noun classes(Individuals (animals) vs. Collectives (insects)) and visualprompts (singleton vs. multiples) as fixed effects and participants and items as random effects. The model includedthe maximal random effects structure justified by the databased on model comparison (Jaeger, 2008). There was anexpected significant main effect of visual prompts such thatparticipants were more likely to produce the optional pluralmarker ka for multiples (p .001). Critically, the interaction between the noun class and the visual prompts was alsosignificant (p .03): Learners (inversely) conditioned pluralproduction on plural inferability. They did so despite the factthat they were exposed to three times as many instances of-ka with the Collectives (insects) compared to the Individuals(animals).Experiment 2What is deriving the observed difference between Individualsand Collectives? Under our hypothesis, it is at least partiallydue to the expectation that animals are less likely to be represented with the plural meaning, and hence the meaning isless inferable. (And the opposite is the case for insects). InExperiment 1, however, it is not clear if the inferability of theplural meaning (the conditional probability of multiples giventhe referent) is learned within the experiment or it is carriedover from participants’ prior semantic knowledge that insectsare more likely to occur, and be referred to, as multiples.To separate these two factors, in Experiment 2, we used thelexical items from Experiment 1 while associating them withnovel, semantically bleached, items to remove effects of priorsemantic knowledge. If participants exhibit the same asymmetric use of the plural marker for Individuals and Collectives, that will yield support for the idea that the inferabilityis likely extrapolated in this experiment.Participants20 native speakers of American English at University ofRochester participated in this study. They received 10 fortheir participation.The languageThe lexicon was identical to that used in Experiment 1. Theonly difference is that the visual images consisted of 12 geometrical shapes with no commonly known names. To equatethe visual features of the referents (e.g., size, spacial distributions, complexity of visual scenes), we created two classes ofreferents (Fig. 1). One of the classes (Individuals) consists ofsix relatively large geometrical shapes spatially distributed ina manner similar to how the animals were presented in Experiment 1. The other class (Collectives) consists of six smallershapes that replace the insects in Experiment 1.ProcedureThe same as Experiment 1.Results and discussionComprehension Accuracy The mean accuracy in the 4AFCtask was 68%, suggesting that the word learning was slightlymore difficult in Experiment 2 compared to Experiment 1,presumably due to the overall unfamiliarity with the geometrical shapes. One subject could not achieve the cut off rate of

Figure 3: Proportions of plural marker use by conditions.Dots present by-participant averages. Error-bars show 95%Confidence Intervals.Figure 4: Proportions of plural marker use by conditions.Dots present by-participant averages. Error-bars show 95%Confidence Intervals.65% and was removed from the analysis. The mean accuracyin the word production phase (3) was 80%.Plural marker use in Production We excluded three(15.7%) of the participants who failed to produce 50% ofthe sentences in the final sentence production phase. As inExperiment 1, all the results we report below remain unchanged with the complete set of data. We then further removed 99 (25.9%) sentences that included wrong nouns. Thefinal dataset included 16 subjects and 283 sentences.Proportions of participants’ plural marker use in Experiment 2 are illustrated in Figure 4. The same model from Experiment 1 was used to predict participants’ use of -ka. Whilethe main effect of visual prompt was significant (p .001),the interaction between the noun classes and visual promptwas not (p .2). Learners were equally likely to produce theoptional plural marker for both Individuals and Collectives,suggesting that the effect in Experiment 1 was likely drivenby prior semantic knowledge of the semantic classes (animalsvs. insects). The relative conditional probability of multiplesin the input was not sufficient to induce this effect, perhapsrequiring longer exposure (to be tested).guage. Critically, English does not have the optional pluralmarking (OPM) system. Still, when native speakers of English are exposed to an OPM language with no bias to markplurality for low-inferability items, they end up producingmore plural marking for less inferable items. As such this isone of the first studies showing a systematic effect of semantic knowledge on the morpho-syntactic encoding of speech.This body of research including the current study constitutes strong support for the view that language production isoptimized to maximize the efficiency of information transmission (Levy & Jaeger, 2007, Jaeger, 2010). The asymmetrical uses (and non-uses) of -ka cannot be accounted forin terms of availability of an upcoming linguistic element orother sources of speaker-internal production or planning difficulties (Ferreira & Dell, 2000; MacDonald, 2013). All thesentences were produced with the same verb and no participant failed to learn to produce it.It is possible, however, that the difference between Experiment 1 and Experiment 2 stems from the differential levels of mastery in word learning. Participants learned thenonce labels better in Experiment 1 than in Experiment 2,presumably due to more easily memorable visual referents(animals/insects as opposed to geometrical shapes). The facilitation in word learning might have made it easier for participants to modulate their production with respect to communicative considerations. In a future study, we intend toincrease the amount of of 1) linguistic and 2) non-linguistic(inferability) information in the input in Experiment 2 to seeif participants would show an asymetrical use of the OPM.Lastly, this study has broad implications for understanding typologically attested morpho-syntactic variations. It haslong been hypothesized that conceptual markedness plays aguiding role in grammaticalization of morpho-syntactic elements. The current experimental paradigm using an artificiallanguage allows us to dissociate the effects of input in termsGeneral DiscussionOur results suggest that native speakers of American Englishprefer to produce an NP without overt marking of pluralitywhen the meaning is more inferable given the semantics ofthe noun classes (e.g., animals vs. insects). The effect wasnot present when the nonce shapes were used even thoughwithin-experiment statistics as well as visual features of referents (size, spacial arrangements, movement patterns) wereheld constant. This suggests that learners have knowledge ofthe relative inferability of plural meaning for different typesof referents (e.g., How often do you describe animals/insectsas singletons vs. plural referents?), and this knowledge supports the learning of morphological systems of a novel lan-

