MiniCorp Info Page

What is MiniCorp?

Last updated on January 24, 2008

MiniCorp (currently implemented only as MiniCorpJS 0.5) is a tool for small-scale corpus analysis. It attempts to bridge the gap between traditional phonological analytical methods with quantitative corpus analysis. Though the focus in the current version is on analyses within the theoretical framework of Optimality Theory, MiniCorp differs from other OT programs in that it is not an automatic grammar learner, but rather a tool for exploring dictionary-based corpora and for testing prespecified phonological hypotheses. MiniCorp helps with the tagging of lexical items, partially with the help of regular expression matching, and counting lexical neighbors. The statistical analyses test both each proposed OT constraint and their ranking. Currently the statistics is handled with R. Though intended primarily for phonological analyses, MiniCorp may also be useful for morphological analyses as well.

Why phonological corpus analysis?

Despite the traditional skepticism about corpus data in generative linguistics, most contemporary phonological research is actually a form of informal corpus linguistics. Most of the data cited in phonological research aren't elicited native-speaker judgments of novel forms, as in syntax, but fixed sets of lexical items or phrases, usually taken from dictionaries. These data sets are corpora, since they are preexisting rather than generated through experimental manipulation. Even Chomsky and Halle (1968), written by fluent English speakers, was based primarily on dictionary data.

Phonologists have much to gain from experiments, whether they involve native-speaker judgments (e.g. Frisch and Zawaydeh 2001), phonetic measurements (e.g. Port and Leary 2005), or something else (e.g. Ohala 1986), but most phonological argumentation is still built on the reasonable assumption that corpus patterns reveal something about the mental grammar that went into producing them. Moreover, since judgments of word-sized items, even nonsense words, are affected by lexical properties, work on phonological judgment experiments must start with corpus analyses. (MiniCorp can help with this sort of corpus analyses as well, helping you to develop materials for MiniJudge experiments.)

Though phonologists traditionally do not use the sophisticated quantitative methods commonly associated with corpus analysis (e.g. Frisch et al. 2004, Uffmann 2006), traditional phonological analyses are implicitly quantitative because they assume that more common word types are more likely to reflect grammatical patterns, and less common word types are more likely to be exceptional. This follows directly from the notion that grammar is systematic, and evidence for systematicity is (by definition?) stronger the higher the proportion of conforming examples relative to exceptional examples. Of course frequency is not identical with grammaticality, as has often been pointed out by generative linguists (starting with Chomsky 1957), but this is simply because frequency is an aspect of performance, and so is affected by forces in addition to grammar, both systematic (e.g. analogy) and accidental (e.g. historical borrowing). The relationship between grammaticality and frequency is exactly parallel to that between grammaticality and acceptability; acceptability is affected by factors other than grammar, but this doesn't mean that acceptability cannot reflect grammar as well.

For example, consider vowel harmony in a subset of the lexicon of the Formosan language Pazih (data from Li and Tsuchida 2001; see further discussion in Myers 2007). Pazih highlights the importance of corpus analysis in phonology, since like many languages of traditional societies, it is in danger of extinction. Thus there is no way to run proper experiments to test the synchronic productivity of its patterns, and we must do the best we can with the corpus data alone.

Pazih has 45 morphemes containing a reduplicated CVC and an intervening vowel (a complete list is given in this RTF file, encoded in Doulos SIL font). In most cases (33/45), the intervening vowel is identical with that in the CVC components, as in (1) below. However, there are 12 exceptional non-harmonizing examples, as in (2).

(1) hur-u-hur "steam, vapor"
(2) hur-a-hur "bald"

In terms of Optimality Theory (OT), the harmonizing examples suggest a markedness constraint that we may called AgreeV. The existence of exceptions may be handled by a faithfulness constraint like IdentIO. If we assume that the intervening vowel is prespecified in the exceptional items but not in the harmonizing items, we can analyze the data set with the grammar in (3).

(3) IdentIO >> AgreeV

This is an entirely ordinary OT analysis, but note the essential role that corpus frequency plays in the argument. If the proportion of exceptions were not 12/45 but rather 40/45 or so, it would be much harder to argue that this language shows evidence of AgreeV. Similarly, if the proportion of exceptions were 1/1000, we might dismiss the single exception as a random accident, a bit of "performance noise" irrelevant to the underlying grammar.

