2.3: Corpus linguistics as a scientific method
-
- Last updated
- Save as PDF
Say we have noticed that English speakers use two different words for the forward-facing window of a car: some say windscreen , some say windshield . It is a genuinely linguistic question, what factor or factors explain this variation. In line with the definition above, we would now try to determine their distribution in a corpus. Since the word is not very frequent, assume that we combine four corpora that we happen to have available, namely the BROWN, FROWN, LOB and FLOB corpora mentioned in Section 2.1.2 above. We find that windscreen occurs 12 times and windshield occurs 13 times.
That the two words have roughly the same frequency in our corpus, while undeniably a fact about their distribution, is not very enlightening. If our combined corpus were representative, we could at least conclude that neither of the two words is dominant.
Looking at the grammatical contexts also does not tell us much: both words are almost always preceded by the definite article the , sometimes by a possessive pronoun or the indefinite article a . Both words occur frequently in the PP [ through NP], sometimes preceded by a verb of seeing, which is not surprising given that they refer to a type of window. The distributional fact that the two words occur in very similar grammatical contexts is more enlightening: it suggests that we are, indeed, dealing with synonyms. However, it does not provide an answer to the question why there should be two words for the same thing.
It is only when we look at the distribution across the four corpora, that we find a possible answer: windscreen occurs exclusively in the LOB and FLOB corpora, while windshield occurs exclusively in the BROWN and FROWN corpora. The first two are corpora of British English, the second two are corpora of American English; thus, we can hypothesize that we are dealing with dialectal variation. In other words: we had to investigate differences in the distribution of linguistic phenomena under different conditions in order to arrive at a potential answer to our research question.
Taking this into account, we can now posit the following final definition of corpus linguistics:
Definition (Final Version)
Corpus linguistics is the investigation of linguistic research questions that have been framed in terms of the conditional distribution of linguistic phenomena in a linguistic corpus.
The remainder of Part I of this book will expand this definition into a guideline for conducting corpus linguistic research. The following is a brief overview.
Any scientific research project begins, obviously, with the choice of an object of research – some fragment of reality that we wish to investigate –, and a research question – something about this fragment of reality that we would like to know.
Since reality does not come pre-packaged and labeled, the first step in formulating the research question involves describing the object of research in terms of constructs – theoretical concepts corresponding to those aspects of reality that we plan to include. These concepts will be provided in part by the state of the art in our field of research, including, but not limited to, the specific model(s) that we may choose to work with. More often than not, however, our models will not provide fully explicated constructs for the description of every aspect of the object of research. In this case, we must provide such explications.
In corpus linguistics, the object of research will usually involve one or more aspects of language structure or language use, but it may also involve aspects of our psychological, social or cultural reality that are merely reflected in language (a point we will return to in some of the case studies presented in Part II of this book). In addition, the object of research may involve one or more aspects of extralinguistic reality, most importantly demographic properties of the speaker(s) such as geographical location, sex, age, ethnicity, social status, financial background, education, knowledge of other languages, etc. None of these phenomena are difficult to characterize meaningfully as long as we are doing so in very broad terms, but none of them have generally agreed-upon definitions either, and no single theoretical framework will provide a coherent model encompassing all of them. It is up to the researcher to provide such definitions and to justify them in the context of a specific research question.
Once the object of research is properly delineated and explicated, the second step is to state our research question in terms of our constructs. This always involves a relationship between at least two theoretical constructs: one construct, whose properties we want to explain (the explicandum ), and one construct that we believe might provide the explanation (the explicans ). In corpus linguistics, the explicandum is typically some aspect of language structure and/or use, while the explicans may be some other aspect of language structure or use (such as the presence or absence of a particular linguistic element, a particular position in a discourse, etc.), or some language external factor (such as the speaker’s sex or age, the relationship between speaker and hearer, etc.).
In empirical research, the explicandum is referred to as the dependent variable and the explicans as the independent variable – note that these terms are actually quite transparent: if we want to explain X in terms of Y, then X must be (potentially) dependent on Y. Each of the variables must have at least two possible values . In the simplest case, these values could be the presence vs. the absence of instances of the construct, in more complex cases, the values would correspond to different (classes of) instances of the construct. In the example above, the dependent variable is WORD FOR THE FORWARD-FACING WINDOW OF A CAR with the values WINDSHIELD and WINDSCREEN; the independent variable is VARIETY OF ENGLISH with the values BRITISH and AMERICAN (from now on, variables will be typographically represented by small caps with capitalization, their values will be represented by all small caps). 5 The formulation of research questions will be discussed in detail in Chapter 3, Section 3.1.
