Category Archives: math

Do humanists get their ideas from anything at all?

Posted on by
11

My reaction to Stanley Fish’s third column on digital humanities was at first so negative that I thought it not worth writing about. But in the light of morning, there is something here worth discussing. Fish raises a neglected issue that I (and a bunch of other people cited at the end of this post) have been trying to foreground: the role of discovery in the humanities. He raises the issue symptomatically, by suppressing it, but the problem is too important to let that slide.

Fish argues, in essence, that digital humanists let the data suggest hypotheses for them instead of framing hypotheses that are then tested against evidence.

The usual way of doing this is illustrated by my example: I began with a substantive interpretive proposition … and, within the guiding light, indeed searchlight, of that proposition I noticed a pattern that could, I thought be correlated with it. I then elaborated the correlation.

The direction of my inferences is critical: first the interpretive hypothesis and then the formal pattern, which attains the status of noticeability only because an interpretation already in place is picking it out.

The direction is the reverse in the digital humanities: first you run the numbers, and then you see if they prompt an interpretive hypothesis. The method, if it can be called that, is dictated by the capability of the tool.

The underlying element of truth here is that all researchers — humanists and scientists alike — do need to separate the process of discovering a hypothesis from the process of testing it. Otherwise you run into what we unreflecting empiricists call “the problem of data dredging.” If you simply sweep a net through an ocean of data, and frame a conclusion based on whatever you catch, you’re not properly testing anything, because you’re implicitly testing an infinite number of hypotheses that are left unstated — and the significance of any single test is reduced when it’s run as part of a large battery.

That’s true, but it’s also a problem that people who do data mining are quite self-conscious about. It’s why I never stop linking to this xkcd comic about “significance.” And it’s why Matt Wilkens (mistargeted by Fish as an emblem of this interpretive sin) goes through a deliberately iterative process of first framing hypotheses about nineteenth-century geographical imagination and then testing them. (For instance, after noticing that certain states seem especially prominent in 19c American fiction, he tests whether this remains true after you compensate for differences in population size, and then proposes a pair of hypotheses that he suggests will need to be evaluated against additional “test cases.”)

Wiliam Blake, "Satan, Sin, and Death"


More importantly, Fish profoundly misrepresents his own (traditional) interpretive procedure by pretending that the act of interpretation is wholly contained in a single encounter with evidence. On his account we normally begin with a hypothesis (which seems to have sprung, like Sin, fully-formed from our head), and test it against a single sentence.

In reality, of course, our “interpretive proposition” is often suggested by the same evidence that confirms it. Or — more commonly — we derive a hypothesis from one example, and then read patiently through dozens of books until we have gathered enough confirming evidence to write a chapter. This process runs into a different interpretive fallacy: if you keep testing a hypothesis until you’ve confirmed it, you’re not testing it at all. And it’s a bit worse than that, because in practice what we do now is go to a full-text search engine and search for terms that would go together if our assumptions were correct. (In the example Fish offers, this might be “bishops” and “presbyters.”) If you find three sentences where those terms coincide, you’ve got more than enough evidence to prop up an argument, using our richly humanistic (cough, anecdotal) conception of evidence. And of course a full-text search engine can find you three examples of just about anything. But we don’t have to worry about this, because search engines are not tools that dictate a method; they are transparent extensions of our interpretive sensibility.

The basic mistake that Fish is making is this: he pretends that humanists have no discovery process at all. For Fish, the interpretive act is always fully contained in an encounter with a single piece of evidence. How your “interpretive proposition” got framed in the first place is a matter of no consequence: some readers are just fortunate to have propositions that turn out to be correct. Fish is not alone in this idealized model of interpretation; it’s widespread among humanists.

Fish is resisting the assistance of digital techniques, not because they would impose scientism on the humanities, but because they would force us to acknowledge that our ideas do after all come from somewhere — whether a search engine or a commonplace book. But as Peter Stallybrass eloquently argued five years ago in PMLA (h/t Mark Sample) the process of discovery has always been collaborative, and has long — at least since early modernity — been embodied in specific textual technologies.

References
Stallybrass, Peter. “Against Thinking.” PMLA 122.5 (2007): 1580-1587.
Wilkens, Matthew. “Geolocation Extraction and Mapping of Nineteenth-Century U.S. Fiction.” DHCS 2011.
On the process of embodied play that generates ideas, see also Stephen Ramsay’s book Reading Machines (University of Illinois Press, 2011).

A brief outburst about numbers.

Posted on by
10

In responding to Stanley Fish last week, I tried to acknowledge that the “digital humanities,” in spite of their name, are not centrally about numbers. The movement is very broad, and at the broadest level, it probably has more to do with networked communication than it does with quantitative analysis.

The older tradition of “humanities computing” — which was about numbers — has been absorbed into this larger movement. But it’s definitely the part of DH that humanists are least comfortable with, and it often has to apologize for itself. So, for instance, I’ve spent much of the last year reminding humanists that they’re already using quantitative text mining in the form of search engines — so it can’t be that scary.* Kathleen Fitzpatrick recently wrote a post suggesting that “one key role for a ‘worldly’ digital humanities may well be helping to break contemporary US culture of its unthinking association of numbers with verifiable reality….” Stephen Ramsay’s Reading Machines manages to call for an “algorithmic criticism” while at the same time suggesting that humanists will use numbers in ways that are altogether different from the way scientists use them (or at least different from “scientism,” an admittedly ambiguous term).

