Each general is allowed to state his case in his own words and the author additionally draws upon a multitude of third party opinions in the form of politicians, civilians, fellow general officers, and private soldiers. As might be expected, the centerpiece of A Crisis in Confederate Command is the Red River Campaign, the aftermath of which expedited the final break between Smith and Taylor. Prushankin has consulted a thorough array of primary sources and a well-chosen set of secondary sources to craft an excellent command history of the campaign.
E. Kirby Smith | American Battlefield Trust
His fresh ideas and insights are welcomed. Contemporary commentary from outside the Trans-Mississippi theater is thoughtfully included as well. Readers interested in Civil War command relationships in general and the Trans-Mississippi theater in particular should reserve a space on their bookshelf for this excellent study. This review is reprinted with the permission of North and South Magazine , originally appearing in vol.
However, in some cases as in this first model , the maximal model will not converge; as a minimal step toward parsimony, we thus do not require a correlation parameter between random slopes and intercepts Bates et al. Table 1 shows the resulting mixed-effects model for the KCS Experiment 2 data.
Figure 1a summarizes the raw compositionality scores that form the basis for this model, and Fig. The Fig. Note that the effects plotted in this figure fixed effects, as well as chain-specific random effects show model predictions, rather than raw compositionality values. Fixed effects are shown in black, with random-effect estimates for the four individual transmission chains shown in gray. Overall, the model clearly exhibits an increase in compositionality via iteration, corresponding to the positive, linear term for generation. Moreover, by the second generation of iteration, compositionality scores consistently surpass the threshold designating random structure the horizontal line in Fig.
The model, however, also includes a negative quadratic term, corresponding to a downward curve across generations.
In the fixed effects as well as the random effects of Fig. That is, two of the four transmission chains decline markedly in compositionality toward the end of the transmission chain, and the model is unduly influenced by these presumably chance events. Although the basic findings in KCS are highly noteworthy, we would argue that more data is altogether appropriate. The current study thus aims to replicate KCS, with a substantially larger amount of data. Next we consider an additional open issue in the KCS data.
Earlier we noted that in KCS Experiment 1, participants tend to neglect some meaning dimensions, while others are encoded more systematically. This raises the question of whether there are any patterns of semantic biases across the three dimensions shape, color, motion in KCS Experiment 1, and whether related biases may be observed in Experiment 2. Thus we conduct a new analysis of the KCS data to consider this question, as follows. The same measure is then calculated for the other two dimensions of meaning shape, motion.
The resulting within-category similarity score is meaningful insofar as there may be differences between dimensions—that is, at a given generation t , some dimensions show more category-internal similarity than others. The averaged by-dimension results of our reanalysis are shown in Fig. Reanalysis of data from KCS Experiment 2, showing the timecourse of development for within-category similarity for different dimensions of meaning color vs. The evident pattern is indeed statistically significant, as borne out by a mixed-effects regression analysis, using within-category similarity as the dependent variable.
Table 2 Mixed-effects regression model for amount of within-category string similarity, in data for our reanalysis of KCS Experiment 2, by dimension across generations. Dimension is a categorical variable, with shape as the reference level. Model: in. Mixed-effects regression model for amount of within-category string similarity, in data for our reanalysis of KCS Experiment 2, by dimension across generations.
In the KCS data, within-category similarity generally increases, that is, generation is a significant predictor, with a small negative quadratic effect. The generation interaction effects for color red, black , or blue and shape circle, square , or triangle are not significantly different from one another. These analyses indicate that one dimension in particular motion tends to lead the way in the emergence of structure in the KCS artificial languages.
In the remainder of this article, we revisit these open questions with a new iterated learning experiment, drawing data from a much larger group of participants. Our experiment is in many respects fashioned after Experiment 2 in KCS, and has been adapted to online data collection methods using AMT. AMT has proven to be a reliable platform for behavioral research, and provides resources for the recruitment of a large and diverse population of participants Munro et al.
At the same time, online data collection using AMT presents challenges that are not present when participants are highly educated and motivated university students tested in the lab as in KCS. The elaborations of the original paradigm described here reflect extensive pilot testing designed to overcome these challenges. We wanted to ensure completion of the task within a maximum time frame and minimize the number of dropouts, while also automatically eliminating the few participants who respond at random without following instructions.
Several innovations of our experiment's setup are similar to but developed independently from approaches used in Carr That study failed to find evidence of increased compositionality via iteration, but introduced several variations on the original KCS design. Our current experiment shares with Carr 1 a redefined meaning space, which is focused on unambiguously noun-like units; 2 a considerably broader syllable repertoire in the initial state, and 3 the use of a fixed training set size.
