On p-language, g-language, and Naturalistic Inquiry:
Sailing the Seas of Confusion with Language Typologist Martin Haspelmath
by Peter Ludlow
In (2019) I published an article on best theory criteria in linguistics and in that article, I took issue with Martin Haspelmath’s “framework free” approach to the study of language. I argued that it amounted to a “bottom-up approach” approach to linguistic theorizing (hence, not a viable methodology), and I expressed concern that Haspelmath’s stated method is merely a mode of disguising some ossified theoretical assumptions.
I recently discovered that Martin Haspelmath objected to my remarks his blog Diversity Linguistics Comment (https://dlc.hypotheses.org/1801). The crux of his objection, or one of the cruxes of his objection was my “sloppy failure to distinguish between p-linguistics and g-linguistics.”
Ludlow is not a prominent figure and I could simply ignore him, but my paper has been misunderstood by others (see here and here), so it is probably worth emphasizing again that all I’m saying is that p-theories should not make use of speculative ideas of what UG might possibly be like. And what is the source of the misunderstanding? I think it’s the sloppy failure to distinguish between p-linguistics and g-linguistics. [His emphases.]
Apart from the question of whether one should stoop to conquer the non-prominent, what is the core issue that needs attention? As the quote above suggests, Haspelmath believes that there is a kind of confusion in the linguistics literature (apparently also the philosophy of linguistics literature) that extrudes from a failure to distinguish p-theories from g-theory. What are those you ask?
p-theory is a theory of a p-language and a p-language is a particular language, which is to be contrasted with g-language, which is the covering notion of “human language” writ large. From Haspelmath:
Particular languages (p-languages) are immediately observable, in the sense that we can hear people talk in them (i.e. we can observe speech produced in them). But the general phenomenon of Human Language (g-language) is a more elusive concept — even in the 18th century, not all intellectuals were sure that all human groups had a p-language (see Levelt 2018), and it takes some abstraction to say that every p-language is somehow a manifestation of a more general notion of “Human Language”. In other domains, the analogous difficulties are serious: For example, are Christianity, Shintoism and Confucianism manifestations of a more general phenomenon of “Human Religion”? [His emphases.]
Now, I’m not entirely sure why Haspelmath believes I have confused my p’s and g’s, since that is a little bit like confusing a ham sandwich with the concept of food, and I really don’t think that the distinction has somehow snuck by me. Mind you, I do see confusion here, but I don’t see a confusion of p-language and g-language. In fact, the only confusion I see here is in the assumption that p-languages and g-language are what I or any generative linguist take to be the ultimate object of enquiry. For that matter, I doubt that any serious scientific investigation into language would take p-languages or g-language to be the ultimate objects of enquiry. As we will see, the same will be true of prefixing p- and g- to any other areas of scientific enquiry. p-rocks and g-rock are not ultimate objects of enquiry in geology, p-animals and g-animal are not ultimate objects of enquiry in biology, etc. People are not confusing these p- and g- things. They are just investigating other things.
Now, I’m not saying that there aren’t such things as individual languages, and I’m not saying we don’t talk about studying them and about learning them, but sometimes when we talk this way in the scientific context, it is loose talk. There is an important distinction to be made between the facts and phenomena you are collecting or investigating (let’s say the facts and phenomena related to the dialect of Bello in the suburbs of Medellin in Antioquia, Colombia), and the object of study, which is the thing that you are ultimately constructing a theory of. For a generative linguist that would be a parametric setting of the language faculty, or UG, or I-language, or whatever you want to call it. Other linguists will have some other ultimate objects of investigation, but whatever they are, they should not be confused with a p-language (or g-language).
So, to be clear, I am not hating on p-languages, I’m really not. What I am saying is that they are not the principal object of study. They are a way of getting at that object of study. A p-language is a pretheoretical way of organizing facts and phenomena. How we bundle up the facts/phenomena and what we call that bundle of facts/phenomena really depends on our interests. Today, it might be useful to talk about the bundle of facts/phenomena that we call the dialect of Bello, whereas tomorrow we might talk about the dialect of the Antioquia region in Colombia, and the next day we might be talking about the Colombian dialect of Spanish, and the next day about Latin American Spanish, and then in the fullness of time Romance languages and Indo-European languages. So, yes, I believe in p-languages, but I understand them to be ways of organizing linguistic facts and phenomena based on apparent similarities and usefulness to our projects, whether those projects be political or scientific or whatever.