I think all three of us (Stephen, Kathleen, and myself) are making strategically necessary moves. Because if you tell humanists that we do (also) need to use numbers the way scientists use them, your colleagues are going to mutter about naïve quests for certainty, shake their heads, and stop listening. So digital humanists are rhetorically required to construct positivist scapegoats who get hypothetically chased from our villages before we can tell people about the exciting new kinds of analysis that are becoming possible. And, to be clear, I think the people I’ve cited (including me) are doing that in fair and responsible ways.

However, I’m in an “eppur si muove” mood this morning, so I’m going to forget strategy for a second and call things the way I see them. <Begin Galilean outburst>

In reality, scientists are not naïve about the relationship between numbers and certainty, because they spend a lot of time thinking about statistics. Statistics is the science of uncertainty, and it insists — as forcefully as any literary theorist could — that every claim comes accompanied by a specific kind of ignorance. Once you accept that, you can stop looking for absolute knowledge, and instead reason concretely about your own relative uncertainty in a given instance. I think humanists’ unfamiliarity with this idea may explain why our critiques of data mining so often taken the form of pointing to a small error buried somewhere in the data: unfamiliarity with statistics forces us to fall back on a black-and-white model of truth, where the introduction of any uncertainty vitiates everything.

Moreover, the branch of statistics most relevant to text mining (Bayesian inference) is amazingly, almost bizarrely willing to incorporate subjective belief into its definition of knowledge. It insists that definitions of probability have to depend not only on observed evidence, but on the “prior probabilities” that we expected before we saw the evidence. If humanists were more familiar with Bayesian statistics, I think it would blow a lot of minds.

I know the line about “lies, damn lies, and so on,” and it’s certainly true that statistics can be abused, as this classic xkcd comic shows. But everything can be abused. The remedy for bad verbal argument is not to “remember that speech should stay in its proper sphere” — it’s to speak better and more critically. Similarly, the remedy for bad quantitative argument is not “remember that numbers have to stay in their proper sphere”; it’s to learn statistics and reason more critically.

possible shapes of the Beta distribution, from Wikpedia


None of this is to say that we can simply borrow tools or methods from scientists unchanged. The humanities have a lot to add — especially when it comes to the social and historical character of human behavior. I think there are fascinating advances taking place in data science right now. But when you take apart the analytic tools that computer scientists have designed, you often find that they’re based on specific mistaken assumptions about the social character of language. For instance, there’s a method called “Topics over Time” that I want to use to identify trends in the written record (Wang and McCallum, 2006). The people who designed it have done really impressive work. But if a humanist takes apart the algorithm underlying this method, they will find that it assumes that every trend can be characterized as a smooth curve called a “Beta distribution.” Whereas in fact, humanists have evidence that the historical trajectory of a topic is often more complex than that, in ways that really matter. So before I can use this tool, I’m going to have to fix that part of the method.

The diachronic behavior a topic can actually exhibit.


But this is a problem that can be fixed, in large part, by fixing the numbers. Humanists have a real contribution to make to the science of data mining, but it’s a contribution that can be embodied in specific analytic insights: it’s not just to hover over the field like the ghost of Ben Kenobi and warn it about hubris.

</Galilean outburst>

For related thoughts, somewhat more temperate than the outburst above, see this excellent comment by Matthew Wilkens, responding to a critique of his work by Jeremy Rosen.

* I credit Ben Schmidt for this insight so often that regular readers are probably bored. But for the record: it comes from him.

Exploring the relationship between topics and trends.

Posted on by
Reply

I’ve been talking about correlation since I started this blog. Actually, that was the reason why I did start it: I think literary scholars can get a huge amount of heuristic leverage out of the fact that thematically and socially-related words tend to rise and fall together. It’s a simple observation, and one that stares you in the face as soon as you start to graph word frequencies on the time axis.1 But it happens to be useful for literary historians, because it tends to uncover topics that also pose periodizable kinds of puzzles. Sometimes the puzzle takes the form of a topic we intuitively recognize (say, the concept of “color”) that increases or decreases in prominence for reasons that remain to be explained:

At other times, the connection between elements of the topic is not immediately intuitive, but the terms are related closely enough that their correlation suggests a pattern worthy of further exploration. The relationship between terms may be broadly historical:

Or it may involve a pattern of expression that characterizes a periodizable style:

Of course, as the semantic relationship between terms becomes less intuitively obvious, scholars are going to wonder whether they’re looking at a real connection or merely an accidental correlation. “Ardent” and “tranquil” seem like opposites; can they really be related as elements of a single discourse? And what’s the relationship to “bosom,” anyway?

Ultimately, questions like this have to be addressed on a case-by-case basis; the significance of the lead has to be fleshed out both with further analysis, and with close reading.

But scholars who are wondering about the heuristic value of correlation may be reassured to know that this sort of lead does generally tend to pan out. Words that correlate with each other across the time axis do in practice tend to appear in the same kinds of volumes. For instance, if you randomly select pairs of words from the top 10,000 words in the Google English ngrams dataset 1700-1849,2 measure their correlation with each other in that dataset across the period 1700-1849, and then measure their tendency to appear in the same volumes in a different collection3 (taking the cosine similarity of term vectors in a term-document matrix), the different measures of association correlate with each other strongly. (Pearson’s r is 0.265, significant at p < 0.0005.) Moreover, the relationship holds (less strongly, but still significantly) even in adjacent centuries: words that appear in the same eighteenth-century volumes still tend to rise and fall together in the nineteenth century.

Why should humanists care about the statistical relationship between two measures of association? It means that correlation-mining is in general going to be a useful way of identifying periodizable discourses. If you find a group of words that correlate with each other strongly, and that seem related at first glance, it's probably going to be worthwhile to follow up the hunch. You’re probably looking at a discourse that is bound together both diachronically (in the sense that the terms rise and fall together) and topically (in the sense that they tend to appear in the same kinds of volumes).

