Reconciling inconsistency in encoded morphological distinctions in an artificial language, in Proceedings of the 36th Annual Conference of the Cognitive Science Society. The visuals separate it from training. In Experiment V, by contrast, each query picture displays an item and each target picture displays three of the same item. We investigate the learning of contextual meaning by adults in an artificial language. Both the patterns and the contextual differences are more abstract than those explored in studies of phonetic adaptation such as Kraljic and Samuel (2006). doi: 10.1177/0956797615597912, Milroy, J., and Milroy, L. (1993). The former is typical of most cases of social-indexical variation, in which we know our conversation partner's principal characteristics. Increasing the amount of evidence available by lengthening the training phase enables more participants to modify their strategy and succeed at the task. Work on language variation and change provides an important body of evidence on how people learn linguistic patterns that are associated with non-linguistic contexts. We had three hypotheses for Experiment II: (i) Participants would learn the diminutive pattern and extend it to new items in the gender condition (ii) learning and extension would be poorer in the view condition (iii) participants would be more likely to assign the correct pattern to items in the test phase if they have seen them in the training phase. Lab. Our paradigm demonstrates differences in adult learning of socially salient vs. accidental non-linguistic contextual cues.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Training takes longer (in terms of trial counts) in Experiment II (m = 66, sd = 25) than in Experiment I (m = 42, sd = 18) (a significant difference according to a Wilcoxon rank sum test, W = 3450, p < 0.001). doi: 10.1177/0165025412454028, Van der Zande, P., Jesse, A., and Cutler, A. They are more likely to rely on the social contextual cue if it is salient. 116, 19051908. Our results also show that many participants have difficulties learning a vocalic cue to an alternation. Language learners restructure their input to facilitate efficient communication. Beckner et al. The hero will meet people who stand on the roofs and ask questions. Of particular interest in this paper is the degree to which learners may attend differently to different types of non-linguistic context. They are able to learn this cue with adults and extend it to children and vice versa. We will use the term social salience to compare the usefulness of a context for linguistic learning andas a consequencepeople's ability to rely on it. The main effects did not change considerably. Rev. Participants are also told when they enter the test phase. J. Lang. Cognition 107, 5481. We also expected that they would generalize to new items. Yu and Smith, 2007). In the training phase, we look at the number of trials it takes a participant to finish the experiment. The gender condition of Experiment IV is more interesting. Distributions of participant responses on test items in the view and gender conditions, Experiment III. Campbell-Kibler, K. (2011). Through a series of short games, played online, we test how well adults can learn different contextual meanings for a word-formation pattern in an artificial language. Language and Social Networks. B. This difference does not vary with participant age or gender. doi: 10.1353/lan.2011.0016, Beckner, C., Rcz, P., Brandstetter, J., Hay, J., and Bartneck, C. (2016). Linguist. Where a combined model was too complex we fit interactions of participant-level predictors (age and gender) and experimental conditions (cue type, item presence in training, etc.) The effect plot can be seen in Figure 14. U.S.A. 109, 1789717902. In Experiment IV, we see that participants who focus on the suffix in test also finish training faster. Comparing type counts: the case of women, men and -ity in early English letters, in Corpus Linguistics: Refinements and Reassessments, eds A. Renouf and A. Kehoe (Amsterdam: Rodopi), 87109. Groningen: Noord-Hollandsche Uitgevers Maatschappij. In short, participant accuracy is highly variable in learning tasks and this variation derives from a complex set of cognitive differences. (2008) extend this generalization to a new accent: their participants are able to use a context-specific vowel adaptation mechanism to process phonetic variation coming from a speaker and then, in turn, re-use the adapted vowel categories when encountering a similar speaker. Third, listeners do not treat all available information the same way. They are also able to generalize this pattern to items not seen during training.

In contrast, a sizeable proportion of participants has high accuracy for the gender cue.

Each experiment has items in the test phase that were seen in training as well as new items.