The alien objects used in the current experiment are illustrated in Fig. Graphics: Visual Voice vvlab. It is arguable whether the linguistic units of interest in this study are morphological or syntactic in nature. The referent stimuli are objects and their properties; the relevant linguistic units are nominal, and could alternately be viewed as a noun phrase, or as a noun stem with affixes. By comparison, recall that the meaning space of the KCS experiments involves types of motion for items differing in shape and color.
The productions in the KCS experiments are thus presumably morpho- syntactic, involving a subject and a predicate. The linguistic forms in the current study, as in the KCS experiments, appeared with no internal spaces. This artificial grammar generates linguistic forms such as vilkantin, tinkalsol , and kalvonsi.
See a Problem?
The language initialization differs from that in KCS by having forms of approximately constant length, but comprised of a broader range of different syllables. These changes in the protocol were made to avoid possible artifacts from the KCS initialization method; see Supplementary Text S3 for further discussion. For each run of the experiment, twenty-seven linguistic forms were selected at random from the set of all possible forms, and assigned randomly to the twenty-seven meaning configurations. For the first generation of players, N items were randomly chosen out of the full set of twenty-seven; these constituted the training items for first-generation participants.
An alien avatar on the screen asked players to attempt to learn an alien language; the full game instructions are presented in Supplementary Text S1. After each training block, participants also completed an interim task requiring open-text responses. Full details regarding the sequence of training rounds are provided in Supplementary Text S2.
As in KCS, the instructions never alert participants to the fact that they are being tested on items which they have not encountered during training. As in KCS, participants in successive generations were not informed that the language they were learning had been produced by other players. However, as part of the consent process, participants were informed that their answers during the game could be used to create new versions of the game.
Before transmission to the next generation, the output languages were adjusted as follows. As in KCS, the training set for each new learner comprised a subset of the language output from the previous generation. However, some selection processes were imposed by experimenters at the intergenerational stage.
As in KCS Experiment 2, no two identical forms were allowed in any training set, to avoid a tendency toward underspecification. On this point, the procedures in our study for randomly selecting training items are different from those of KCS. In KCS Experiment 2, when identical forms occurred and happened to be randomly selected for the next generation , all homonyms but one was removed from the training set. However, these filtered items were not replaced, thus leading to variable training set sizes dipping as low as eight items out of twenty-seven depending on how often the language repeated the exact same form.
For each removed item, we then selected replacement candidates at random from the same language while continuing to disallow identical forms as candidates , and used these as training items for the next generation. We introduced a fixed size in our experiment for several reasons. Additionally, imposing consistency across different chains and generations is helpful for interpretive purposes. One of the KCS metrics—the amount of intergenerational change, by generation—is difficult to evaluate meaningfully if the amount of training input fluctuates.
Holding the training set size constant imposes an additional layer of data filtering beyond the filtering implemented by KCS.
- Putting your Business Online: The Small Business DIY Guide (A Do It Yourself Guide for Small Business Owners Book 1).
- Edmund Kirby Smith - Wikipedia;
- The paladins of Edwin the Great (1908)?
- The Kingdom Scroll.
In our experiment, in cases where the output language contained fewer than N items, we discarded the output and reran the exact same experiment setup with a new participant. Thus, in effect, the filtering processes in our current approach imposes more than one selection pressure on the output. Second, the procedure removes altogether any output from participants who have more severe tendencies toward homonymy. However, the removal of at least some participants is unavoidable if we impose the requirement of a fixed training set size, while also filtering homonyms.
The experiment was limited to native speakers of English, aged 18 and older. The Mechanical Turk assignment was set up such that each participant identified by a Mechanical Turk ID could complete the experiment only once. Thus, there were a total of twenty-four distinct transmission chains initialized with random form-meaning mappings the zeroth generation ; these structureless languages were provided as training for twenty-four participants, and the process iterated for a total of ten generations of participants.
However, the requirement of a fixed training set size led to the rejection of some candidate participants, and replacement with new participants prior to iteration, since these participants failed to provide a sufficient number of distinct answers to be used on the next generation. For the runs with a training set size of twelve, eleven participants were discarded for having fewer than twelve unique linguistic forms in the final testing round rejection rate: 8. Among participants with a training set of fifteen, sixteen were discarded for providing fewer than fifteen unique forms rejection rate: Two participants were replaced for failing to follow experiment instructions.
In addition to the foregoing, a number of participants elected not to complete the experiment. In the training-set-twelve condition, thirty-nine participants began the study but dropped out before finishing In the training-set-fifteen condition, sixty-seven participants began the study but dropped out before finishing For the trials with a training set of twelve items, the final set of participants consisted of 72 women and 48 men.