Ultimately, literary historians are going to want to assess correlation within different genres; a dataset like Google's, which mixes all genres in a single pool, is not going to be an ideal tool. However, this is also a domain where size matters, and in that respect, at the moment, the ngrams dataset is very helpful. It becomes even more helpful if you correct some of the errors that vitiate it in the period before 1820. A team of researchers at Illinois and Stanford4, supported by the Andrew W. Mellon Foundation, has been doing that over the course of the last year, and we're now able to make an early version of the tool available on the web. Right now, this ngram viewer only covers the period 1700-1899, but we hope it will be useful for researchers in that period, because it has mostly corrected the long-s problem that confufes opt1cal charader readers in the 18c — as well as a host of other, less notorious problems. Moreover, it allows researchers to mine correlations in the top 10,000 words of the lexicon, instead of trying words one by one to see whether an interesting pattern emerges. In the near future, we hope to expand the correlation miner to cover the twentieth century as well.

For further discussion of the statistical relationship between topics and trends, see this paper submitted to DHCS 2011.

UPDATE Nov 22, 2011: At DHCS 2011, Travis Brown pointed out to me that Topics Over Time (Wang and McCallum) might mine very similar patterns in a more elegant, generative way. I hope to find a way to test that method, and may perhaps try to build an implementation for it myself.

References
1) Ryan Heuser and I both noticed this pattern last winter. Ryan and Long Le-Khac presented on a related topic at DH2011: Heuser, Ryan, and Le-Khac, Long. “Abstract Values in the 19th Century British Novel: Decline and Transformation of a Semantic Field,” Digital Humanities 2011, Stanford University.

2) Jean-Baptiste Michel*, Yuan Kui Shen, Aviva Presser Aiden, Adrian Veres, Matthew K. Gray, William Brockman, The Google Books Team, Joseph P. Pickett, Dale Hoiberg, Dan Clancy, Peter Norvig, Jon Orwant, Steven Pinker, Martin A. Nowak, and Erez Lieberman Aiden*. “Quantitative Analysis of Culture Using Millions of Digitized Books.” Science (Published online ahead of print: 12/16/2010)

3) The collection of 3134 documents (1700-1849) I used for this calculation was produced by combining ECCO-TCP volumes with nineteenth-century volumes selected and digitized by Jordan Sellers.

4) The SEASR Correlation Analysis and Ngrams Viewer was developed by Loretta Auvil and Boris Capitanu at the Illinois Informatics Institute, modeled on prototypes built by Ted Underwood, University of Illinois, and Ryan Heuser, Stanford.

Identifying diction that characterizes an author or genre: why Dunning’s may not be the best method.

Posted on by
8

Most of what I’m about to say is directly lifted from articles in corpus linguistics (1, 2), but I don’t think these results have been widely absorbed yet by people working in digital humanities, so I thought it might be worthwhile to share them, while demonstrating their relevance to literary topics.

The basic question is just this: if I want to know what words or phrases characterize an author or genre, how do I find out? As Ben Schmidt has shown in an elegantly visual way, simple mathematical operations won’t work. If you compare ratios (dividing word frequencies in the genre A that interests you by the frequencies in a corpus B used as a point of comparison), you’ll get a list of very rare words. But if you compare the absolute magnitude of the difference between frequencies (subtracting B from A), you’ll get a list of very common words. So the standard algorithm that people use is Dunning’s log likelihood,

— a formula that incorporates both absolute magnitude (O is the observed frequency) and a ratio (O/E is the observed frequency divided by the frequency you would expect). For a more complete account of how this is calculated, see Wordhoard.

But there’s a problem with this measure, as Adam Kilgarriff has pointed out (1, pp. 237-38, 247-48). A word can be common in a corpus because it’s very common in one or two works. For instance, when I characterize early-nineteenth-century poetic diction (1800-1849) by comparing a corpus of 60 volumes of poetry to a corpus of fiction, drama, and nonfiction prose from the same period (3), I get this list:

Much of this looks like “poetic diction” — but “canto” is poetic diction only in a weird sense. It happens to be very common in a few works of poetry that are divided into cantos (works for instance by Lord Byron and Walter Scott). So when everything is added up, yes, it’s more common in poetry — but it doesn’t broadly characterize the corpus. Similar problems occur for a range of other reasons (proper nouns and pronouns can be extremely common in a restricted context).

The solution Kilgarriff offers is to instead use a Mann-Whitney ranks test. This allows us to assess how consistently a given term is more common in one corpus than in another. For instance, suppose I have eight text samples of equal length. Four of them are poetry, and four are prose. I want to know whether “lamb” is significantly more common in the poetry corpus than in prose. A simple form of the Mann-Whitney test would rank these eight samples by the frequency of “lamb” and then add up their respective ranks:

Since most works of poetry “beat” most works of prose in this ranking, the sum of ranks for poetry is higher, in spite of the 31 occurrences of lamb in one work of prose — which is, let us imagine, a novel about sheep-rustling in the Highlands. But a log-likelihood test would have identified this word as more common in prose.

In reality, one never has “equal-sized” documents, but the test is not significantly distorted if one simply replaces absolute frequency with relative frequency (normalized for document size). (If one corpus has on average much smaller documents than the other does, there may admittedly be a slight distortion.) Since the number of documents in each corpus is also going to vary, it’s useful to replace the rank-sum (U) with a statistic ρ (Mann-Whitney rho) that is U, divided by the product of the sizes of the two corpora.