Participants who go through longer training are more accurate in the test phase. The joint analysis gives us no ground to reject the null hypothesis that the plural does not differ from the diminutive. Learning the view cue requires participants to notice that changes in language use are correlated with changes in the direction the partner is facing. There are two suffixes, and the correct one includes a vowel that matches the stem vowel. This means that each of them had to make at least a few mistakes before they learned the pattern. Experiment III expands the scope of contextual learning to examine whether participants generalize on the basis of conversation partner gender. Experiment IV. (2005). We hypothesized that participants would fare better at learning the link between the type of conversation partner and morphological pattern if the categorization of conversation partners was socially salient. Sidestepping the combinatorial explosion: an explanation of n-gram frequency effects based on naive discriminative learning. 37, 1320. Our method removes the slowest 2.5%. The summary for the best model can be seen in Table 4. Linguistics 48, 865892. We fit a mixed-effects binomial regression model on the test data using item and conversation partner presence in training, type of cue, and participant age and gender as predictors, with a participant random intercept. We test for all interactions of our terms. Mean age is 35 years, with a standard deviation of 10.98. Foulkes (2010, p. 6) laments the lack of understanding of learning and storage of social meaning, stating that: it now seems uncontroversial to conclude that social information is retained in memory alongside linguistic knowledge. Some higher level knowledge is needed to realize that an adult and a child share the same gender. We then compare the test data from Experiment VI to test data from Experiment II. J. Sociolinguist. We created our plots using ggplot (Wickham, 2009). For each participant, two syllables are assigned as suffixes. Mean accuracy is higher for items seen in training (seen items, right) than for items participants only encountered in test (unseen items, left). PLoS ONE 8:e57410. Experiment I tests for a main effect of a single across-subject condition while Experiment II has two across-subject conditions and tests for interactions as well. The morphological pattern is recognized and extended to previously unseen words after training. As in the survey of statistical learning in various domains by Siegelman and Frost (2015), we found that people vary in their individual ability to learn from training datasome people have high accuracy, and others perform worse. ggplot2: Elegant Graphics for Data Analysis. doi: 10.1080/09658416.1996.9959890. We also assumed that this salient link would be generalized to new items and new conversation partners. Event-related potentials for the older participants indicate better ability to focus attention on feedback. This work shows that visual decision tasks speed up if the trial (with a given visual context) is shown repeatedly, despite the fact that participants are unable to identify the contexts afterwards, suggesting that the effect of the context on learning can be implicit. Dixon, R. M. W. (1980). The quadratic-chi histogram distance family, in European Conference on Computer Vision (New York, NY: Springer), 749762. Another important question arising from our results across experiments I-V is the role of individual participant characteristics. Outliers for the separate conditions add up to the sum of outliers for each experiment. Effect of cue type (view cue, gender cue, or linguistic cue) and item presence in training in the test data for Experiments I and II. New York, NY: Springer-Verlag. B. (2014). In the test phase, overall, participants pick the answer containing the correct suffix significantly more often than the answer with the correct stem (the original name) (59% of the time). The core meaning is bottom-up and perceptual (a salient entity differs from its environment). Table 12. Acad. All are native speakers of American English. Mean age is 32 years, with a standard deviation of 9.99. The stem of the two available responses is always the same, the name of the query, which is also visible on the screen.

B., Bent, T., Munson, B., Bradlow, A. R., and Bailey, J. M. (2004). Hoboken, NJ: Wiley-Blackwell. For models for which age was justified as a predictor, the reported models exclude the few participants for which we have no age data. Three waves of variation study: the emergence of meaning in the study of sociolinguistic variation. One grouping, gender, is the identity of the conversation partnerwho is either male or female. Rapid word learning under uncertainty via cross-situational statistics. Those are also comparable. Social salience has a top-down effect. 2. In this paper we introduce an experimental paradigm that facilitates investigation into the contextual learning of morphological patterns. The editor and reviewer's affiliations are the latest provided on their Loop research profiles and may not reflect their situation at the time of review. Berlin: De Gruyter - Mouton. Table 6. If the non-linguistic context is salient, it is more readily available as a factor in variation. doi: 10.1111/1467-9280.00476, Goujon, A., Didierjean, A., and Thorpe, S. (2015). Ramscar et al. These studies do not look at extension or generalization to different speakers. However, in order to attempt an objective test that there were not strong visual differences between the different dimensions, we computed the Levenshtein distance between uniformly binned histograms of the grayscale versions of the images using Matlab (Mathworks, 2016). The nonce language we used is similar to Experiment I, except that the vowel harmony pattern is absent. Good learners and poor learners across cue type, Experiment II. Our hypotheses were based on the the results of Experiments II and V. We expected that participants would learn the contextual association more easily in the gender condition than in the view condition. Three important findings emerge from it.