The above remarks track an important observation in the philosophy of science from Bogen and Woodward (1988): Facts and phenomena are not the same as the object of study in any field of scientific enquiry. Facts and phenomena provide evidence for the nature of the object of study. For a generative linguist, facts about languages/dialects/etc. provide evidence for the structure of the language faculty (UG). Meanwhile getting a grasp on what the language faculty is and how it works should help to explain the facts/phenomena that you witness during your investigation. This is the distinction that Haspelmath seems to have lost track of — the distinction between the object of study and the facts/phenomena that provide evidence for our theory of the object of study. If you are a generative linguist, the object of investigation is the language faculty. As I said, other kinds of linguists may have other objects of investigation.
This brings us back to my beef with Haspelmath and Haspelmath’s framework-free methodology. First, let’s hear Haspelmath out on how he understands his framework free project:
In a (2010) paper entitled “Framework-free grammatical theory”, I argued that linguists wishing to analyze/describe a single language (i.e. “p-linguists”) should not adopt one of these innate frameworks, but should build a custom framework for the purposes of their language (in other words: “describe each language in its own terms”, as urged by fieldworkers since Boas’s time). [His emphasis.]
OK, what is the problem with studying a p-language “in its own terms?” Here is one worry. Some naturalists might become very smitten with jade and devote themselves to studying it “in its own terms” — in a way that is not corrupted by all the gobbledygook dished up by those mineralogists. But here is the problem: That class of phenomena that is called “jade” is not actually a single thing. Our jade-loving naturalists were actually studying two things — jadeite (which is NaAlSi2O6) and nephrite (which is Ca2(Mg, Fe)5Si8O22(OH)2) — not the same things! If you studied “jade” in “its own terms” you would never understand that it is not itself a natural kind. Learning this requires a whole lot of theory. And no, no one is confusing p-minerology and g-minerology here. There aren’t two kinds of minerology. The goal is to understand particular things with the help of theory that you have developed in the study of other minerals, and for that matter, studying the basic principles of chemistry and atomic theory that underly those minerals. Theory is what you use to investigate particular things.
There is a general version of this point. Is there actually anything which can or should be studied “in its own terms”? Or another way to put the point is this: If you are studying something in “its own terms” are you even studying? After all, one way to understand the study of the natural order is to understand how it fits into the greater scheme of things. That is, to study something is to learn how it is integrated with the natural order. Note, we aren’t talking about g-linguistics or g-minerology or g-physics or g-anything here. We are talking about how the facts and phenomena we are looking at hook up with the world as we understand it. Is there another way to look at things? Sure, you can look at the facts and phenomena in a way in which they can’t possibly hook up with the world as we understand it, or you might look at it in a way that the world used to be understood but is no longer, but in that case, you are not describing things “in their own terms.” You are just describing them incorrectly.
It is hard to imagine Darwin, perhaps the greatest naturalist of all, saying, “I study each animal in its own terms.” To the contrary, after looking at the strange new creatures in the Galapagos Islands, he said this: “The naturalist, looking at the inhabitants of these volcanic islands in the Pacific, distant several hundred miles from the continent, yet feels that he is standing on American land. Why should this be so?” Indeed, why should that be so?
The voyage of the Beagle did not produce some catalogue of observations each disconnected from the next. Darwin saw each in relation to the next. In doing this he was not doing g-biology. He was coming up with a theory of the principles that could explain how each species he encountered came to have the properties that it did. And he also came to understand that superficial observations of animals (p-animals if you will) were not the actual objects of study. In fact, those p-animals did not even line up with the natural kinds — they did not line up with the actual species (even Darwin did not recognize how many species of finches he had collected).
Darwin was of course very good at seeing the interconnection between things, but even when he zeroed in on a particular p-animal it was to acquire insight into the more important underlying principles of nature. Thus, when Darwin spent ten long years studying barnacles (see Stott, 2003) he did not study barnacles “in their own terms.” He studied them, armed with the theory of evolution (by then written up and sitting in his desk) and he used that theory to inform his observations of that class of phenomena. And armed with that background theory he made an enormous number of discoveries about various species (loosely lumped together and called barnacles) and how they evolved and how they fit into the broader evolutionary theory.
Here is another way to illustrate the point, from Darwin, in a letter to the naturalist Henry Fawcett:
About 30 years ago there was much talk that geologists ought to observe and not to theorize; and I well remember someone saying that at this rate a man might as well go into a gravel pit and count the pebbles and describe the colors. How odd it is that anyone should not see that all observation must be for or against some view if it is to be of any service! (Darwin and Seward (1903)).