Using this measure of over-representation in a corpus produces a significantly different model of “poetic diction”:

This looks at first glance like a better model. It demotes oddities like “canto,” but also slightly demotes pronouns like “thou” and “his,” which may be very common in some works of poetry but not others. In general, it gives less weight to raw frequency, and more weight to the relative ubiquity of a term in different corpora. Kilgarriff argues that the Mann-Whitney test thereby does a better job of identifying the words that characterize male and female conversation (1, pp. 247-48).

On the other hand, Paul Rayson has argued that by reducing frequency to a rank measure, this approach discards “most of the evidence we have about the distribution of words” (2). For linguists, this poses an interesting, principled dilemma, where two statistically incompatible definitions of “distinctive diction” are pitted against each other. But for a shameless literary hack like myself, it’s no trouble to cut the Gordian knot with an improvised algorithm that combines both measures. For instance, one could multiply rho by the log of Dunning’s log likelihood (represented here as G-squared) …

I don’t yet know how well this algorithm will perform if used for classification or authorship attribution. But it does produce what is for me an entirely convincing portrait of early-nineteenth-century poetic diction:

Of course, once you have an algorithm that convincingly identifies the characteristic diction of a particular genre relative to other publications in the same period, it becomes possible to say how the distinctive diction of a genre is transformed by the passage of time. That’s what I hope to address in my next post.

UPDATE Nov 10, 2011: As I continue to use these tests in different ways (using them e.g. to identify distinctively “fictional” diction, and to compare corpora separated by time) I’m finding the Mann-Whitney ρ measure more and more useful on its own. I think my urge to multiply it by Dunning’s log-likelihood may have been the needless caution of someone who’s using an unfamiliar metric and isn’t sure yet whether it will work unassisted.

References
(1) Adam Kilgarriff, “Comparing Corpora,” International Journal of Corpus Linguistics 6.1 (2001): 97-133.
(2) Paul Rayson, Matrix: A Statistical Method and Software Tool for Linguistic Analysis through Corpus Comparison. Unpublished Ph.D thesis, Lancaster University, 2003, p. 47. Cited in Magali Paquot and Yves Bestgen, “Distinctive words in academic writing: A comparison of three statistical tests for keyword extraction,” Corpora: Pragmatics and Discourse Papers from the 29th International Conference on English Language Research on Computerized Corpora (ICAME 29), Ascona, Switzerland, 14-18 May 2008, p. 254.
(3) The corpora used in this post were selected by Jordan Sellers, mostly from texts available in the Internet Archive, and corrected with a Python script described in this post.

LSA is a marvellous tool, but literary historians may want to customize it for their own discipline.

Posted on by
4

Right now Latent Semantic Analysis is the analytical tool I’m finding most useful. By measuring the strength of association between words or groups of words, LSA allows a literary historian to map themes, discourses, and varieties of diction in a given period.

This approach, more than any other I’ve tried, turns up leads that are useful for me as a literary scholar. But when I talk to other people in digital humanities, I rarely hear enthusiasm for it. Why doesn’t LSA get more love? I see three reasons:

1. The word “semantic” is a false lead: it points away from the part of this technique that would actually interest us. It’s true that Latent Semantic Analysis is based on the observation that a word’s distribution across a collection of documents works remarkably well as a first approximation of its meaning. A program running LSA can identify English synonyms on the TOEFL as well as the average student applying to college from a non-English-speaking country. [1]

But for a literary historian, the value of this technique does not depend on its claim to identify synonyms and antonyms. We may actually be more interested in contingent associations (e.g., “sensibility” — “rousseau” in the list on the left) than we are in the core “meaning” of a word.

I’ll return in a moment to this point. It has important implications, because it means that we want LSA to do something slightly different than linguists and information scientists have designed it to do. The “flaws” they have tried to iron out of the technique may not always be flaws for our purposes.

2. People who do topic-modeling may feel that they should use more-recently-developed Bayesian methods, which are supposed to be superior on theoretical grounds. I’m acknowledging this point just to set it aside; I’ve mused out loud about it once already, and I don’t want to do more musing until I have rigorously compared the two methods. I will say that from the perspective of someone just getting started, LSA is easier to implement than Bayesian topic modeling: it runs faster and scales up more easily.

3. The LSA algorithm provided by an off-the-shelf package is not necessarily the best algorithm for a literary historian. At bottom, that’s why I’m writing this post: humanists who want to use LSA are going to need guidance from people in their own discipline. Computer scientists do acknowledge that LSA requires “tuning, which is viewed as a kind of art.” [2] But they also offer advice about “best practices,” and some of those best practices are defined by disciplinary goals that humanists don’t share.

For instance, the power of LSA is often said to come from “reducing the dimensionality of the matrix.” The matrix in question is a term-document matrix — documents are listed along one side of the matrix, and terms along the other, and each cell of the matrix (tfi,j) records the number of times term i appears in document j, modified by a weighting algorithm described at the end of this post.

A (very small) term-document matrix.


That term-document matrix in and of itself can tell you a lot about the associations between words; all you have to do is measure the similarity between the vectors (columns of numbers) associated with each term. But associations of this kind won’t always reveal synonyms. For instance, “gas” and “petrol” might seem unrelated, because they substitute for each other in different sociolects and are rarely found together. To address that problem, you can condense the matrix by factorizing it with a technique called singular value decomposition (SVD). I’m not going to get into the math here, but the key is that condensing the matrix partially fuses related rows and columns — and as a result, the compressed matrix is able to measure transitive kinds of association. The words “gas” and “petrol” may rarely appear together. But they both appear with the same kinds of other words. So when dimensionality reduction “merges” the rows representing similar documents, “gas” and “petrol” will end up being strongly represented in the same merged rows. A compressed matrix is better at identifying synonyms, and for that reason at information retrieval. So there is a lot of consensus among linguists and information scientists that reducing the number of dimensions in the matrix is a good idea.

