Distribution and meaning
Yet given that the treatment of meaning is often regarded as the primary deficiency of the Bloomfieldian (and Descriptivist) model, any rehabilitation of this model must also address semantic concerns. A discriminative perspective again suggests a reinterpretation of earlier proposals. The distributional approach to meaning explored by Harris (1954) is particularly amenable to a discriminative construal on which same-different semantic contrasts are grounded in distributional variation rather than speaker judgments. The meaningful expressions in a language can likewise be discriminated by contextual vectors, which can be automatically obtained from corpora by distributional semantic models such as Marelli and Baroni (2015).
Significantly, these models address the concerns originally expressed by Bloomfield, along with the criticisms provoked by his cautious approach to the study of meaning. These criticisms incorporate two main claims and a number of auxiliary assumptions. The first claim is that Bloomfieldian analyses provided purely formal descriptions that took no account of meaning. The second was that the behaviourist orientation of the Bloomfieldians led them to dismiss ‘mentalist’ conceptions of meaning out of hand. The linkage between these points is clearly articulated in the familiar remarks by Firth:
Certain leading linguists especially in America find it possible to exclude the study of what they call ‘meaning’ from scientific linguistics, but only by deliberately excluding anything in the nature of mind, thought, idea, concept. ‘Mentalism’ is taboo. (Firth 1951: 82)
Implicit in this criticism is the assumption that the study of ‘meaning’ is inextricably bound up with ‘concepts’ and other ‘mentalistic’ notions. Accompanying this assumption is the idea that there is a fundamental opposition between distributional analysis and semantic analysis, and that even the identification of units is expected to make some appeal to meaning. These assumptions are explicit in Carroll’s survey of Descriptivist contemporaries:
A general characteristic of the methodology of descriptive linguistics, as practiced by many American linguists today, is the effort to analyse linguistic structure without reference to meaning. It is thought possible in theory that one could identify the phonemes and morphemes of a language purely on the basis of their distribution, that is, by noting the linguistic environment in which they occur. (Carroll 1953:31f.)
There is no shortage of passages from Descriptivist works, which, particularly when lifted out of context, can be construed as endorsing an extreme distribution- alist position and/or as adopting a doctrinaire view of the role of meaning. In formulating a ‘set of postulates’, Bloch (1948) entertains the possibility of establishing phonemic analysis on a purely distributional basis:
Theoretically it would be possible to arrive at the phonemic system of a dialect entirely on the basis of phonetics and distribution, without any appeal to meaning—provided that in the utterance of the dialect not all possible combinations of phonemes actually occurred. (Bloch 1948:5,61.8)
A more strident view of the primacy of distributional criteria is expressed in the approach to grammatical analysis outlined in Harris (1951):
The main research of descriptive linguistics, and the only relation which will be accepted as relevant in the present survey, is the distribution or arrangement within the flow of speech of some parts or features relative to others. (Harris 1951: 5)
Yet even the positions expressed in these passages are less radical than they may appear in isolation. In the continuation of the footnote quoted above, Bloch (1948) places his distributional approach in the context of the “non-semantic” tradition of phonemic analysis he associates with Daniel Jones:
An important point to notice is that the phoneme is essentially a phonetic conception. The fact that certain sounds are used in a language for distinguishing the meanings of words doesn’t enter into the definition of a phoneme. It would indeed be possible to group the sounds of a language into phonemes without knowing the meaning of any words. (Jones 1929:44)
The elaboration of Harris’s model of grammatical analysis likewise makes it clear that his goal was not to repudiate meaning but to operationalize it, at least partially, in terms that were susceptible to analysis using established methods:
If we know that life and rife are not entirely repetitions of each other, we will then discover that they differ in distribution (and hence in ‘meaning’). It may be presumed that any two morphemes A and B having different meanings also differ somewhere in their distribution: there are some environments in which one occurs and the other does not. (Harris 1951:7, fn.4)
Like the decompositional models of morphemic analysis discussed in Chapter 2.2 and Section 8.4.1, genuinely ‘meaning-free’ models of grammatical analysis cannot be traced back to Bloomfield’s own work but stem from the elaborations of his successors. This development is noted in the quotation on p. 24 above, where Matthews (1991:148) observes that “the origin of... Bloomfieldian constituency analysis... was what remained of Bloomfield’s model when, first, grammatical arrangement is reduced to selection and order and, secondly, all reference to meaning is taken out”. The central role assigned to meaning in the Bloomfieldian model is stated explicitly in the immediate continuation of the passage on p. 216 above, which goes on to state that “A phonetic form with its meaning is a linguistic form; a tactic form with its meaning is a grammatical form” (p. 166; emphasis in original).
The types of meanings that Bloomfield proposes, ‘sememes’ and ‘episememes’, function essentially as placeholders in his system. This does not reflect a hostility towards the study of meaning or even agnosticism about its importance but instead an assessment that “ [t]he statement of meanings is... the weak point in language- study and will remain so until human knowledge advances very far beyond its present state” (Bloomfield 1933:140). The consideration of “mentalistic psychology” in Bloomfield (1933: §9.4) likewise does not treat ‘mentalism’ as ‘taboo, but disputes the usefulness of accounts that “define the meaning of a linguistic form as the characteristic mental event that occurs in every speaker and hearer in connection with the utterance or hearing of the linguistic form” (Bloomfield 1933:142). Far from excluding meaning, the Bloomfieldian model makes essential reference to meaning, but is forced to acknowledge the limitations of techniques for analyzing meaning.
It is thus misleading to describe Bloomfield as hostile to meaning or to the study of meaning. His caution was grounded in reservations about approaches that sought to study meaning without providing techniques for operationalizing meaning. Since operationalization has not been a prominent focus of most subsequent approaches, it is not addressed by accounts that analyze meaning in terms of paraphrases, frames, model-theoretic interpretations, etc.
Instead it is in the context of the distributional model of Harris (1951, 1954) that a general attempt to address Bloomfield’s concerns takes shape. What has come to be known as the ‘Distributional Hypothesis’ links meanings to patterns of occurrence by proposing that linguistic items with similar distributions will have similar meanings. The basic idea, in one form or another, is attributed to a range of sources, including Weaver (1955) and (ironically) Firth (1957), among others. But the hypothesis is particularly compatible with the general distributional methodology developed by Harris.
In effect, this perspective cancels the presupposition that any model can be ‘purely distributional’, or that distributional analysis is generally incompatible with meaning analysis, given that distributional contrasts will correlate with meaning contrasts. Current models of distributional (or vector) semantics have been most extensively developed in the domain of natural language processing. However, there have also been studies that explore the relevance of the Distributional Hypothesis to learning (McDonald and Ramscar 2001) and others that develop sophisticated models of morphological meaning (Lazaridou et al. 2013; Marelli and Baroni 2015). A distribution operationalization of meaning not only addresses Bloomfield’s methodological qualms but provides an observable correlate of meaning that can be acquired by a discriminative learning model based on the primary linguistic data available to a speaker.