Preliminary Notes on Conceptual Issues Affecting Interpretation of Topic Models

by Mauro Carassai
Published July 31, 2018

The WE1S Research Blog posts discoveries, observations, and questions by project members bearing on WE1S's themes and methods. (For context, see "About" WE1S.)
Any process of interpretation of textual data relates, to some extent, to the fundamental interplay between observable features belonging to the object of our inquiry and the specific “perspective” that we use in observing those features, i.e. the specific point of view that makes us see what we see during such an observational process.

If we agree to take each and every aspect of such fundamental interplay into consideration while trying to develop both our “topic” and “topic model” interpretation methodology, then we need to start from analyzing the problem of what we see.

The initial consideration might therefore be that what you see, say, in the DFR Browser are not the empirical or directly observable data, but rather estimated data resulting from a quantitative processing of the actual words (tokens) in the corpus. As tempting as it might be to consider the word-image list in front of us as the object of analysis in itself, it might be worth keeping in mind that such list of words represents rather a possible picture of our object of analysis based on the so-called posterior probability. In Bayesian statistics, elements are affected by posterior probability when these further conditions are assigned after the relevant evidence or background is taken into account. That is, the list of words constituting a topic is not just merely the count of actual word occurrences (a topic model from this point of view is not in itself just concrete quantitative reading) but a probabilistic distribution of words occurring in the corpus that uses quantitative counting of words occurrences and frequencies. Topic modeling, when considered from this point of view, might come into view as possibly different from other forms of distant reading insofar as it does not primarily use, statistically speaking, just actual physical evidence but rather uses physical evidence and frequencies counting as mostly a means to an end.

As we know, Bayesian statistics works not with physical probabilities but with evidential probabilities, i.e. degrees of belief. As a result, in topic modeling the computer, rather than being in charge of simply counting or retrieving data beyond human possibilities, is in charge of making an initial wild (i.e. large-scale) computational-based guess and then of making iterations of probabilistic-based of hypotheses based on the data counting. Conceptually, this means that we are (mostly) putting the computer in charge of a subjective, speculative function based on data evidence. Generally speaking, any recognition of the computer performance in these terms should ideally lead us to fundamentally question the dubious usefulness of pinning down the concrete substance of a (computer visualized) speculative attempt.

Keeping in mind all of the above, we can move to addressing a pressing concern that comes into view when relying on previous scholarship about “topic labeling.” The problem of most papers discussing attempts at both automating or standardizing single topic naming/interpretation concerns the fact that scholars seem to consider the meaning of a topic as univocal (see, for example, Sievert & Shirley “What is the meaning of a topic?” and other articles). Despite our use of digital technologies, such conception of meaning represents a particular problematic conceptual node. On the one hand, the humanities – and literary theory in particular – have traditionally been fairly suspicious of this specific stance when addressing natural language. On the other hand, it might be useful to assume that, as a probability-based aggregation, a topic is always within the realm of a “possibility of signification,” a status that literary theory has, for good or bad, extensively addressed over the past fifty years.

If we keep this issue in mind, the “problem” – to an extent – seems to disappear into possibly becoming an apparent one. Interpretation of topics cannot be automated for the simple reason that a topic can still be fundamentally seen as a text produced by the computation-based imagination of a computer. The etymology of the word “text” (from Latin texere = to weave) in humanistic settings accounts for the ways in which the interpreter can implement (i.e. make prominent) various different configurations among the possible ways to connect the text’s specific elements.

Just like a literary classic, a topic is “news that stays news.” In aggregating (a list of) disparate terms in ways not obvious to the human eye, the topic keeps “radiating meaning,” that is, keeps meaning different things to different people in different settings and for different purposes. Unless our purpose in digital humanities settings has surreptitiously become to constrain natural language into the fixity of formal languages for the (possibly weak) reason that language has operationally (and temporarily) been transformed into and processed as data, this aspect should constitute no real problem whatsoever. On the contrary, the remarkable amount of scholarship addressing meaning ambiguity and polysemy comes as an extremely valuable help and actual life-saving toolkit here. The problem is, in other words, an old one and, as such, incredibly familiar to all scholarly discussion of interpreting what we see in front of our eyes – perhaps remembering that the most difficult things to see are the ones in plain view.

Just like with interpretation of traditional texts, of course we want an interpretive protocol of both topics and topic models that might avoid the extremes, i.e. one that avoids the random unfounded take of the subjective impression as well as one that avoids the tautological nonsense of the univocal formalized expression (it might be worth reminding ourselves that in natural languages, the mathematical expression “a = a” means, in the end, virtually nothing).

The problem of topic interpretation, therefore, seems to already shift back to a basic one: an interpretation differs from a mere opinion insofar as it is supported by identified (locatable) evidence that can be shared inter-subjectively and that can become visible not only to the interpreter but also made visible to others. From this point of view, the interpretation (naming/labeling) of a topic model can be made inter-subjective only to the extent that evidence for the specific, contingent, “crystallized label” at hand are conspicuously and clearly articulated by means of the interpreter’s description.

However, in its own turn, pointing to evidence, can become relevant, meaningful, and ultimately intelligible only to the extent that the interpreter and her audience are all interested in working together on the same “mystery case.” In other words, visible evidence becomes clear in front of our eyes only if we explicitly lay out what is the specific point of view the interpreter adopts and what is the specific lens of analysis that we are willing to accept in our provisional, circumstantial, and purpose-oriented interpretation. Which leads us to the next blog post, and possibly provides justification for its tentative title: “Reading a Topic: Theoria means ‘view,’ ‘contemplation.'”