But literary historians approach this technique with a different set of goals. We care a lot about differences of sociolect and register, and may even be more interested in those sources of “noise” than we are in purely semantic relations. “Towering,” for instance, is semantically related to “high.” But I could look that up in a dictionary; I don’t need a computer program to tell me that! I might be more interested to discover that “towering” belongs to a particular subset of poetic diction in the eighteenth century. And that is precisely the kind of accident of distribution that dimensionality-reduction is designed to filter out. For that reason, I don’t think literary applications of LSA are always going to profit from the dimensionality-reduction step that other disciplines recommend.

For about eight months now, I’ve been using a version of LSA without dimensionality reduction. It mines associations simply by comparing the cosine-similarity of term vectors in a term-document matrix (weighted in a special way to address differences of document size). But I wanted to get a bit more clarity about the stakes of that choice, so recently I’ve been comparing it to a version of LSA that does use SVD to compress the matrix.

Comparing 18c associations for "delicacy" generated by two different algorithms.


Here’s a quick look at the results. (I’m using 2,193 18c volumes, mostly produced by TCP-ECCO; volumes that run longer than 100,000 words get broken into chunks that can range from 50k-100k words.) In many cases, the differences between LSA with and without compression are not very great. In the case of “delicacy,” for instance, both algorithms indicate that “delicate” has the strongest association. “Politeness” and “tenderness” are also very high on both lists. But compare the second row. The algorithm with compression produces “sensibility” — a close synonym. On the left-hand side, we have “woman.” This is not a synonym for “delicacy,” and if a linguist or computer scientist were evaluating these algorithms, it would probably be rejected as a mistake. But from a literary-historical point of view, it’s no mistake: the association between “delicacy” and femininity is possibly the most interesting fact about the word.

The 18c associations of "high" and "towering," in an uncompressed term-document matrix.


In short, compressing the matrix with SVD highlights semantic relationships at the cost of slightly blurring other kinds of association. In the case of “delicacy,” the effect is fairly subtle, but in other cases the difference between the two approaches is substantial. For instance, if you measure the similarity of term vectors in a matrix without compression, “high” and “towering” look entirely different. The main thing you discover about “high” is that it’s used for physical descriptions of landscape (“lies,” “hills”), and the main thing you discover about “towering” is that it’s used in poetic contexts (“flowery,” “glittering”).

The 18c. associations of "high" and "towering," as measured in a term-document matrix that has undergone SVD compression.


In a matrix that has undergone dimensionality reduction with SVD, associations have a much more semantic character, although they are still colored by other dimensions of context. Which of these two algorithms is more useful for humanistic purposes? I think the answer is going to depend on the goals being pursued in a given research project — if you’re interested in “topics” that are strictly semantic, you might want to use an algorithm that reduces dimensionality with SVD. If you’re interested in discourses, sociolects, genres, or types of diction, you might use LSA without dimensionality reduction.

My purpose here isn’t to choose between those approaches; it’s just to remind humanists that the algorithms we borrow from other disciplines are often going to need to be customized for our own disciplinary purposes. Information scientists have designed topic-modeling algorithms that produce semantically unified topics, because semantic categorization is important for them. But in literary history, we also care about other dimensions of language, and we don’t have to judge topic-modeling algorithms by strictly semantic criteria. How should we judge them? It will probably take decades for us to answer that question fully, but the short answer is just — by how well, in practice, they help us locate critically and historically interesting patterns.

A couple of technical notes: A fine point of LSA that can matter a great deal is how you weight the individual cells in the term-document matrix. For the normal LSA algorithm that uses dimensionality reduction, the consensus is that “log-entropy weighting” works well. You take the log of each frequency, and multiply the whole term vector by the entropy of the vector. I have found that this also works well for humanistic purposes.

For LSA without dimensionality reduction, I would recommend weighting cells by subtracting the expected frequency from the observed frequency. This formula “evens the playing field” between common and uncommon words — and it does so, vitally, in a way that gives a word’s absence from a long document more weight than its absence from a short one. (Much of LSA’s power actually comes from learning where a given word tends not to appear. [3]) I have tried various ways of applying log-entropy weighting without compressing the matrix, and I do not recommend it. Those two techniques belong together.

For reasons that remain somewhat mysterious (although the phenomenon itself is widely discussed), dimensionality reduction seems to work best when the number of dimensions retained is in the range of 250-350. Intuitively, it would seem possible to strike a sort of compromise between LSA methods that do and don’t compress the matrix by reducing dimensionality less drastically (perhaps only, say, cutting it by half). But in practice I find that doesn’t work very well; I suspect compression has to reach a certain threshold before the noise inherent in the process starts to cancel itself out and give way to a new sort of order.

[1] Thomas K. Landauer, Peter W. Foltz, and Darrell Latham, An Introduction to Latent Semantic Analysis, Discourse Processes 25 (1998): 259-84. Web reprint, p. 22.
[2] Preslav Nakov, Elena Valchanova, and Galia Angelova, “Towards Deeper Understanding of Latent Sematic Analysis Performance,” Recent Advances in Natural Language Processing, ed. Nicolas Nicolov (Samokov, Bulgaria: John Benjamins, 2004), 299.
[3] Landauer, Foltz, and Latham, p. 24.

Words that appear in the same 18c volumes also track each other over time, through the 19c.