Suppose that a geologist decided that they were going to do p-geology for that gravel pit and thus decided that they were going to study it “in its own terms.” Well, ok, such a person could sit around all day and randomly describe rocks and rock-like things, dutifully ignoring theoretical questions about how rocks are formed and what their origins might be based on glacial movements, and how those glacial movements might have ground those rocks into the shapes that they have and oh yeah, how some of those rock-looking things might be fossilized animals etc. etc. Ignoring all of those theoretical questions from outside the p-geology silo, what is one actually looking for? What is worth describing? You can spend a whole day describing one rock if you want. Or you can choose rocks at random and describe random properties. The question is, where exactly are you going with that enterprise? Spoiler alert: nowhere.
This raises the question of how Haspelmath’s idea of “studying things in their own terms,” might smuggle in some very nasty theoretical assumptions. The question of course, is what does “studying things in their own terms” entail? For example, had Darwin studied barnacles in their own terms would he have exorcised all thinking about evolution? Had he done so, what would be the alternative?
This is the point where Haspelmath insists that he is not engaged in bottom-up methodology.
…what I am concerned with in the 2010 paper is p-theories: comprehensive descriptions of particular languages (such as my 1993 grammar of Lezgian). We may not often call grammatical descriptions/analyses “theories”, but they are certainly not atheoretical data collections. Everyone who has thought in some depth about p-linguistic analysis knows that it is a complex interaction of hypothesis formation, data collection for hypothesis testing, reformuation of hypotheses, and so on. P-linguistic analysis is an eminently theoretical enterprise. [His emphases.]
First let’s note that Haspelmath’s issue with me doesn’t actually have anything to do with distinctions between p-language and g-language; at least not as far as I can tell. It is really fundamentally an issue about Haspelmath’s relationship to theory. In the above quote, Haspelmath insists that he is all about theory, and come to think of it, my point was that Haspelmath’s approach was theoretical as all Sam Hades — “eminently theoretical” in the new vernacular. And my second point was that Haspelmath was attempting to disguise his theoretical assumptions and present his approach as being somehow more honest or wholesome or kinder to our understanding of individual languages, when in point of fact it was anything but that. It was fossilized theory in atheoretical sheep’s clothing.
But now that Haspelmath proclaims that he is eminently theoretical, is he off the hook? Are his theoretical intentions transparent now? Well, sorry, no. Haspelmath and I might be on the same page about one thing — some sort of theory must be in play when you study things “in their own terms,” but what sort of theory? In the passage quoted above, Haspelmath says it works like this: “it is a complex interaction of hypothesis formation, data collection for hypothesis testing, reformuation of hypotheses, and so on.”
Here is my take on that passage. I think that what Haspelmath is saying is “yes of course there is theory, but it is theory that is constructed without the outside influence of general theoretical interests, it is a theory that is constructed for the data set or facts and phenomena in question. Or, if you prefer, it is a theory that is constructed for that very p-language, ignoring all other theoretical concerns. Again, had Darwin proceeded in this way with barnacles, it would be a sad day for our understanding of the species he was investigating. Put a pin in that thought.
What are we to make of Haspelmath’s insistence that his method is not bottom-up? What it would mean is that he doesn’t begin with data collection. He begins with a hypothesis. So, the idea would be that you have some version of what philosophers call the Nomological-Deductive (N-D) model, which is exactly the model that Haspelmath describes: you form a hypothesis, collect data to test the hypothesis, revise the hypothesis, repeat, etc. The method is not bottom up (in theory) because you don’t get the hypothesis directly from observation (and good thing, because no one believes that is possible anymore). Instead, you don’t worry about where the hypothesis comes from. You just worry about whether you can confirm (or in Popper’s case, falsify) the hypothesis in question. Haspelmath’s wrinkle is that you put on blinders or silo the enterprise so that it isn’t corrupted by outside influences.
Now again, I first want to come back to the point of how weird it would be if Darwin had proceeded in that way: “I aim to study barnacles, but I shall ignore that manuscript “The Origin of the Species” that is even now sitting in my desk. And this leads to a further question: If Haspelmath is allowing his siloed enterprise to be top-down, where is he getting his hypotheses from?