The third step in a research project is to derive a testable prediction from the hypothesis. Crucially, this involves defining our constructs in a way that allows us to measure them, i.e., to identify them reliably in our data. This process, which is referred to as operationalization , is far from trivial, since even well-defined and agreed-upon aspects of language structure or use cannot be straightforwardly read off the data. We will return to operationalization in detail in Chapter 3, Section 3.2.
The fourth step consists in collecting data – in the case of corpus linguistics, in retrieving them from a corpus. Thus, we must formulate one or more queries that will retrieve all (or a representative sample of) cases of the phenomenon under investigation. Once retrieved, the data must, in a fifth step, be categorized according to the values of the variables involved. In the context of corpus linguistics, this means annotating them according to an annotation scheme containing the operational definitions. Retrieval and annotation are discussed in detail in Chapter 4.
The fifth and final step of a research project consists in evaluating the data with respect to our prediction. Note that in the simple example presented here, the conditional distribution is a matter of all-or-nothing: all instances of windscreen occur in the British part of the corpus and all instances of windshield occur in the American part. There is a categorical difference between the two words with respect to the conditions under which they occur (at least in our corpora). In contrast, the two words do not differ at all with respect to the grammatical contexts in which they occur. The evaluation of such cases is discussed in Chapter 3, Section 3.1.2.
Categorical distributions are only the limiting case of a quantitative distribution: two (or more) words (or other linguistic phenomena) may also show relative differences in their distribution across conditions. For example, the words railway and railroad show clear differences in their distribution across the combined corpus used above: railway occurs 118 times in the British part compared to only 16 times in the American part, while railroad occurs 96 times in the American part but only 3 times in the British part. Intuitively, this tells us something very similar about the words in question: they also seem to be dialectal variants, even though the difference between the dialects is gradual rather than absolute in this case. Given that very little is absolute when it comes to human behavior, it will come as no surprise that gradual differences in distribution will turn out to be much more common in language (and thus, more important to linguistic research) than absolute differences. Chapters 5 and 6 will discuss in detail how such cases can be dealt with. For now, note that both categorical and relative conditional distributions are covered by the final version of our definition.
Note also that many of the aspects that were proposed as defining criteria in previous definitions need no longer be included once we adopt our final version, since they are presupposed by this definition: conditional distributions (whether they differ in relative or absolute terms) are only meaningful if they are based on the complete data base (hence the criterion of completeness ); conditional distributions can only be assessed if the data are carefully categorized according to the relevant conditions (hence the criterion of systematicity ); distributions (especially relative ones) are more reliable if they are based on a large data set (hence the preference for large electronically stored corpora that are accessed via appropriate software applications); and often – but not always – the standard procedures for accessing corpora ( concordances , collocate lists , frequency lists ) are a natural step towards identifying the relevant distributions in the first place. However, these preconditions are not self-serving, and hence they cannot themselves form the defining basis of a methodological framework: they are only motivated by the definition just given.
Finally, note that our final definition does distinguish corpus linguistics from other kinds of observational methods, such as text linguistics, discourse analysis, variationist sociolinguistics, etc., but it does so in a way that allows us to recognize the overlaps between these methods. This is highly desirable given that these methods are fundamentally based on the same assumptions as to how language can and should be studied (namely on the basis of authentic instances of language use), and that they are likely to face similar methodological problems.
5 Some additional examples may help to grasp the notion of variables and values. For example, the variable INTERRUPTION has two values, PRESENCE (an interruption occurs) vs. ABSENCE , (no interruption occurs). The variable SEX, in lay terms, also has two values (MALE vs. FEMALE). In contrast, the value of the variable GENDER is language dependent: in French or Spanish it has two values (MASCULINE vs. FEMININE), in German or Russian it has three (MASCULINE vs. FEMININE vs. neuter) and there are languages with even more values for this variable. The variable VOICE has two to four values in English, depending on the way that this construct is defined in a given model (most models of English would see ACTIVE and PASSIVE as values of the variable VOICE, some models would also include the MIDDLE construction, and a few models might even include the ANTIPASSIVE).