Posted on by
Reply

I wrote a long post last Friday arguing that topic-modeling an 18c collection is a reliable way of discovering eighteenth- and nineteenth-century trends, even in a different collection.

But when I woke up on Saturday I realized that this result probably didn’t depend on anything special about “topic modeling.” After all, the topic-modeling process I was using was merely a form of clustering. And all the clustering process does is locate the hypothetical centers of more-or-less-evenly-sized clouds of words in vector space. It seemed likely that the specific locations of these centers didn’t matter. The key was the underlying measure of association between words — “cosine similarity in vector space,” which is a fancy way of saying “tendency to be common in the same 18c volumes.” Any group of words that were common (and uncommon) in the same 18c volumes would probably tend to track each other over time, even into the next century.

Six words that tend to occur in the same 18c volumes as 'gratify' (in TCP-ECCO), plotted over time in a different collection (a corrected version of Google ngrams).


To test this I wrote a script that chose 200 words at random from the top 5000 in a collection of 2,193 18c volumes (drawn from TCP-ECCO with help from Laura Mandell), and then created a cluster around each word by choosing the 25 words most likely to appear in the same volumes (cosine similarity). Would pairs of words drawn from these randomly distributed clusters also show a tendency to correlate with each other over time?

They absolutely do. The Fisher weighted mean pairwise r for all possible pairs drawn from the same cluster is .267 in the 18c and .284 in the 19c (the 19c results are probably better because Google’s dataset is better in the 19c even after my efforts to clean the 18c up*). At n = 100 (measured over a century), both correlations have rock-solid statistical significance, p < .01. And in case you're wondering … yes, I wrote another script to test randomly selected words using the same statistical procedure, and the mean pairwise r for randomly selected pairs (factoring out, as usual, partial correlation with the larger 5000-word group they’re selected from) is .0008. So I feel confident that the error level here is low.**

What does this mean, concretely? It means that the universe of word-frequency data is not chaotic. Words that appear in the same discursive contexts tend, strongly, to track each other over time, and (although I haven’t tested this rigorously yet), it’s not hard to see that the converse proposition is also going to hold true: words that track each other over time are going to tend to have contextual associations as well.

To put it even more bluntly: pick a word, any word! We are likely to be able to define not just a topic associated with that word, but a discourse — a group of words that are contextually related and that also tend to wax and wane together over time. I don’t imagine that in doing so we prove anything of humanistic significance, but I do think it means that we can raise several thousand significant questions. To start with: what was the deal with eagerness, gratification, and disappointment in the second half of the eighteenth century?

* A better version of Google’s 18c dataset may be forthcoming from the NCSA.

** For people who care about the statistical reliability of data-mining, here’s the real kicker: if you run a Benjamini-Hochberg procedure on these 200 randomly-generated clusters, 134 of them have significance at p < .05 in the 19c even after controlling for the false discovery rate. To put that more intelligibly, these are guaranteed not to be xkcd’s green jelly beans. The coherence of these clusters is even greater than the ones produced by topic-modeling, but that’s probably because they are on average slightly smaller (25 words); I have yet to test the relative importance of different generation procedures while holding cluster size rigorously constant.

Topics tend to be trends. Really: p < .05!

Posted on by
Reply

While I’m fascinated by cases where the frequencies of two, or ten, or twenty words closely parallel each other, my conscience has also been haunted by a problem with trend-mining — which is that it always works. There are so many words in the English language that you’re guaranteed to find groups of them that correlate, just as you’re guaranteed to find constellations in the night sky. Statisticians call this the problem of “multiple comparisons”; it rests on a fallacy that’s nicely elucidated in this classic xkcd comic about jelly beans.

Simply put: it feels great to find two conceptually related words that correlate over time. But we don’t know whether this is a significant find, unless we also know how many potentially related words don’t correlate.

One way to address this problem is to separate the process of forming hypotheses from the process of testing them. For instance, we could use topic modeling to divide the lexicon up into groups of terms that occur in the same contexts, and then predict that those terms will also correlate with each other over time. In making that prediction, we turn an undefined universe of possible comparisons into a finite set.

Once you create a set of topics, plotting their frequencies is simple enough. But plotting the aggregate frequency of a group of words isn’t the same thing as “discovering a trend,” unless the individual words in the group actually correlate with each other over time. And it’s not self-evident that they will.

The top 15 words in topic #91, "Silence/Listened," and their cosine similarity to the centroid.


So I decided to test the hypothesis that they would. I used semi-fuzzy clustering to divide one 18c collection (TCP-ECCO) into 200 groups of words that tend to appear in the same volumes, and then tested the coherence of those topics over time in a different 18c collection (a much-cleaned-up version of the Google ngrams dataset I produced in collaboration with Loretta Auvil and Boris Capitanu at the NCSA). Testing hypotheses in a different dataset than the one that generated them is a way of ensuring that we aren’t simply rediscovering the same statistical accidents a second time.

To make a long story short, it turns out that topics have a statistically significant tendency to be trends (at least when you’re working with a century-sized domain). Pairs of words selected from the same topic correlated significantly with each other even after factoring out other sources of correlation*; the Fisher weighted mean r for all possible pairs was 0.223, which measured over a century (n = 100) is significant at p < .05.

In practice, the coherence of different topics varied widely. And of course, any time you test a bunch of hypotheses in a row you're going to get some false positives. So the better way to assess significance is to control for the "false discovery rate." When I did that (using the Benjamini-Hochberg method) I found that 77 out of the 200 topics cohered significantly as trends.

