3.4.3. Discussion

The distribution of grammatical constructions suggests that nouns which on average appear in a smaller number of more lexically diverse constructions will receive more support from lexical frames, resulting in less variance in the conditional probability between nouns and the words on which they are conditioned. That is, the more high-frequency nouns tend to appear in larger, high frequency contexts and thus tend to be further subcategorized by smaller lexical subcategorization frames. In contrast, the extent to which the variety of grammatical contexts serves to reduce uncertainty across a smaller (by type count) set of verbs will lead to more variance in the conditional probability between verbs and verb arguments.

In the next section, we explore the effects that the distinct patterns of covariance between high-frequency verbs and high-frequency nouns and their collocates have on the variety of articulated variants we find in the speech corpus.

#### **4. From Information Structure to Speech**

The results we have described so far sugges<sup>t</sup> that the structure of speech serves to facilitate efficient message transmission over multiple nested levels of description. The distribution of lexical and grammatical contrasts indicates that information structure *depth* increases over message sequences, supporting gradual increases in the degree to which low-level sublexical contrasts contribute to resolving uncertainty about a message. Consistent with this, it has been shown experimentally that speech rates are perceived as being faster and target words as being longer when cognitive load is increased [47], a response pattern that suggests that speakers adapt their response to the relative uncertainty resulting from utterance context.

The notion that the timescale variance captured in speaker and listener performances reflects adaptation to uncertainty is further supported by evidence showing that the sublexical variation in speech sequences increases with sequence length, a phenomenon characterized by the strengthening of word initial consonants and the lengthening of final vowels. While both effects increase cumulatively as a function of utterance length [48], the interaction between lengthening and strengthening is weak, indicating that hyperarticulation and vowel space expansion are not equally affected by context. Moreover, while low-probability and word initial segments are more likely to be stressed and while segmen<sup>t</sup> deletion is more likely in high-frequency phonemes and in latter positions, the frequency effects actually observed in very frequent segments depart from this pattern. Also, the correlation between duration and extreme articulation, and duration and frequency declines as a function of utterance position [49].

The analyses presented in Section 3.4 indicate that average grammatical uncertainty peaks in words that are more likely to occur in utterance initial positions and that average lexical uncertainty peaks in categories which are more likely at utterance final positions. It has also been shown that slow-downs in articulation are associated with uncertainty and that uncertainty leads to an increase in articulatory variance. These effects have been observed both within [23] and across word boundaries [18] as a consequence of syntactic irregularities [22] and appear functional in lexical decision [24] and discourse [26]. Since our analyses show substantial differences across parts of speech in both the extent to which words are predicted by the previous context and the extent to which they serve to predict the upcoming part of the message across the frequency range, this seems to imply that the apparently inconsistent effects of frequency that have been previously observed are both predictable and systematic with respect to the structure of the grammatical context.

This in turn can be taken to sugges<sup>t</sup> that sublexical variance follows as a consequence of an increase in lexical and grammatical variety in which words are embedded and that the variants we observe aim to increase the efficiency in transmission of informative contrast at multiple levels of description. In the next section, we conduct a statistical analysis of the effects of variation in the collocations of words on the number of distinct forms found in the speech corpus.

#### *4.1. Effects of Frequency and Collocate Diversity on Variation*

#### 4.1.1. The Distinct Effects of Collocate Diversity And Frequency

Wedel and colleagues have shown that the number of competing minimal pairs in lexical context predict likelihood of vowel merger [50] and voice onset time duration [17], suggesting that what drives speech contrast loss is the extent to which minimal pair competition is resolved in context. In line with this, Piantadosi et al. [51] observe that the relative probability of a word in a lexical context (defined as word sequences ranging between 2–4 words) is a far better predictor of word length than word frequency.

This raises questions: Does this hold for variance too? Is the diversity of collocate contexts across which a word appears a better predictor of the extent to which a type will vary across a speech sample than frequency?

The probability of a known word appearing in a previously unattested context increases with the average word count so that word frequency and collocate diversity are strongly correlated (*r*(9190) = 0.70, *p* < 0.0001). High-frequency words are more likely to be preceded by a larger number of different words and thus tend to appear across a larger number of communicative contexts that vary in size. By implication, there is more variance in the conditional probability between high-frequency words and their collocates. In contrast, words from the mid-frequency range will appear in a smaller number of distinct communicative contexts, leading to less variance in the conditional probability between mid-frequency words and their collocates. In line with this, an analysis by Arnon and Priva [52] shows that, in contrast to results reported by Bell et al. [18], duration in content words is affected by both word and multiword frequency as well as the transitional probability of both the preceding and following collocates when high- and low-frequency trigrams; sequences interrupted by pauses and word final sequences are excluded from the analysis. Finally, the increase in lexical diversity over utterance length (Section 3.3) suggests that low-frequency words tend to appear in a larger number of distinct message contexts, again leading to more variance in the conditional probabilities of low-frequency words at different positions within the sequence with respect to the likelihood of the message.

The discriminative nature of learning predicts that this variance will increase within-context competition over exposure time and that this will minimize the informativeness of contextual cues which predict a large number of lexical contrasts. This in turn predicts more sublexical variation in words that serves as cues to a larger number of collocates, reflecting the uncertainty of the relative context. Taken together, these factors predict distinct patterns of variance across frequency ranges.