If Haspelmath is not working bottom up, then he has to draw his initial hypotheses from somewhere, and the usual view of people working within the N-D model is that hypotheses can come from anywhere you wish. So the question is, why would the initial hypotheses for a p-language investigation not come from some version of UG — for example from Chomsky’s Minimalist Program? As stated, his description of his mode of investigation cannot be complete, because as it stands it would not rule out the very methods he is rejecting. (As near as I can make out Haspelmath’s ultimate point here, it is that if you try to bring concepts from UG into your study of a particular language you are engaged in bad methodology.)
What Haspelmath needs is to specify which initial hypotheses are permitted into his investigation and which are not. Is he going to rule out standard Minimalist hypotheses by fiat before the investigation even starts? If so, then what are the permissible initial working hypotheses? And it is here that I thought I saw an attempt to smuggle in some sort of ossified theoretical assumptions, if only by deliberate exclusion of others.
For example, I made the point that while Haspelmath is usually scrupulous about excluding theoretical machinery from contemporary generative linguistics, he apparently adores theoretical machinery that is a couple thousand years old. Haspelmath readily helps himself to notions like predicate and topic, even if he assigns them new roles and meanings. One wonders why they get to count as “framework free.”
The immediately preceding remarks are intended to express concern about the origin of hypotheses in Haspelmath’s attempt to carry out the Nomological-Deductive method within a siloed domain of phenomena — what he likes to call a p-language. But there is a much bigger problem than the question of where the hypotheses are supposed to come from. The bigger problem is that the N-D method is considered to be, as the kids say, rekt. And the bigger bigger problem is that the received way out of the wreckage is to delve into the nature of what explanation is, and once you start thinking about explanation, you realize that it has everything to do with how your theory hooks up with everything else. That is to say, if you want to escape the wreckage of the N-D model, you have to escape any siloes that attempt to make you investigate things “in their own terms.”
Here, a little bit of philosophy of science will be necessary. Philosophers of science like Harman (1965), Boyd (1981, 1984), Harré (1986, 1988), Lipton (2004), Psillos (1999), and McMullin (1992) advocated an alternative to the N-D model which is sometimes called Inference to the Best Explanation (IBE). For example, you don’t abandon your theory in the face of a counterexample; you seek ways explain the alleged counterexample that are consistent with your theory. Even though the evidence is equally good in either direction you say that the height of the flagpole determines the length of the shadow and not the other way around because the former theory is explanatory (see Bromberger (1966) for a discussion of the N-D model-killing objection). And why should we not lean on the explanatory ability of theories? — it is not merely a “get out of jail free” card for theorists. Nor is it merely a desideratum for our theories. It is a primary goal of scientific theorizing.
In other words, the growing consensus about scientific theorizing is that we want more than theories that fit the facts; we want theories that explain the facts. This is of course a point that Chomsky (1964) made in his distinction between descriptive adequacy and explanatory adequacy. But what does this come to? When is a theory explanatory?
Seemingly, the answer is that there is more to a good explanatory theory than just surviving attempts at (dis)confirmation. The theory needs to have something else. Kitcher (1989), for example, offered that the more explanatory theory was the one that allowed us to unify our explanations.
Science advances our understanding of nature by showing us how to derive descriptions of many phenomena, using the same pattern of derivation again and again, and in demonstrating this, it teaches us how to reduce the number of facts we have to accept as ultimate. (p.432).
In other words, the explanatory theory is the one that can show us how to explain multiple things in terms of a repeating pattern of explanation.
It isn’t just explanation that requires this connection to broader theory. It is recognized that even best-theory criteria like simplicity rely on this is well. Thus, Eliot Sober (1975) argued that the only really coherent notion of simplicity in scientific enquiry is a notion of simplicity that involves ease of integration with the rest of science. The point is, that if you were doing siloed science, even concepts like simplicity would be outside of your grasp.
Returning to the connection between explanation and integration with broader science, I believe that this is part of the reason Darwin’s work is recognized as being deeply explanatory; the various macro-level patterns that we see in nature are a function of a very simple pattern at the micro level.
What is the big point here? It is this: You can’t silo an area of investigation, because if you do that you aren’t doing science, because science isn’t just about confirmation of hypotheses — that doesn’t work. It can’t work unless you also rope in the idea that your ultimate goal is explanatory. And explanation, at bottom, is an attempt to show how the thing you are ultimately studying connects with everything else.