The primary purpose of MiniCorp is to make the quantitative nature of such arguments fully explicit. For example, in the Pazih data, do we really have enough evidence to support the hypothesized AgreeV and IdentIO constraints, as opposed to the null hypothesis that the apparent patterns are due to chance? (After all, 12/45 is still a quite high proportion of exceptions!) Even if both constraints can be shown to be relevant, are we really justified in ranking them as in (3)? The following sections explain how MiniCorp answers these questions.

Why focus on Optimality Theory?

The first reason MiniCorp focuses on Optimality Theory (OT) (Prince and Smolensky 2004) is that it has become the lingua franca of theoretical phonology, so most phonologists are quite comfortable thinking in OT terms.

The second reason is that an OT analysis is essentially the application of a paper-and-pencil computational algorithm, one that has relatively straightforward mathematical properties (see e.g. Smolensky and Legendre 2006). In particular, the OT constraint ranking is a special case of the constraint weighting used in many computational and statistical models. OT constraints also provide "surface-true" descriptions of data, in the sense that they can only be violated on the surface if blocked by another (higher-ranked) constraint. This means that the relationship between an OT grammar and the data is more transparent than is the case for derivational grammars. (This doesn't mean that OT is necessarily more "correct", only that it is computationally more convenient.) These aspects of OT are exploited by MiniCorp, as well as several other recent OT programs, though they do so in rather different ways.

The third reason for focusing on OT is that the notion of universal structural constraints helps solve the fundamental logical problem of corpus analysis, namely, how to restrict the hypothesis space. The inherently proper analysis of an experiment is defined by its design; if we're testing factor [+F] vs. [-F], then we analyze the results to see how likely it is that our observed difference between [+F] vs. [-F] could have arisen by chance. For a corpus of naturally occurring items, there is no inherently proper framework for testing whether a pattern is statistically significant. For example, the chance probability of choosing any particular card from a deck of cards is p = 1/52 = 0.019, which is "statistically significant". Clearly we must specify ahead of time what counts as a "winner". OT constraints make this possible, assuming that they have some a priori theoretical reason for being plausible. Derivational theories, like pre-OT generative phonology, do not offer as explicit a set of hypotheses for inherent frameworks for corpus patterns.

Despite the advantages of OT for the purposes of MiniCorp, derivational approaches remain important, even in the OT literature (e.g. in stratal OT; Kiparsky 2000). Hence future versions of MiniCorp will permit tests of rule ordering as well, perhaps building on statistical models like that proposed by Sankoff & Rousseau (1989).

How is MiniCorp different?

Given the historical relationship of OT with computational models, as well as its emergence during an era of ever-growing computer power, it's no surprise that there are many computer programs designed with OT in mind. Up till now, however, all of them are intended primarily as automatic grammar learners. That is, the user inputs a corpus, and the program outputs a grammar. The most commonly used OT (-related) programs and best-cited references are the following:

Praat (Boersma and Weenink 2007): Implements various OT learning algorithms, including the Error-Driven Constraint Demotion for standard OT grammars (Tesar and Smolensky 1998) and the Gradual Learning Algorithm for stochastic OT grammars (Boersma and Hayes 2001).
OTSoft (Hayes et al. 2003): Implements some of the same learning algorithms as Praat, among other tools for creating OT analyses.
HaLP (Potts et al. 2007): Implements a learner for harmonic grammars, where constraints are weighted rather than ranked.
Goldwater and Johnson (2003): Describes another type of OT/harmonic grammar learner.
Hayes and Wilson (to appear): Describes a harmonic grammar learner using an algorithm similar to that of Goldwater and Johnson (2003), but which can also learn constraints (the other models listed above require prespecified constraints).

There are presumably two reasons why OT programs focus on grammar learning. First, given that corpus data are collected without any prespecified hypothesis, it makes sense to analyze it without any prespecified hypothesis. Second, since Chomsky (1965), generative linguistics has emphasized the theoretical importance of language acquisition, since a universal, innate learning algorithm represents the one "true" corpus analysis relevant to the description of adult mental grammars.