MS, University of Rochester (Accessed March 17, 2014). Experiment IV focuses on the way participants rely on the denotative and the contextual aspect of the naming pattern. Figure 5. With a spatial cue, they overwhelmingly prefer the denotative aspect. 3, 5575. Listeners can, in turn, use such patterns to infer speaker characteristics or to adapt to different speakers when processing speech (for review of relevant literature see Hay and Drager, 2007; Foulkes and Hay, 2015). The contextual learning literature that is most relevant to this paper focusses particularly on the role of the broader extra-linguistic contextabove and beyond the referentin learning a linguistic category. Trends Cogn. We find these results with more types of conversation partners (such as children and adults) and with two distinct morphological patterns, the diminutive and the plural. doi: 10.1146/annurev-anthro-092611-145828, Fagyal, Z., Swarup, S., Escobar, A. M., Gasser, L., and Lakkaraju, K. (2010). A reviewer pointed out that the lower participant count is problematic given that we compare this experiment to Experiment III. It includes the addressee and the discourse situation, as well as the speaker's attitudes and ideologies, which are conjoined to give social meaning to a given utterance (Eckert, 2008). B. However, the link between contextual language learning and the observed structural complexity of social language use is far from completely understood. Foulkes, P., and Hay, J. There were substantial individual differences in learning. Unlike in Experiment I, item presence in training has no effect on response accuracy. One possible response for the target has the stem which is the name of the query of the target (as seen in training) but a context-inappropriate suffix (this is a choice present in the previous experiments). The two possible names of the response picture are the name of the query picture plus one of two suffixespek and pak. B., and Drager, K. (2010). We stepwise fit a binomial mixed-effects regression model on the test data, using response to individual items (correct or not correct) as an outcome variable and the interaction of cue type (gender or view) and item presence in training, conversation partner presence in training, conversation partner type in training (children or adults), and participant age and gender as predictors, with a participant random intercept. Four participants were excluded based on training length. People who go on to pick the suffix in test are much faster to finish training than people who go on to pick the stem. Further research using other images and other contrasts is therefore still required. J. Mem. The term relevance often implies conscious decision making on which contexts to consider and which to discard. doi: 10.3765/bls.v19i1.1531, Kraljic, T., Brennan, S. E., and Samuel, A. G. (2008a). Johnson, K., Strand, E. A., and D'Imperio, M. (1999). This, in our experience, is not an excessively high number for an online experiment. Six occur only in test. Small horizontal lines show individual values; longer if multiple individuals have the same average. Increased training improves participant accuracy in test. Phonol. However, even participants who were good learners needed to make a number of mistakes to learn the pattern. Work by Docherty et al. Experiment II looks at the association of a morpho-phonological pattern and a non-linguistic context. Relatively few studies have investigated learning that involves novel social groups, or learning of socio-linguistic cues other than ones at the phonetic level. The next toolkit. This is true even when little evidence of generality is actually given. (2014), who also recruited participants on Amazon Mechanical Turk. However we will explicitly test the degree to which the learning to generalized to other speakers on the basis of gender in later experiments. Does a varying non-linguistic context aid the learning of a linguistic pattern? 34, 409438. Our experiments use a training-test paradigm based on a simplified version of adaptive tracking. If we tabulate good learners across cue type, we find that the gender cue is easier to learn. Spat. Brooks, P. J., Kwoka, N., and Kempe, V. (2016). Participants were paid three dollars upon completion of the task. We return to the training phase and discuss our exclusion criteria in detail in the Results section of Experiment I. The pattern type is the diminutive rather than the plural, but Experiment V provided little evidence that this would be a relevant dimension. We fit a regression model on the test phase of Experiment VI following the same procedure as in the previous experiments. Training consists of six items, so it has 4 6 = 24 targets in total. However, the test phase is different. Statistical learning as an individual ability: theoretical perspectives and empirical evidence. If context is important, how do we learn to use it? (2012). They are told that the bird is the protagonist (our hero), and that they need to help our hero return to its nest by flying from roof to roof. Distributions of training trial counts for the two conditions for all participants, Experiment II. doi: 10.1111/j.1467-9841.2008.00374.x, Eckert, P. (2012). The experiment was hosted on Amazon Mechanical Turk. If we look at the distribution of good learners across cue type, we find that most good learners are to be found in the gender condition (cf. Since stem vowel is no longer relevant, we used the five English vowel letters to make the syllables more distinct. Four participants were removed based on training speed. We therefore evaluated the same hypotheses for Experiment V as for Experiment III. In Experiment II the same morphological pattern is presented, again with different conversation partners.

US worker IDs are independently verified by Amazon, making it very difficult for the same person to operate multiple accounts. The adaptive tracking paradigm described in Leek (2001). doi: 10.3758/BF03194543, Leung, J. H., and Williams, J. N. (2012). Language and the free-rider problem: an experimental paradigm. However they do provide some first steps toward shedding some light on the complex cognitive mechanisms that must be at play in such learning. Mean accuracy is much higher for the gender condition.