From this a number of conclusions start cascading. If science is really about explanation and if explanation is really about connecting with more general patterns, then any scientific observations, if they really are scientific observations, need to have an eye to the project of connecting the thing you are observing with more general principles and properties of the world — or, as Kitcher put it, to find the patterns that occur elsewhere. If you aren’t looking for those general patterns, then you are like the p-geologist in Darwin’s gravel pit just randomly looking at rocks.
Now, setting aside all this business about the philosophy of science and explanation and methodology, maybe we should just pause and think about what Haspelmath’s actual concern is, abstracting away from his explicit theoretical writing, which probably isn’t serving his cause that well. And I can imagine one line of thought, which on first blush seems reasonable. Perhaps the worry is that by bringing theory to your investigation of some class of phenomena (i.e., some p-language) you run the risk that the theory will color your description of the phenomena, possibly in a nonveridical way. That is, maybe you can be so deep into a linguistic framework (say the Minimalist Program) that what looks like a y to you is actually an x.
Now this might sound plausible from the outset, but it is only plausible if you forget that any sort of theory might color your description in this way, including the theory that you constructed while ignoring everything outside of your p-language silo, and recall that Haspelmath admits that his approach is “eminently theoretical.” Even if he hadn’t admitted this, there is no escaping the problem. No one gets to claim “the view from nowhere” because there is no view from nowhere.
In practical terms, Haspelmath’s worry seems to be a very strange boogieman. Anyone who sat in on a Chomsky syntax lecture at MIT during the 1970s and 1980s and 1990s and maybe to his last lecture at MIT (I don’t know) has witnessed Chomsky making a theoretical proposal, followed by students and visitors attempting to counterexample him with facts from about a dozen or more languages that they speak natively. And the point is this: approaching linguistic phenomena with a theory in hand does not make the counterexamples to the theory disappear. It makes the counterexamples pop into existence (it is that elixir from which LI squibs are born). It highlights facts and data that might otherwise have escaped our attention. MITs Ken Hale knew a lot about the GB program in generative linguistics, but it did not prevent him from observing thousands of facts about Warlpiri and other languages that appeared to be puzzles for GB, and remain puzzles to this day.
Many years ago, I spoke with a colleague from a language department who said that he had gotten out of linguistics. When I asked him why, he said “I thought the counterexamples to the theory were the most interesting part.” I then asked him, “ok, but now that you have abandoned theory, what examples are interesting to you,” and he stared at me like the proverbial deer in the headlights. The unspoken answer, of course, was that nothing was interesting. If you don’t come at phenomena with theory in hand, nothing is really inherently interesting. You are like the guy in Darwin’s gravel pit, picking up rocks at random and describing them with no apparent rhyme or reason. Any sane person would rather just sit around and eat glue.
One final thought. You might think that in view of all of the forgoing discussion, Haspelmath probably did mess up in dragging the concepts of p-language and g-language out of retirement, and likewise you might think that yeah, he did mess up in thinking he could silo a field of investigation off from broader theory and from integration with the other sciences, and on top of all that you might think he probably also screwed up in trying to bring the defunct N-D model of explanation to his rescue, but in the end perhaps he was right when he insisted that (contra Ludlow) he is not working bottom-up.
That certainly sounds plausible on first glance. After all, he did say that he would allow the forming of hypotheses before he got to the good business of fact collecting. We can invent a name for his method. Let’s called the Siloed Nomological-Deductive method. It is siloed for all the reasons that Haspelmath said — he is studying languages “in their own terms,” unencumbered by pesky theoretical considerations from outside the data silo. But here is the problem. If he is going to allow hypotheses to come in from elsewhere — let’s say from the MIT linguistics department — then his position just collapses into everyone else’s. If he tries to constrain the class of permissible outside hypotheses, then he is playing favorites for no good reason. This seems to suggest that at the end of the day, if he doesn’t want his proposal to be either vacuous or arbitrarily restrictive, he seems to be left with hypotheses that are also siloed, and that would have to mean that they are in some sense tied to or directly inspired by the phenomena in the siloed domain of investigation.
OK, now we come to a point where we are talking about a view that is top-down on paper, but it isn’t top-down in the usual sense of top-down inquiry where the hypotheses of investigation can be drawn from elsewhere in the natural sciences. No, once we are restricting ourselves to hypotheses that can only be directly inspired by the phenomena immediately under investigation, we are actually engaged in bottom-up enquiry. If you didn’t start with the facts on the ground, you would have no way of knowing what sorts of hypotheses are admissible. You have just disguised your bottom-up enquiry as something else. And now the problem is that you really can’t engage in bottom-up scientific investigation in any meaningful way, so if you are engaged in science at all, the enterprise has to be top-down and explanatory, whether you want it to be or not. At the end of the day, there is no such thing as purely descriptive science.