Yet most phonological analyses in the theoretical literature do not use corpus data in this way, as mere input to learning algorithms. Instead, as we saw in the Pazih example above, phonologists first posit a set of generalizations about the corpus and then attempt to convince their audience that these generalizations are the best way to describe the data set. That is, a corpus is seen as performance data generated by a process involving grammar as one of its components, and the task is to test claims about what this grammar is like. The primary purpose of MiniCorp is to help automate this task.

Though some of the programs and algorithms described above use techniques similar to that of MiniCorp, none of them is designed to test the statistical reliability of grammars. This is not their goal, of course, but note that these models depend on the assumption that actual children also ignore statistical reliability. The result is that these models all show one-trial learning. That is, if they are trained on a corpus with some large number of items, where one and only one item requires the ranking Constraint₁ >> Constraint₂ and the rest say nothing, then this ranking will become part of the mature grammar. This implausible result follows from the fact that these models seem to neglect the difference between competence and performance, and treat raw corpus data as direct (not noisy) reflections of the underlying grammar.

What is corpus tagging?

One way in which corpus analysis is never really a purely "bottom-up" empiricist exercise is that corpus linguists rarely, if ever, analyze raw corpora. Instead, they annotate the corpus items first, in particular demarcating structure (e.g. word boundaries in a corpus of fluent language use) and tagging demarcated items for the classes they belong to (e.g. indicating part of speech). This job requires native-speaker competence and/or a priori assumptions about what aspects of the language productions recorded in the corpus are linguistically relevant. Subsequent analyzes are then conducted in terms of the tagged properties, not solely in terms of the raw corpus itself.

For the purposes of MiniCorp, we want a tagging system that indicates the properties of the lexical items most relevant to the OT analysis that we want to test. Fortunately, it turns out that OT constraint violations themselves can serve as tags. This was first discovered by Golston (1996), who showed that the representation of a form is equivalent to the constraints that it violates. For example, the form ab can be represented in terms of violations of Onset, NoCoda, *a, *b; ba as violations of *a, *b; a as violations of Onset, *a; and so on.

MiniCorp uses a similar idea. After entering a corpus, the user chooses the constraints defining the analysis, and then marks each item for whether or not it violates that constraint. (In the initial version of MiniCorp, each constraint can receive at most one * mark per item, though this limitation will hopefully be fixed soon.) Thus in the Pazih data set, we would tag items as in (4). Note that by hypothesis, IdentIO is never violated: not in harmonizing items because we assume the intervening vowel is not specified, and not in exceptional items because there the output vowel matches the prespecified input vowel.

(4)

Corpus IdentIO AgreeV

hur-u-hur

hur-a-hur *

... ... ...

Note that (4) is not an OT tableau, but merely a tagged corpus; the items aren't related to each other and the ranking doesn't matter (yet). Moreover, though in this case there are no violations of faithfulness constraints, and Golston (1996) argued against faithfulness constraints in general, MiniCorp allows the user to tag items for violations of faithfulness constraints as well as markedness constraits.

What are regular expressions?

Though tagging is essential to corpus analysis, it is very difficult to do. With a corpus as tiny as the Pazih lexicon, it is not hard to hand-tag each item, but this option is not very practical with a full dictionary of tens of thousands of items. In quantitative corpus linguistics, sophisticated algorithms have been developed to automate the process as much as possible, as for part of speech tagging.

One tagging assistant offered by MiniCorp makes use of another familiar component of corpus linguistics: regular expressions. A regular expression is a notation system for representing a set of strings. The most familiar example is the "wildcard" symbol offered by many search systems: if ? is the wildcard symbol, then phonolog? would represent the set phonology, phonological, phonologist, phonologists. As demonstrated by Karttunen (1998), regular expressions are useful for defining OT constraint violations because such violations involve classes of strings (i.e. all the items that violate the constraints). Another reason they work so well for phonology is that unlike syntax, phonology and (most) morphology can be expressed in terms of regular grammars (Savitch et al. 1987).

