8.1 Grammar in corpora
There are two major problems to be solved when searching corpora for grammatical structures. We discussed both of them to some extent in Chapter 4, but let us briefly recapitulate and elaborate some aspects of the discussion before turning to the case studies.
First, we must operationally define the structure itself in such a way that we (and other researchers) can reliably categorize potential instances as manifesting the structure or not. This may be relatively straightforward in the case of simple grammatical structures that can be characterized based on tangible and stable characteristics, such as particular configurations of grammatical morphemes and/or categories occurring in sequences that reflect hierarchical relations relatively directly. It becomes difficult, if not impossible, with complex structures, especially in frameworks that characterize such structures with recourse to abstract, non-tangible and theory-dependent constructs (see Sampson 1995 for an attempt at a comprehensive annotation scheme for the grammar of English).
Second, we must define a query that will allow us to retrieve potential candidates from our corpus in the first place (a problem we discussed in some detail in Chapter 4). Again, this is simpler in the case of morphologically marked and relatively simple grammatical structures, for example, the s -possessive (as above) is typically characterized by the sequence ⟨ [pos="noun"] [word="'s"%c] [pos="adjective"]* [pos="noun"] ⟩ in corpora containing texts in standard orthography; it can thus be retrieved from a POS-tagged corpus with a fairly high degree of precision and recall. However, even this simple case is more complex than it seems. The query will produce false hits: in the sequence just given, ’s may also stand for the verb be (Sam’s head of marketing). The query will also produce false misses: the modified nominal may not always be directly adjacent to the ’s (for example in This office is Sam’s or in Sam’s friends and family), and the s -possessive may be represented by an apostrophe alone (for example in his friends’ families).
Other structures may be difficult to retrieve even though they can be characterized straightforwardly: most linguists would agree, for example, that transitive verbs, are verbs that take a direct object. However, this is of very little help in retrieving transitive verbs even from a POS-tagged corpus, since many noun-phrases following a verb will not be direct objects (Sam slept the whole day) and direct objects do not necessarily follow their verb (Sam, I have not seen); in addition, noun phrases themselves are not trivial to retrieve.
Yet other structures may be easy to retrieve, but not without retrieving many false hits at the same time. This is the case with ambiguous structures like the of - possessive, which can be retrieved by a query along the lines of ⟨ [pos="noun"] [pos="determiner"]? [pos="adjective"]* [pos="noun"] ⟩, which will also retrieve, among other things, partitive and quantitative uses of the of -construction.
Finally, structures characterized with reference to invisible theoretical constructs (“traces”, “zero morphemes”, etc.) are so difficult to retrieve that this, in itself, may be a good reason to avoid such invisible constructs whenever possible when characterizing linguistic phenomena that we plan to investigate empirically.
These difficulties do not keep corpus linguists from investigating grammatical structures, including very abstract ones, even though this typically means retrieving the relevant data by mind-numbing and time-consuming manual analysis of the results of very broad searches or even of the corpus itself, if necessary. But they are probably one reason why so much grammatical research in corpus linguistics takes a word-centered approach.
A second reason for a word-centered approach is that it allows us to transfer well-established collocational methods to the study of grammar. In the preceding chapter we saw that while collocation research often takes a sequential approach to co-occurrence, where any word within a given span around a node word is counted as a potential collocate, it is not uncommon to see a structure-sensitive approach that considers only those potential collocates that occur in a particular grammatical position relative to each other – for example, adjectives relative to the nouns they modify or vice versa. In this approach, grammatical structure is already present in the design, even though it remains in the background. We can move these types of grammatical structure into the focus of our investigation, giving us a range of research designs where one variable consists of (part of) the lexicon (with values that are individual words) and one variable consists of some aspect of grammatical structure. In these studies, the retrieval becomes somewhat less of a problem, as we can search for lexical items and identify the grammatical structures in our search results afterwards, though identifying these structures reliably remains non-trivial. We will begin with word-centered case studies and then move towards more genuinely grammatical research designs.