Participant gender is not a significant predictor. doi: 10.1515/labphon.2011.007, Walker-Andrews, A. S., Bahrick, L. E., Raglioni, S. S., and Diaz, I. Sex and gender have a complex effect on the use of social meaning in general (Milroy and Milroy, 1993; Cheshire, 2002). The number of participants removed for each experimental condition ranges between 1 and 2. Brooks et al. Rcz, P. (2013). Models of category learning and memory retrieval (Ratcliff, 1978; Grossberg, 1987) operate using notions of context-specificity. (2016), for example, have shown that morphological learning and generalization varies across individuals, in a way that correlates with measures of non-verbal intelligence and general statistical learning abilities. In this phase, participants need to guess the names given to small objects, just like they did in the first part. indiv. 1, ed A. Duranti (Oxford: Blackwell), 382401. A participant cannot finish too quickly, and so the distribution of training trial counts has no left tail. Age is a significant predictorolder participants are more likely to give correct answers. While individual trials vary in duration (there is no time limit on trial length, that is, people can spend as much time as they want on their decision), they do so to a modest degree (in Experiment I, mean (m) = 16 seconds, standard deviation (sd) = 12 seconds). Bull. Good learners across cue type, Experiment III. If future work establishes that socio-indexical associations for plural patterns are indeed more difficult to learn than for diminutive patterns, the reasons for this difference would be of considerable interest. J. Appl. Those who focus on the stem ignore social contextual cues. If they give an incorrect response, they have to return to the previous target. We remove outliers to safeguard against participants with very poor attention. Spatial context and top-down strategies in visual search. In Experiment I, which has one across-subject condition, we removed 2 participants and report data from 45 participants. We explain our design in detail in Section 4.2 and address changes to it in subsequent Methods sections. People learn both deictic expressions (denoting spatial relations, such as here and that) and words with implicit spatial relations (such as wide or tall) easily, since these are frequent forms of every language. (1991). As we note in Section 3, the perceptual differences between the images are unlikely to affect their categorization. Six occur in training and then also in the test. Linguist. We have given a review of the literature to show that the non-linguistic context is extremely influential in learning linguistic constructions. For each across-subject condition, we removed the 2.5% of participants who took the most trials to finish trainingi.e., the slowest ones. The correct response no longer depends on the vowel of the prompt. Note that, if our participants in this task learn the association of the linguistic pattern with conversation partner, we have no way of knowing whether they are imputing a person-specific pattern, or a more general distinction based on person gender. 19 (Berkeley, CA). doi: 10.1080/03640210802035357, Metcalfe, J., Casal-Roscum, L., Radin, A., and Friedman, D. (2015).

However, only about half the participants exposed to the salient cue show high accuracy in test, while the other half of this group resorts to guessing, much like participants learning the non-salient cue. For each participant, two syllables are randomly selected as suffixes (marking conversation partner gender or spatial orientation, depending on the condition) while the rest are randomly assigned as item names. (2014). Environmental context-dependent memory: a review and meta-analysis. Linguistic constructions have denotative meaning and social meaning. In all experiments, we filtered participants within the across-subject conditions because we expected these to vary in length. Figure 9 shows a bean plot of participant test responses for the gender condition and for the view condition. Contextual cueing: implicit learning and memory of visual context guides spatial attention. (2015). Small horizontal lines show individual values; longer if multiple individuals have the same average. Experiment V is identical to Experiment III except for the prompt and target images. We look at this question in Experiment VI. doi: 10.1371/journal.pone.0019009, PubMed Abstract | CrossRef Full Text | Google Scholar, Baayen, R. H., Hendrix, P., and Ramscar, M. (2013). Age and gender differences in social problem-solving ability. 35, 158179. In the rest of the experiments, it depends on the conversation partner (as in this example). Participants were paid three dollars upon completion of the task. Docherty, G. J., Langstrof, C., and Foulkes, P. (2013). The aim of this design is to teach naming patterns in conjunction with the contextual cue provided by the grouping. We also evaluated age and gender as potential predictors of performance. Each occurs once with each item in training. Distributions of participant responses on test items in the view and gender conditions, Experiment VI. Kraljic and Samuel (2006) show that a phonetic category distinction is generalized to a new linguistic context and also to a new speaker. In the training phase, the stimuli were generated from the same pool as in Experiment III. As we note above, training takes longer. Sily, T., and Suomela, J. Conversation partners can differ across various dimensions. The test data distributions generally showed two distinct modes, one for good learners and one for poor learners. The adaptive tracking training enabled us to examine the differences between good learners and poor learners in more depth. The best model has age and cue type as significant predictors. Auditoryvisual integration of talker gender in vowel perception. The instructions told the participant to identify the name of the small item based on the larger item. The same logic was applied to subsequent experiments. What remains unclear is whether participants make a by and large random choice at the beginning to focus on either context and then remain with it, or whether the effect of the non-linguistic context can be increased by expanding training. doi: 10.1177/0261927X15584682. Are they able to generalize the pattern to new words? R Core Team (2016). (2006). Small horizontal lines show individual values; longer if multiple individuals have the same average.