The above account is an empirical hypothesis about certain psychological correlates of grammaticalization. It does not subsume grammaticalization under automation, but constructs an indirect relation between the two processes. The general hypothesis is that grammatical operations are processed in the individual mind with a higher degree of automaticity and a lesser degree of consciousness than lexical and discourse operations. This hypothesis should be testable by methods of cognitive psychology and psycholinguistics. Three of the observable correlates of automatic (as opposed to controlled) behavior enumerated in the table of controlled and automatic processing may be taken out here and proposed as test criteria. For each of the three tests, two variants are proposed. Hypothesis #a concerns the processing of grammatical as opposed to lexical material, but does not directly relate to grammaticalization. This, in turn, is afforded by hypothesis #b.

Given that the purpose of the present lecture is to put forward the hypothesis of the correlation between grammaticalization and automation and to motivate it on theoretical grounds, it should be understood that the types of experiment suggested here to test it can only be sketched. Their purpose is to show that the hypothesis is falsifiable. Before any of the tests can actually be executed, the variables involved have to be firmly controlled. Needless to say, the degree of grammaticalization of linguistic units is determined on purely linguistic grounds (most reliably, by the parameters proposed in Lehmann 2015), thus, in complete independence from their production.

Speech errors

Automatic behavior is freer of errors than controlled behavior. One test therefore involves counting the speech errors (slips of the tongue) in a corpus of connected speech. The hypothesis is then the following:

  1. Generic: There will be relatively less errors in selection and combination of grammatical units than in selection and combination of lexical units.
  2. Specific: For any given linguistic unit which has a more and a less grammatical use in the language, the more grammatical occurrences will be affected by lesser speech errors.

A common method of heightening the number of speech errors in some stretch of speech in order to widen the basis for statistics is to distract the speaker. Now one of the claimed correlates of automatic processing is precisely its relative insusceptibility to interference from simultaneous tasks. Therefore, this method is directly applicable to the present test: Have subjects perform some unrelated task, e.g. potato peeling, while they produce speech. Again, predictions #a and b# should be born out.

One thing to be kept in mind while implementing these tests is that grammatical items differ much from lexical items in their token frequency in running text. One will therefore have to count the number of errors per 100 lexical items and the number of errors per 100 grammatical items, or alternatively the number of errors in the total of occurrences of an item.

Production speed

Another empirical claim associated with automatic as opposed to controlled behavior is that it runs faster. Thus, hesitation pauses preceding grammatical units should be shorter than pauses preceding lexical units. However, this expectation must be modified because a speech pause does not necessarily come immediately before a problematic word, but often before the constituent containing a problematic word. The relevant hypothesis should therefore be based on measuring, in a corpus of connected speech, the length of pauses immediately preceding the last word of a constituent. The hypothesis will then run as follows:

  1. Generic: The pauses preceding grammatical units will be shorter, on average, than the pauses preceding lexical units.1
  2. Specific: For any given linguistic unit which has a more and a less grammatical use in the language, length of speech pauses preceding the former will be shorter, on average, than length of the latter.

A variant of this test counts hesitation interjections instead of speech pauses.

Memory of speech production

Yet another empirical claim about automatic behavior is that its memory imprint is weaker than for controlled behavior. This may be tested as follows: Have subjects engage in a conversation. Afterwards, check their memory of what they said.2

  1. For instance:

    • Did you say war or conflict?
    • Did you say this war or that war?

  2. For instance:

    • Did you say I have a car or I possess a car?
    • Did you say I have gone to Boston or I went to Boston?

The relevant hypothesis is the following: Recall of the subject’s choice in the first of the paired alternatives of #a and #b will be better than recall of choice in the second alternative.

1 Levelt (1989:203) reports on empirical evidence confirming this hypothesis.

2 Givón (2002, ch. 7.4) reports on an experiment in which subjects were asked to report on a conversation they just had. They remembered well the shifting speaker-hearer roles, the shifting intended speech acts and the explicitly stated shifting epistemic modalities associated with the recalled information. They did not recall any content associated exclusively with grammatical structure.