MiniCorpJS uses the built-in regular expression engine of JavaScript, which follows the internationally accepted Basic Regular Expressions (BRE) standard. This system uses a variety of symbols (for example, the wildcard symbol is "."). Perhaps the quickest way to become familiar with it is to study examples. Here are some markedness constraints in a hypothetical language with the segment inventory {p, t, k, b, d, g, a, i, u}:

No_tp: tp
xVoice: [bdg] ([xy] = "x or y")
AgrVoice: ([ptk][bdg])|([bdg][ptk]) (| = "or")
NoGeminates: [ptkbdg]* (* = "any number of previous segment")
AgrV: (i.*[au])|(a.*[iu])|(u.*[ai]) (. = "any segment")
xFinalLab: [pb]$ ($ = "end of string")
xInitialLab: ^[pb] (^ = "start of string")
Onset: (^|\.)[aiu] (\ = "treat following segment literally"; assuming syllable breaks are marked "." in corpus)
NoCoda: [ptkbdg]($|\.) (assuming syllable breaks are marked "." in corpus)

Faithfulness constraint violations are harder to represent since the inputs are not expressed directly in the corpus (at least in the current version of MiniCorp), but MiniCorp can do half the job by marking all items that are potential violations. If the process is neutralizing, you have to toggle off violations for the faithful items manually. For example:

IdentIOVoice: g (assuming [g] is always derived from /k/ and nothing else becomes voiced)
DepV: ($|\.)[ptkbdg][aiu](\.)[ptkbdg] (for vowels epenthesized in underlying onset clusters; marks for items with underlying vowels in this position must be toggled off manually)

Given this notation system, the regular expression defining violations of AgreeV in Pazih is as shown in (5).

(5) ([eiu].-a)|([aiu].-e)|([aeu].-i)|([aei].-u)

Entering this into the appropriate box in the MiniCorp tagging table will automatically add * marks to all of the Pazih items violating AgreeV.

What are lexical neighbors?

Since a corpus represents preexisting data, it is natural to want to explore it in a more unstructured way in order to discover any latent patterns worthy of more careful follow-up. MiniCorp currently has only one tool along these lines (though more are planned): finding lexical neighbors.

Lexical neighbors are lexical items that are "sufficiently similar" to some target item. The notion of "sufficiently similar" can be defined in various ways, but most definitions start with the notion of edit distance, which counts the minimum number of deletions, insertions, and/or replacements of units needed to change one item into another. In OT terms, edit distance can be thought of as measuring violations of a sort of output-output correspondence that applies between any pair of items, related or not (Myers 2002). Thus a target item, as compared with a lexical item, earns one * mark for each violation of MaxOO (deletion), DepOO (insertion), and IdentOO (replacement).

Given the notion of edit distance, the simplest definition of neighbor is a lexical item that has edit distance of 1 from the target (Luce 1986). Neighborhood density of a target item is then the total number of the target's lexical neighbors. This is the definition assumed in the current version of MiniCorp, though more complex variants exist. For example, the lexical frequency of the lexical neighbors can be included in the computation of neighborhood density, different costs can be assigned to different replacements (e.g. /p/ vs. /b/ is less "different" than /p/ vs. /d/), and/or nonadjacent neighbors can be taken into account (though more weakly the farther away from the target), as in the definition advocated by Bailey and Hahn (2001).

The main reason for phonologists to be concerned with neighborhood density is that it represents a model of exemplar-driven analogy, rather than true grammar. Thus if target A has lexical item B as a neighbor, this means only that A and B share superficial similarities, not that both form a natural class as defined by a grammar. If A ends up gaining another property associated with B, this may be an analogical effect that should be factored out of a grammatical analysis of the corpus. For example, English drive has long had the irregular past tense form drove, but only relatively recently has dive taken the irregular past tense form dove for some speakers. The fact that dive has irregularized, but not arrive, is presumably related to the fact that dive has drive as a lexical neighbor. A phonologist would therefore be justified in setting aside these facts as irrelevant to English grammar.

The confound between analogy and grammar also plays a role in the interpretation of acceptability judgments, since it is known that judgments of word-sized nonlexical items (also known as wordlikeness judgments) are affected by neighborhood density (e.g. Bailey and Hahn 2001). This is why studies testing hypotheses about phonological grammar tend to control for neighborhood density, among other factors (e.g. Frisch and Zawaydeh 2001).

MiniCorp makes it possible to compute neighborhood densities for the materials planned for a MiniJudge experiment. The user can then try to adjust the materials so that neighborhood densities are better controlled (i.e. not confounded with the experimental factors), or can can keep the original materials but include the neighborhood densities themselves in the statistical analysis, so that their effects on judgments can be distinguished from those due to the theoretically interesting grammatical factors.

