When we analyze a single poem, the basic unit is fairly easy to picture. VerseVAD identifies the words that are eligible for a particular measure, applies the selected scope and weighting rules, and summarizes what it finds. At the corpus level, things get more interesting because we are no longer asking only about words. We are also asking how individual poems should contribute to the larger result. That is where token-pooled and equal-work outputs come in.
Both are legitimate ways to summarize a corpus. They simply answer different questions. One gives more influence to works that contribute more eligible words. The other gives every work the same influence, regardless of length. A surprisingly useful way to think about the difference is through the U.S. Congress.
The House and Senate Analogy
In the U.S. House of Representatives, states do not all have the same number of representatives. Larger states receive more seats because representation is tied to population. California contributes more members than Wyoming because California has far more people.
Token-pooled corpus analysis works in a similar way. A longer poem contributes more eligible observations to the corpus than a shorter poem. If one poem contributes 1,000 matched lexical tokens and another contributes 50, those 1,000 observations are part of the same pool. The longer poem therefore has more influence on the corpus-level result.
Equal-work analysis is closer to the Senate. Every state gets two senators regardless of population. California and Wyoming enter the Senate on equal footing.In an equal-work corpus summary, every poem gets one vote. VerseVAD first calculates the relevant result for each individual work, then averages those work-level results. A twenty-line poem and a two-hundred-line poem therefore contribute equally to the final corpus mean.
Neither system is inherently more correct. They represent two different ideas about what the corpus is.
What Token-Pooled Results Tell Us
A token-pooled result asks something like:
What does the language of this corpus look like when all eligible observations are considered together?
Suppose we have three poems containing 100, 500, and 1,000 eligible tokens. A token-pooled mean treats those 1,600 observations as one larger body of evidence. The third poem contributes much more to the result because it contributes much more language. That can be exactly what we want.
If we are interested in the overall lexical environment of a collection, token pooling gives us a natural way to describe it. A word that occurs repeatedly across a long poem really does occupy more of the corpus. If a poet devotes half a book to one very long work, it may be reasonable for that work to exert more influence on a description of the language found across the volume.
The important thing is to understand what the number represents. It is not the average poem. It is the average eligible observation across the corpus.
What Equal-Work Results Tell Us
Equal-work analysis asks a different question:
What does the typical work in this corpus look like when every work contributes equally?
VerseVAD calculates the relevant metric for each poem first. Those poem-level values are then averaged without giving longer poems additional weight. Imagine a corpus containing one enormous narrative poem and twenty short lyrics. In a token-pooled analysis, the narrative poem might dominate simply because it contains most of the words. In an equal-work analysis, it contributes one result alongside the twenty lyrics.
That makes equal-work summaries especially useful when the unit we care about is the poem itself.
If we are comparing poets, collections, periods, or groups of works, we may want to know what an average poem in each corpus looks like rather than what the pooled vocabulary looks like. Equal-work analysis gives us that perspective.
The Token Pool Depends on the Scope
There is one important detail hidden inside the phrase token pool: not every analysis uses the same pool of tokens.
VerseVAD allows lexical analyses to operate under different scopes. A corpus might be analyzed using all lexical tokens, stopwords excluded, or content words only. That means the token pool changes with the scope. If the selected scope is all lexical tokens, the pool contains every eligible lexical token under that analysis. If stopwords are excluded, those observations are removed from the pool. If the analysis is restricted to content words, the pool becomes narrower still.
So an all-lexical-token pool will normally be larger than a stopword-excluded pool, and a stopword-excluded pool will normally be larger than a content-word-only pool. This matters because “token-pooled” does not mean “every word in every poem no matter what.” It means every eligible observation under the currently selected analytical scope.
If we change the scope, we change the population being pooled.
A Simple Example
Imagine a tiny corpus with two poems.
Poem A is long and contributes 900 eligible content-word tokens. Its mean valence is 0.60.
Poem B is short and contributes 100 eligible content-word tokens. Its mean valence is 0.40.
Under an equal-work approach, both poems contribute equally:
0.60 and 0.40 produce an equal-work mean of 0.50.
Under a token-pooled approach, Poem A contributes nine times as many eligible observations as Poem B. The result will therefore sit much closer to 0.60 than to 0.40.
That difference is not a contradiction. It tells us something important about the corpus.
The longer poem has a different lexical profile from the shorter one, and it also accounts for most of the language in the collection.
When the Two Results Agree
Sometimes token-pooled and equal-work results are nearly identical. That can happen when the poems are similar in length, when their metric values are broadly similar, or when length and the measured characteristic are not strongly related in the corpus. This kind of agreement is useful. It suggests that the corpus-level story is relatively stable regardless of whether we think in terms of total language or equal poem-level contribution.
If both methods tell us that a corpus has roughly the same average valence, concreteness, frequency, or other characteristic, then our interpretation does not depend very much on how individual works are weighted. That does not make the result more “true,” but it does make it less sensitive to this particular methodological choice.
When the Two Results Differ
The more interesting situation is often when they do not agree. A meaningful difference between token-pooled and equal-work results tells us that poem length and the measured characteristic are interacting somehow.
Perhaps the longest poems in the corpus use less familiar vocabulary. Perhaps the short poems are unusually concrete. Perhaps one very long work has much higher arousal than the rest of the collection. Perhaps a handful of long narrative poems are pulling the token pool in one direction while many shorter lyrics cluster somewhere else.
That discrepancy is not something to average away. It is information.
The House and Senate analogy helps here. If the House and Senate produce different pictures of the country, part of the reason may be that they represent the population differently. One emphasizes population size. The other emphasizes states as equal units.
Token-pooled and equal-work analysis do something similar for a corpus. One emphasizes how much eligible language each work contributes. The other emphasizes the works themselves as equal units.
When they diverge, VerseVAD is effectively telling us, the size of the works matters to this result. That is a very useful thing to know.
Neither View Needs to Win
It can be tempting to ask which result should be reported as the “real” corpus mean. Usually, that is the wrong question.
If the research question concerns the language distributed throughout the entire corpus, token pooling may be the more natural view. If the question concerns the typical poem or equal comparison among works, equal-work may be more informative.
In many cases, showing both is better than choosing one and hiding the other. The two values provide context for each other. Agreement suggests stability. Difference reveals structure inside the corpus that a single mean could conceal. That is why VerseVAD treats them as complementary outputs rather than competing answers.
What to Look for in Practice
When reading a corpus report, it helps to ask a few simple questions. Are the poems roughly similar in length, or does the corpus contain some unusually long works? Are the token-pooled and equal-work values close together? If they differ, which poems are contributing the most eligible tokens? Do those longer works have unusually high or low values for the metric being examined?
It is also worth checking the active scope. A content-word-only token pool may tell a somewhat different story from an all-lexical-token pool because it is summarizing a different set of observations.
Those questions turn what might look like two competing numbers into a small map of the corpus.
Two Views of the Same Collection
The point of offering token-pooled and equal-work outputs is not to make corpus analysis more complicated than it needs to be. It is to avoid hiding an important choice.
A corpus can be understood as a large collection of language, in which case longer works naturally contribute more evidence. It can also be understood as a collection of individual works, in which case each poem may deserve an equal voice. VerseVAD shows both because sometimes the difference between those perspectives matters.
Token-pooled analysis gives the longer works more representatives. Equal-work analysis gives every poem two senators. And when those two chambers agree, we have a fairly stable picture. When they do not, that disagreement is not a nuisance. It is an invitation to look more closely at how the corpus is built.