This version of the experiment lasted an average of For the trials with a training set of fifteen items, the final set of participants consists of 56 women and 64 men. The mean age is The experiment lasted an average of Participant responses in the final test round comprise a total of 6, open-text responses 3, responses per training-set condition. Note that in a few instances in the dataset, participants failed to provide a response within the second timeout see Section 2. In the training-set-fifteen condition, a total of nine nonresponses occurred again out of 3, entries.
In this condition, one participant provided two different NA responses, and a different participant provided three different NA responses. The four remaining NAs were the only nonresponses for the participant. We quantify language compositionality using the same methods as KCS and Cornish , which adapts a test developed by Mantel The Mantel score is devised so as to quantify the relationship between forms and meanings in a language; across different linguistic items in a compositional language, similarities in meaning should correspond to similarities in form.
Monte Carlo methods are used to determine a threshold for identifying likely compositional languages. These quantitative methods are discussed in greater detail in Supplementary Text S4. Summaries of the compositionality scores across ten generations plus the random initial state are presented in the boxplots of Fig.
Monte Carlo investigations see Supplementary Text S4 confirm that as intended, none of the twenty-four languages in the current study have significant compositional structure at the randomly initialized zeroth generation. Boxplots of language compositionality over ten generations plus random initial state. The plots summarize twelve chains using a training set size of twelve Fig. The compositionality metrics displayed are z -scored with respect to 1, randomized rearrangements of the linguistic forms and meanings. The horizontal dotted line at 1. However, the plots in Fig.
The structure does seem to emerge somewhat slower than in earlier work. In KCS Experiment 2, there was a rapid increase in compositionality on early iterations; nonrandom structure was evident on two of the four chains by the first generation, and on all four chains by the second generation. The general trend for increasing compositionality in the current dataset can be verified, again using mixed-effects linear regression. Our model investigates the effect of generation , as well as higher order generation terms to allow for nonlinearity. We performed stepwise regression, examining interactions between generation terms and training size N.
Training set size is not significant, either as a main factor or interaction with linear or higher order terms , and thus it is dropped from the model. The resulting regression model is presented in Table 3 , and fixed and random effects are plotted in Fig. Plot of the compositionality predictions for the model shown in Table 3 , based on our dataset of twenty-four transmission chains. Fixed effects are shown in black, with random effects for the twenty-four individual chains shown in gray. Table 3 Mixed-effects regression model for compositionality. Several things are of note in this model.
Given our initialization of languages with random structure, this aspect of the model is as expected. However, some variation in intercepts is evident in the random effects, which is consistent with slight variations in the initial conditions for our twenty-four chains. Moreover, the model gives evidence of an overall increase in compositionality via iteration, represented by the positive coefficient for generation. Without government funds, which account for the overwhelming bulk of revenue, few of these institutions could attract students or stay in business.
The continuing flow of money illustrates the quandary facing federal education officials. On one hand, they have moved forcefully to try to protect taxpayer funds and prevent students from falling deeply into debt without anything to show for it. On the other, they must avoid running roughshod over private for-profit schools that have not been found guilty of wrongdoing.
Agency officials point out that they cannot withhold money based on accusations, but must have proof of misconduct. Regulators are caught between an industry that says it is being unfairly demonized by opponents and critics who complain not enough is being done to prevent fraud and abuse of vulnerable students. Kinser pointed out that the Education Department had little flexibility under the law when it came to cutting off federal student loan and grant money to potential abusers.
Education officials say they have clamped down on many for-profit schools, restricting their ability to expand their programs or the number of campuses, capping the number of students eligible for student loans, or requiring schools like Education Management to post a letter of credit to gain access to federal student loans and grants. The letter is meant to protect students and taxpayers if the company is unable to cover federal student-aid liabilities.
Still, critics say that even schools with egregious violations have become adept at exploiting loopholes, sidestepping rules or taking advantage of yearslong appeals processes. Companies with several campuses can pool graduation, financial, enrollment, staffing and other statistics to mask weak performers, experts say. The big schools know how to work the numbers to avoid failing.
But several schools have figured out how to lower their rates by getting students temporary deferments or forbearances so they fall outside the three-year window. Nearly schools have a student loan default rate that exceeds 30 percent. Elam said. In recent years, more than two dozen companies that run for-profit colleges have been investigated or sued by state prosecutors.
Related GENERAL KIRBY-SMITH, Annotated and Illustrated
Copyright 2019 - All Right Reserved