There are a lot of technical details, but I'll defer them to a footnote at the end of this post. What I want to emphasize first is the practical significance of the result for two different kinds of researchers. If you're interested in mining diachronic trends, then it may be useful to know that topic-modeling is a reliable way of discovering trends that have real statistical significance and aren’t just xkcd’s “green jelly beans.”

The top 15 terms in topic #89, "Enemy/Attacked," and their cosine similarity to the centroid.


Conversely, if you're interested in topic modeling, it may be useful to know that the topics you generate will often be bound together by correlation over time as well. (In fact, as I’ll suggest in a moment, topics are likely to cohere as trends beyond the temporal boundaries of your collection!)

Finally, I think this result may help explain a phenomenon that Ryan Heuser, Long Le-Khac, and I have all independently noticed: which is that groups of words that correlate over time in a given collection also tend to be semantically related. I've shown above that topic modeling tends to produce diachronically coherent trends. I suspect that the converse proposition is also true: clusters of words linked by correlation over time will turn out to have a statistically significant tendency to appear in the same contexts.

Why are topics and trends so closely related? Well, of course, when you’re topic-modeling a century-long collection, co-occurrence has a diachronic dimension to start with. So the boundaries between topics may already be shaped by change over time. It would be interesting to factor time out of the topic-modeling process, in order to see whether rigorously synchronic topics would still generate diachronic trends.

I haven’t tested that yet, but I have tried another kind of test, to rule out the possibility that we’re simply rediscovering the same trends that generated the topics in the first place. Since the Google dataset is very large, you can also test whether 18c topics continue to cohere as trends in the nineteenth century. As it turns out, they do — and in fact, they cohere slightly more strongly! (In the 19c, 88 out of 200 18c topics cohered significantly as trends.) The improvement is probably a clue that Google’s dataset gets better in the nineteenth century (which god knows, it does) — but even if that’s true, the 19c result would be significant enough on its own to show that topic modeling has considerable predictive power.

Practically, it’s also important to remember that “trends” can play out on a whole range of different temporal scales.

For instance, here’s the trend curve for topic #91, “Silence / Listened,” which is linked to the literature of suspense, and increases rather gradually and steadily from 1700 to the middle of the nineteenth century.

By contrast, here’s the trend curve for topic #89, “Enemy/Attacked,” which is largely used in describing warfare. It doesn’t change frequency markedly from beginning to end; instead it bounces around from decade to decade with a lot of wild outliers. But it is in practice a very tightly-knit trend: a pair of words selected from this topic will have on average 31% of their variance in common. The peaks and outliers are not random noise: they’re echoes of specific armed conflicts.

* Technical details: Instead of using Latent Dirichlet Allocation for topic modeling, I used semi-fuzzy c-means clustering on term vectors, where term vectors are defined in the way I describe in this technical note. I know LDA is the standard technique, and it seems possible that it would perform even better than my clustering algorithm does. But in a sufficiently large collection of documents, I find that a clustering algorithm produces, in practice, very coherent topics, and it has some other advantages that appeal to me. The “semi-fuzzy” character of the algorithm allows terms to belong to more than one cluster, and I use cosine similarity to the centroid to define each term’s “degree of membership” in a topic.

I only topic-modeled the top 5000 words in the TCP-ECCO collection. So in measuring pairwise correlations of terms drawn from the same topic, I had to calculate it as a partial correlation, controlling for the fact that terms drawn from the top 5k of the lexicon are all going to have, on average, a slight correlation with each other simply by virtue of being drawn from that larger group.

Should we model “topics” as associative clusters, or as statistical factors?

Posted on by
5

I should say up front that this is going to be a geeky post about things that happen under the hood of the car. Many readers may be better served by scrolling down to “Trends, topics, and trending topics,” which has more to say about the visible applications of text mining.

I’ve developed a clustering methodology that I like pretty well. It allows me to map patterns of usage in a large collection by treating each term as a vector; I assess how often words occur together by measuring the angle between vectors, and then group the words with Ward’s clustering method. This produces a topic tree that seems to be both legible (in the sense that most branches have obvious affinities to a genre or subject category) and surprising (in the sense that they also reveal thematic connections I wouldn’t have expected). It’s a relatively simple technique that does what I want to do, practically, as a literary historian. (You can explore this map of eighteenth-century diction to check it out yourself; and I should link once again to Sapping Attention, which convinced me clustering could be useful.)

But as I learn more about Bayesian statistics, I’m coming to realize that it’s debatable whether the clusters of terms I’m finding count as topics at all. The topic-modeling algorithms that have achieved wide acceptance (for instance, Latent Dirichlet Allocation) are based on a clear definition of what a “topic” is. They hypothesize that the observed complexity of usage patterns is actually produced by a smaller set of hidden variables. Because those variables can be represented as lists of words, they’re called topics. But the algorithm isn’t looking for thematic connections between words so much as resolving a collection into a set of components or factors that could have generated it. In this sense, it’s related to a technique like Principal Component Analysis.

Deriving those hidden variables is a mathematical task of considerable complexity. In fact, it’s impossible to derive them precisely: you have to estimate. I can only understand the math for this when my head is constantly bathed in cool running water to keep the processor from overheating, so I won’t say much more about it — except that “factoring” is just a metaphor I’m using to convey the sort of causal logic involved. The actual math involves Bayesian inference rather than algebra. But it should be clear, anyway, that this is completely different from what I’m doing. My approach isn’t based on any generative model, and can’t claim to reveal the hidden factors that produce texts. It simply clusters words that are in practice associated with each other in a corpus.

