4.1.3. Discussion

The results of this analysis align with the finding that word counts outside of their communicative context contribute little when it comes to explaining variation in articulated forms. Rather, we observe that the largest part of this variance is explained by the diversity of the lexical contexts in which words appear. The remaining effects of frequency are limited to a relatively small number of high-frequency nouns and words from closed categories (numbers, contractions, and filled pauses).

These results are thus consistent with the differences we find in the distributional patterns of lexical categories in that, unlike high-frequency nouns, it would seem that high-frequency verbs are far less likely to be encountered outside of their argumen<sup>t</sup> frames (supporting the idea that verbs are encountered as arguments rather than lexical items per se).

Given that our results show that the variance in the observed forms is largely explained by the covariance in the collocate structure and that patterns of covariance are systematic, this finally leads us to the question of the systematicity in the sublexical variance: Is the distribution of the observed contrast geometric?

#### *4.2. Distribution of Word Initial Contrast*

#### 4.2.1. Why Word Initial Contrast?

Previous work on sublexical variation shows that the structure of speech sound sequences is such that the probability of speech segments at segmen<sup>t</sup> transitions is not independent [49]. Gating paradigm studies have shown that the informativeness of word medial contrast is mediated by the extent to which both the preceding sentence context and word initial phonetic contrasts have minimized uncertainty about the word [53,54]. Accordingly, the entropy in sublexical contrast peaks at word initial boundaries [49]. This suggests that word initial speech contrasts may serve a distinct communicative function in context.

An initial analysis of word initial phonetic label distributions over both observed and citation forms in the corpus revealed poor fits to both power law and exponential distributions, suggesting that the aggregated distribution of the phonetic labels observed in our corpus may result from mixing the underlying communicative distributions. To examine this, we used parts-of-speech classes to provide a simple, objective method for contextually disaggregating individual communicative distributions from the mixed distribution of phonetic labels in our corpus.