How does MiniCorp use statistics?

(For more information on hypothesis testing in statistics, see the MiniJudge info page here.)

The current version of MiniCorp is primarily designed to test statistical hypotheses relating to standard OT grammars, that is, sets of constraints with a fixed, strict ranking. MiniCorp then computes the probability p that the corpus data conforms to the predictions of each constraint, and their ranking, merely by chance. If the associated p values go below a certain threshold (traditionally 0.05 = 1/20), the constraints and/or their ranking are considered to be statistically significant given the data set. If the results are significant, the researcher (and more importantly, his or her critics) can be more confident that the claimed grammar has some empirical validity. If the results are not significant, this may mean that the analysis is wrong, but it may merely mean that this particular data set and these particular statistical methods were not able to show that it's right ("not show true" is not the same as "show not true").

To run the statistical analyses, MiniCorp takes advantage of the mathematical property of OT noted above, namely that constraint ranking is a special case of constraint weighting (Prince and Smolensky 2004). To see this, consider the three generic OT tableaux in (6), (7), (8), for the grammar C₁ >> C₂. The equations on the right of each tableau convert the * marks into numerical values (* = 1, blank = 0) and convert the constraints into weights (C₁ = 10, C₂ = 1). For each tableau, the equation with the lowest value indicates the optimal candidate.

(6)

In_x C₁ C₂ Evaluations

Out_x1 * 10 x 0 + 1 x 1 = 1

Out_x2 * 10 x 1 + 1 x 0 = 10

(7)

In_z C₁ C₂ Evaluations

Out_z1 * 10 x 0 + 1 x 1 = 1

Out_z2 10 x 0 + 1 x 0 = 0

(8)

In_y C₁ C₂ Evaluations

Out_y1 * * 10 x 1 + 1 x 1 = 11

Out_y2 * 10 x 1 + 1 x 0 = 10

The main difference between OT ranking and weights is that OT ranking is strict, so that lower-ranked constraints cannot override higher-ranked ones. For example, if A >> B, then if A favors one candidate over all others, there is no way that it can lose, even if B very strongly disfavors it. Similarly, if A >> B >> C, the lower-ranked B and C cannot "gang up" to choose a candidate that A disfavors. A weighting scheme that has this effect is called a utility function (Prince 2007).

To understand how MiniCorp tests the statistical reliability of an OT grammar, first notice that the higher-ranked constraint is associated with a larger weight (here, 10 vs. 1). Constraints that play no role in a language would have very low weights. Though this property is an inherent property of an OT grammar as a formal object, it is also conveniently related to the reliability of the constraints as descriptions of corpus data. That is, the less frequently the constraint is violated in the corpus, the greater its weight. This idea is the key to all of the OT learning models listed above, where the learner goes through the corpus and increases the weights (expressed in different ways depending on the model) for obeyed constraints and decreases the weights for violated constraints.

Constraint weights have also long been associated with statistical models. Goldwater and Johnson (2003), Potts et al. (2007), and Hayes and Wilson (to appear) take advantage of this by formalizing their learners in terms of loglinear models. A loglinear model is a type of regression equation designed to handle categorical data, where observations involve binary choices or counts, rather than continuous values. Loglinear models are appropriate here since both MiniCorp and the OT learning models deal with counts of corpus items (i.e. the number that obey a constraint vs. the number that violate it). The name "loglinear" comes from the fact that the equations are linear (additive), like those in (6)-(8), so that they are easy to work with, but they only get this form by transforming originally more complex equations by taking the logarithm of both sides (among other things, this converts products into sums). Well-established methods exist for automatically finding the weights for loglinear models that allow them to best fit the data.

MiniCorp uses a type of loglinear model called Poisson regression, designed to test statistical hypotheses about count data. It gets its name from the Poisson distribution, the shape taken by random samples of counts (unlike the more famous normal distribution, it's not symmetrical, because counts can't go below 0 and because lower counts tend to be more common than higher ones).

MiniCorp first classifies corpus items in terms of all possible combinations of constraint violations. In the case of the Pazih data, the classification is as shown in the table of counts in (9). Note that AgreeV is violated 12 times, and IdentIO is always obeyed.