I haven’t tried the Bayesian approach yet, but it has some clear advantages. For one thing, it should work better for purposes of classification and information retrieval, because it’s looking for topics that vary (at least in principle) independently of each other.* If you want to use the presence of a topic in a document to guide classification, this matters. A topic that correlated positively or negatively with another topic would become redundant; it wouldn’t tell you much you didn’t already know. It makes sense to me that people working in library science and information retrieval have embraced an approach that resolves a collection into independent variables, because problems of document classification are central to those disciplines.

On the other hand, if you’re interested in mapping associations between terms, or topics, the clustering approach has advantages. It doesn’t assume that topics vary independently. On the contrary, it’s based on a measure of association between terms that naturally extends to become a measure of association between the topics themselves. The clustering algorithm produces a branching tree structure that highlights some of the strongest relationships and contrasts, but you don’t have to stop there: any list of terms can be treated as a vector, and compared to any other list of terms.

A fragment of a larger tree, produced by clustering the top 1650 terms in a collection of 2,193 18c documents. Click through for a larger image giving more context.


Moreover, this flexibility means that you don’t have to treat the boundaries of “topics” as fixed. For instance, here’s part of the eighteenth-century tree that I found interesting: the words on this branch seemed to imply a strange connection between temporality and feeling, and they turned out to be particularly common in late-eighteenth-century novels by female writers. Intriguing, but we’re just looking at five words. Maybe the apparent connection is a coincidence. Besides, “cried” is ambiguous; it can mean “exclaimed” in this period more often than it means “wept.” How do we know what to make of a clue like this? Well, given the nature of the vector space model that produced the tree, you can do this: treat the cluster of terms itself as a vector, and look for other terms that are strongly related to it. When I did that, I got a list that confirmed the apparent thematic connection, and helped me begin to understand it.

A list of terms most strongly associated with the group "instantly, cried, felt, moment, longer" in a generically diverse corpus of 2,193 18c documents ranging from sermons to cookbooks. "Similarity" here is technically measured as the cosine of the angle between two vectors.


This is, very definitely, a list of words associated with temporality (moment, hastily, longer, instantly, recollected) and feeling (felt, regret, anxiety, astonishment, agony). Moreover, it’s fairly clear that the common principle uniting them is something like “suspense” (waiting, eagerly, impatiently, shocked, surprise). Gender is also involved — which might not have been immediately evident from the initial cluster, because gendered words were drawn even more strongly to other parts of the tree. But the associational logic of the clustering process makes it easy to treat topic boundaries as porous; the clusters that result don’t have to be treated as rigid partitions; they’re more appropriately understood as starting-places for exploration of a larger associational web.

[This would incidentally be my way of answering a valid critique of clustering — that it doesn't handle polysemy well. The clustering algorithm has to make a choice when it encounters a word like "cried." The word might in practice have different sets of associations, based on different uses (weep/exclaim), but it's got to go in one branch or another. It can't occupy multiple locations in the tree. We could try to patch that problem, but I think it may be better to realize that the problem isn't as important as it appears, because the clusters aren't end-points. Whether a term is obviously polysemous, or more subtly so, we're always going to need to make a second pass where we explore the associations of the cluster itself in order to shake off the artificiality of the tree structure, and get a richer sense of multi-dimensional context. When we do that we'll pick up words like "herself," which could justifiably be located at any number of places in the tree.]

Much of this may already be clear to people in informatics, but I had to look at the math in order to understand that different kinds of “topic modeling” are really doing different things. Humanists are going to have some tricky choices to make here that I’m not sure we understand yet. Right now the Bayesian “factoring” approach is more prominent, partly because the people who develop text-mining algorithms tend to work in disciplines where classification problems are paramount, and where it’s important to prove that they can be solved without human supervision. For literary critics and historians, the appropriate choice is less clear. We may sometimes be interested in classifying documents (for instance, when we’re reasoning about genre), and in that case we too may need something like Latent Dirichlet Allocation or Principal Component Analysis to factor out underlying generative variables. But we’re just as often interested in thematic questions — and I think it’s possible that those questions may be more intuitively, and transparently, explored through associational clustering. To my mind, it’s fair to call both processes “topic modeling” — but they’re exploring topics of basically different kinds.

Postscript: I should acknowledge that there are lots of ways of combining these approaches, either by refining LDA itself, or by combining that sort of topic-factoring approach with an associational web. My point isn’t that we have to make a final choice between these processes; I’m just reflecting that, in principle, they do different things.

* (My limited understanding of the math behind Latent Dirichlet Allocation is based on a 2009 paper by D.M. Blei and J. D. Lafferty available here.)

Seriously geeking out.

Posted on by
Reply

The pace of posts here has slowed, and it may stay pretty slow until I get some new data-slicing tools set up.

I spent the weekend trying to understand when I might want to use a vector space model to compare documents or terms, and when ordinary Pearson’s correlation would be better. Also, I now understand how Ward’s method of hierarchical agglomerative clustering is different from all the other methods.

I know kung fu.



Aside from the sheer fun of geekery, what I’ve learned is that the digital humanities have become *much* easier to enter than they were in the 90s. I attempted a bit of data-mining in the early 90s, and published an article containing a few graphs in Studies in Romanticism, but didn’t pursue the approach much further because I found it nearly impossible to produce the kind of results I wanted on the necessary scale. (You have to remember that my interests lean toward the large end of the scale continuum in DH.)

I told myself that I would get back in the game when the kinds of collections I needed began to become available, and in the last couple of years it became clear to me that they were, if not available, at least possible to construct. But I actually had no idea how transparent and accessible things have become. So much information is freely available on the web, and with tools like Zotero and SEASR the web is also becoming a medium in which one can do the work itself. Everything’s frickin interoperable. It’s so different from the 90s when you had to build things more or less from scratch yourself.