This brings us to the first moral of the story: Thinking you have somehow escaped frameworks or that you are investigating in a frame-work-free way is inherently dangerous. Indeed, it is just as dangerous to the descriptive enterprise as explicit theoretical bias. On the one hand, if you happen to be openly hostile to a framework (and let’s be honest, there is a lot of that going around) then you can fail to see phenomena that might be relevant to a theory. On the other hand, even if you aren’t hostile at all, you can fail to see facts relevant to the theory because you are carrying baggage for some other tacit framework that is so deeply compiled within your thinking that you don’t see it is coloring your investigation.
The ultimate moral about the frame-work-free ideology is this: No one works in a framework-free way. No one can. And more to the point, if you could do it you wouldn’t want to, because without broader frameworks in which to integrate our discoveries, science is simply not possible.
Achinstein, P. (2001). The Book of Evidence, Oxford: Oxford University Press.
Achinstein, P. (2018). Speculation: Within and about Science, Oxford: Oxford University Press.
Boyd, R., (1981). Scientific Realism and Naturalistic Epistemology. In P. Asquith and R. Giere, eds., PSA 1980, (vol. II), East Lansing MI: Philosophy of Science Association, 613–662.
Boyd, R. (1984). The Current Status of Scientific Realism. In J. Leplin, ed., Scientific Realism, Berkeley CA: University of California Press, 41–82.
Bromberger, S. (1966). Why Questions. In R. Colodny, ed., Mind and Cosmos: Essays in Contemporary Science and Philosophy, Pittsburgh: University of Pittsburgh Press.
Chomsky, N. (1964). Current Issues in Linguistic Theory. In J. Fodor, J. and J. Katz, eds., The Structure of Language, Prentice Hall, pp. 50–118.
Harman, G. (1965). The Inference to the Best Explanation. Philosophical Review, 74, 88–95.
Harré, R. (1986). Varieties of Realism, Oxford: Blackwell.
Harré, R. (1988). Realism and Ontology. Philosophia Naturalis, 25, 386–398.
Haspelmath, M. (2010). Framework-Free Grammatical Theory. In B. Heine and H. Narrog, eds., The Oxford Handbook of Linguistic Analysis. Oxford: Oxford University Press, 341–365.
Haspelmath M. (2019). Confusing p-linguistics and g-linguistics: Philosopher Ludlow on “framework-free theory”. Diversity Linguistics Comment. June, 23, 2019. https://dlc.hypotheses.org/1801
Kitcher, P. (1989). Explanatory Unification and the Causal Structure of the World. In P. Kitcher and W. Salmon, eds., Scientific Explanation. Minneapolis: University of Minnesota Press, pp. 410–505.
Ludlow, P. (2011). The Philosophy of Generative Linguistics, Oxford: Oxford University Press.
Ludlow, P. (2019). The philosophy of generative linguistics: Best theory criteria. In A. Kertész, E. Moravcsik, and C. Rákosi (eds.), Current approaches to syntax: A comparative handbook, vol. 3, 521–547. Berlin: De Gruyter Mouton.
MacCawley, J. (1981). Review of F. Newmeyer, Linguistic Theory in America. Linguistics 18.911–30.
McMullin, E. (1992). The Inference that Makes Science, Milwaukee WI: Marquette University Press.
Popper, K. (1959). The Logic of Scientific Discovery, London: Hutchinson.
Psillos, S. (1999). Scientific Realism: How Science Tracks Truth, London: Routledge.
Stott, R. (2003). Darwin and the Barnacle: The Story of one Tiny Creature and History’s Most Spectacular Scientific Breakthrough, New York: W.W. Norton.
 I don’t want to clog up the main thread here, but there is also an important distinction to be made between the facts/phenomena and the data. Data (whether linguistic judgments of native speakers, or corpora data, or field interviews) provides evidence for the facts/phenomena. Those facts in turn provide evidence for the theory of the object of study. Working in the reverse direction, the theory will predict and explain those facts/phenomena. I use ‘facts’ and ‘phenomena’ interchangeably because we are taking phenomena here to be, roughly, stuff that is happening. It should not be conflated with something like phenomenal experience. For more about this I recommend checking out Bogen and Woodward (1988) and/or my application of their position to linguistics in Ludlow (2011).