(9)

Counts IdentIO AgreeV

33 0 0

12 0 1

0 1 0

0 1 1

Poisson regression can automatically find the best-fitting weights for this data set... almost. First we must make some sort of adjustment, since the never-violated constraint IdentIO poses a counter-intuitive computational problem. Namely, the algorithm computing Poisson regression (like that of its cousins logistic regression and GLMM, described here), crashes if there is any perfect correlation between dependent and independent variables (as here, where IdentIO = 1 perfectly predicts that Counts = 0). The current version of MiniCorp solves the problem by brute force, by changing all 0 counts to 1. This slightly weakens the power of the test (especially if any actual count is already 1), but allows the algorithm to run.

Like the loglinear models used in other OT programs, Poisson regression tests the independent contribution of each constraint separately. Since in MiniCorp, constraint violations are marked with positive numbers, a valid constraint should be associated with a negative weight: item types that violate posited constraints should be rarer in the corpus. The main advantage of Poisson regression is that it also allows us to test the statistical reliability of each constraint. Thus with each constraint weight, we also get a p value.

In the case of the Pazih data set, both constraints end up with statistically significant negative weights, as shown below.

MiniCorp also tests the statistical reliability of a posited constraint ranking. The simplest case involves just two constraints, C₁ >> C₂. In this case, it suffices to show that the weight for C₁ is significantly different from that of C₂. (If one or the other is not negative, or if the weight of C₁ is smaller, then the OT grammar must be wrong, and there's no point testing the ranking.) The simplest way to do this is to define two regression equations, one assuming identical weights and one not, and then comparing them on their overall fit with the data set, as in (10) and (11).

(10) Counts = wC₁ + wC₂
(11) Counts = w₁C₁ + w₂C₂

Regression equations can be compared using a standard technique called a likelihood ratio test. The "ratio" part gives a hint that the two equations must be partially overlapping in order for the test to be valid (e.g. y = x₁ + x₂ vs. y = x₁). As it turns out, equation (10) is indeed contained within equation (11), as shown by their algebraic equivalents in (12) and (13), respectively (note that the weights are fitted by the regression algorithm, so their precise values in (11) vs. (13) will be different even though the comparison of (10) vs. (11) and (12) vs. (13) are identical).

(12) Counts = w(C₁ + C₂)
(13) Counts = w₁(C₁ + C₂) + w₂(C₁ - C₂)

Since the data are categorical, the likelihood ratios involve an analysis of deviance (conceptually akin to the more familiar analysis of variance) with p values computed using a chi-square test. In the case of the Pazih data set, the hypothesized ranking IdentIO >> AgreeV turns out to be statistically significant, as shown below.

To test OT grammars with more than two constraints, we must also make sure that the lower-ranked constraints don't collectively "gang up" on the higher-ranked ones. Thus for A>>B>>C>>D, MiniCorp will test the independent claims A>>{B,C,D}, B>>{C,D}, C>>D. The logic involved in testing each of these claims is similar to that for the two-constraint case, merely requiring a bit more algebra, and larger counts tables: the counts table for n constraints has 2ⁿ rows. Users can study the R code generated by MiniCorp to see how it works, and Myers (in prep) provides further justification for the logic.

Extending the logic a bit further, Poisson regression can also be used in cases where constraints can be violated more than once (or gradiently) by a single item. MiniCorp has not yet been given this capability; see Myers (in prep) for a description of how it will work in the next update of MiniCorp.

What is R?

R is free software available at www.r-project.org. R is by far the best free statistics package available, and its power and flexibility (not to mention its freeness) have made it a worldwide standard. If you do any serious quantitative research, it's well worth owning. You can learn more about it in Crawley (2005), Johnson (forthcoming), and Baayen (forthcoming), among many other places.

The main downside to R is that it is not very easy to learn or use. It relies on a command-line interface, not the menus and dialog boxes familiar from most modern programs; this puts a serious burden on memory, especially since its online help can be frustrating. There are currently several ongoing projects to create a simpler interface for it; one of the most advanced is R Commander (Fox 2005).

Since R is difficult to use, MiniJudge does all the communication with R, both generating the necessary R code and translating R's output into an easy-to-read format.

Here are the two key R links:

When you feel psychologically prepared to download R for real, do the following:

Click on the last link above to see a list of downloading locations for R.
On this list, find the location nearest to you and click on it.
In the Download and Install R box, click the link for your operating system.