of the predictability of forms (e.g., How often do you hear aparticular noun with -ka?) and the predictability/inferabilityof meaning (e.g., How likely is it that a given referent is described as a singleton vs. multiples?), making it possible totest a multitude of hypotheses put forward about effects ofmeaning-based predictability. For instance, it has been observed that functionally paired objects (e.g., glasses, chopsticks, a set of pillars) and body-parts (e.g., eyes, ears, hands)are often conceptualized as plural by default, and hence likelyencoded without any additional plural marking morphology(Haspelmath & Karjus, 2017). We can directly test this hypothesis in the current paradigm using objects that differ intheir likelihood of appearing in pairs.In summary, the current results yield support for the hypothesis that the inferability of plurality information guideslearners to restructure the input they receive, as would be expected if language users are biased towards communicatelyefficient systems. Our results thus illuminate the critical roleof distributional information of meanings on language learning, production and typological variation across languages.AcknowledgmentsThanks to Wesley Orth, Joseph Plvan-Franke, Sahed Martinez, Kathleen Ward, and Jennifer Andrews for assistancewith stimuli construction, data collection, and annotation,and to T. Florian Jaeger, Masha Fedzechkina, KurumadaTanenhaus Lab, and Experimental Semantics/Pragmaticsgroup for valuable discussion.ReferencesAissen, J. (2003). Differential object marking: Iconicity vs.economy. Natural Language and Linguistic Theory, 21,435–483.Arensen, J. E. (1998). Murle categorization. In G. J. Dimmendaal & M. Last (Eds.), Surmic languages and cultures(pp. 281–318). Rüdiger Köppe.Aylett, M. P., & Turk, A. (2004). The smooth signal redundancy hypothesis. Language and Speech, 47(1), 31–56.Butler, L. K., Bohnemeyer, J. B., & Jaeger, T. F. (2017).Plural marking in Yucatec Maya at the syntax-processinginterface. In A. Machicao y Priemer, A. Nolda, & A. Sioupi(Eds.), Zwischen Kern und Peripherie (Studia Grammatica,volume 75). Berlin: Akademie-Verlag.Buz, E., Tanenhaus, M. K., & Jaeger, T. F. (2016). Dynamically adapted context-specific hyper-articulation: Feedbackfrom interlocutors affects speakers’ subsequent pronunciations. Journal of Memory and Language, 89, 68–86.Corbett, G. G. (2000). Number. Cambridge: CambridgeUniversity Press.Culbertson, J., Smolensky, P., & Legendre, G. (2012). Learning biases predict a word order universal. Cognition,122(3), 306–329.Fedzechkina, M., Jaeger, T. F., & Newport, E. L. (2012). Language learners restructure their input to facilitate efficientcommunication. PNAS, 44(109), 17897–17902.Fedzechkina, M., Newport, E. L., & Jaeger, T. F. (2016). Balancing effort and information transmission during languageacquisition : Evidence from word order and case marking.Cognitive Science, 1–31.Ferreira, V. S., & Dell, G. S. (2000). The effect of ambiguityand lexical availability on syntactic and lexical production.Cognitive Psychology, 40, 296–340.Grimm, S. (2012). Individuation and inverse number markingin Dagaare. In D. Massam (Ed.), Count and mass acrosslanguages. Oxford: Oxford University Press.Haspelmath, M. (2013). Occurrence of nominal plurality. InM. S. Dryer & M. Haspelmath (Eds.), The world atlas oflanguage structures online. Leipzig: Max Planck Institutefor Evolutionary Anthropology.Haspelmath, M., & Karjus, A. (2017). Explaining asymmetries in number marking: Singulatives, pluratives and usagefrequency. Linguistics.Jaeger, T. F. (2008). Categorical data analysis: Away fromanovas (transformation or not) and towards logit mixedmodels. JML, 59, 434–446.Jaeger, T. F. (2010). Redundancy and reduction: Speakersmanage syntactic information density. Cognitive Psychology, 61, 23-62.Jaeger, T. F. (2013). Production preferences cannot be understood without reference to communication. Frontiers inPsychology, 4, 1–4.Kurumada, C., & Jaeger, T. (2015). Communicative efficiency in language production: Optional case-marking inJapanese. Journal of Memory and Language, 83.Levy, R., & Jaeger, T. F. (2007). Speakers optimize information density through syntactic reduction. In B. Schlökopf,J. Platt, & T. Hoffman (Eds.), Advances in Neural Information Processing Systems (NIPS) (Vol. 19, p. 849-856).Cambridge, MA: MIT Press.MacDonald, M. (2013). How language production shapeslanguage form and comprehension. Frontiers in Psychology, 4, 226.Mifsud, M. (1996). The collective in Maltese. Rivista diLinguistica, 8(1), 29-51.Piantadosi, S. T., Tily, H., & Gibson, E. (2011). The communicative function of ambiguity in language. Cognition,122, 280–291.Plotkin, J., & Nowak, M. (2000). Language evolution andinformation theory. Journal of Theoretical Biology, 205(1),147–159.Stolz, T. (2001). Singulative-collective: Natural morphology and stable cases in Welsh number inflexion of nouns.Sprachtypologie und Universalienforschung, 54(1), 52–76.Tiersma, P. (1982). Local and general markedness. Language, 58, 832–849.Zipf, G. K. (1949). Human behavi

ing, optional morphology, plural marking, communicative ef-ficiency Introduction Producing language is a balancing act. On the one hand, the speaker is biased towards minimizing effort by choos-ing a shorter form and linguistic elements that are readily re-trieved and formulated (e.g., Ferre