What happens next depends on your operating system:

Windows: Choose "base," then the setup program (e.g. "R-2.4.1-win32.exe").
Mac: Download the installer appropriate for your computer.
Linux: Any tips would just hurt your hacker pride.

The specific R package used by MiniJudge for GLMM is called "lme4" (and its prerequisite package "Matrix"), authored by Douglas Bates and Deepayan Sarkar, and maintained by Douglas Bates. The R code generated by MiniJudge will guide you through the downloading and installation of these packages. They are still undergoing modification, so it is possible that in the future MiniJudge will need to be modified to interface properly with it. To get updated versions of your installed R packages, choose the "Update packages..." option in R's "Packages" menu, or paste in the following R code:

update.packages()

What do the statistical results mean?

MiniCorp translates R's output into an easier-to-read format, but it also saves a more detailed report in a file. Here is a line-by-line interpretation of a typical report, based on an analysis using the Pazih corpus file pazihIPA.rtf to test the grammar IdentIO >> AgreeV.

R OUTPUT	EXPLANATION
Hypothesized grammar: IdentIO >> AgreeV	Header.
Constraint test (Poisson regression):	Tests individual contributions of each constraint.
Call: glm(formula = Count ~ IdentIO.c + AgreeV.c, family = poisson)	Poisson regression equation (".c" added to represent counts rather than raw data)
Deviance Residuals: 1 2 3 4 0.07813 -0.12742 -0.39357 0.53897	Measure the difference between the predicted and actual values. Here there are four, one for each row in the counts table. For larger tables, the deviance residuals are summarized in terms of quartiles. Residuals should range randomly around zero, which seems true here.
Coefficients:	Constraint weights.
Estimate Std. Error z value Pr(>\|z\|) (Intercept) 3.4829 0.1742 19.990 < 2e-16 * IdentIO.c -3.1135 0.7226 -4.309 1.64e-05 * AgreeV.c -0.9614 0.3261 -2.948 0.00320 **	Intercept: Baseline skew of data; ignored by MiniCorp. Estimate: Constraint weights; should all be negative, as here. Pr(>\|z\|): p value (two-tailed). z value: measures how far coefficient is from chance expectation of 0. Std. Error: standard error (used to compute z value). (NOTE: 2e-16 = (2/10)¹⁶, a very small number.)
Signif. codes: 0 '*' 0.001 '' 0.01 '*' 0.05 '.' 0.1 ' ' 1	Classification of p values (MiniCorp only checks if p < .05).
(Dispersion parameter for poisson family taken to be 1)	By default R assumes the simplest type of Poisson distribution.
Null deviance: 58.80505 on 3 degrees of freedom Residual deviance: 0.46772 on 1 degrees of freedom AIC: 20.144	More information on fitting. Residual deviance should be low and AIC should be high.
Number of Fisher Scoring iterations: 4	Low value like this means weights were found quickly; a high value hints at a possible problem.
Ranking tests (analysis of deviance with chi-square tests):	Tests constraint rankings.
Analysis of Deviance Table	Summary statistics for likelihood ratio test on categorical data
Model 1: Count ~ I(IdentIO.c + AgreeV.c) Model 2: Count ~ IdentIO.c + AgreeV.c	Model 1: Assumes identical constraint weights Model 2: Doesn't assume identical constraint weights
Resid. Df Resid. Dev Df Deviance P(>\|Chi\|) 1 2 11.0185 2 1 0.4677 1 10.5508 0.0012	Resid. Df: indicates model complexity. Resid. Dev: smaller deviance means complex model is better. Df & Deviance: Used to compute p value. P(>\|Chi\|): p value for model comparison. Here p < 0.05, so ranking matters.
Constraint test: Constraints Weights p IdentIO -3.1135 0 * AgreeV -0.9614 0.0032 * (* significant constraint) Ranking test: Constraints p IdentIO 0.0012 * (* significant ranking)	This simplified summary, generated by MiniCorp, is intended to be mostly self-explanatory. The weights and p values are taken from R's technical output shown above. Ranking tests are summarized with the top constraint of each subhierarchy. Thus for the grammar A>>B>>C, "A" in the summary refers to the test of A>>{B,C} and "B" refers to the test of B>>C.