Volume 4, Issue s2 - Special Issue: The Role of Predictability in Shaping Human Language Sound Patterns / Issue Editors: Jason Shaw and Shigeto Kawahara

Many papers in this special issue grew out of the talks given at the Symposium “The role of predictability in shaping human language sound patterns,” held at Western Sydney University (Dec. 10–11, 2016). Some papers were submitted in response to an open call; others were invited contributions. This introduction aims to contextualize the papers in the special issue within a broader theoretical context, focusing on what it means for phonological theory to incorporate gradient predictability, what questions arise as a consequence, and how the papers in this issue address these questions.

Phonological knowledge is influenced by a variety of cues that reflect predictability (e.g. semantic predictability). Listeners utilize various aspects of predictability when determining what they have heard. In the present paper, we ask how aspects of the acoustic phonetic signal (e.g. speaking rate) interact with other knowledge reflecting predictability (e.g. lexical frequency and collocation strength) to influence how speech is perceived. Specifically, we examine perception of function words by native and non-native speakers. Our results suggest that both native and non-native speakers are sensitive to factors that influence the predictability of the signal, including speaking rate, frequency, and collocation strength, when listening to speech, and use these factors to predict the phonological structure of stretches of ambiguous speech. However, reliance on these cues differs as a function of their experience and proficiency with the target language. Non-native speakers are less sensitive to some aspects of the acoustic phonetic signal (e.g. speaking rate). However, they appear to be quite sensitive to other factors, including frequency. We discuss how these results inform our understanding of the interplay between predictability and speech perception by different listener populations and how use of features reflecting predictability interacts with recovery of phonological structure of spoken language.

A crucial question for historical linguistics has been why some sound changes happen but not others. Recent work on the foundations of sound change has argued that subtle distributional facts about segments in a language, such as functional load, play a role in facilitating or impeding change. Thus not only are sound changes not all equally plausible, but their likelihood depends in part on phonotactics and aspects of functional load, such as the density of minimal pairs. Tests of predictability on the chance of phoneme merger suggest that phonemes with low functional load (as defined by minimal pair density) are more likely to merge, but this has been investigated only for a small number of languages with very large corpora and well attested mergers. Here we present work suggesting that the same methods can be applied to much smaller corpora, by presenting the results of a preliminary exploration of nine Australian languages, with a particular focus on Bardi, a Nyulnyulan language from Australia’s northwest. Our results support the hypothesis that the probability of merger is higher when phonemes distinguish few minimal pairs.

The present study investigates patterns of covariation among acoustic properties of stop consonants in a large multi-talker corpus of American English connected speech. Relations among talker means for different stops on the same dimension (between-category covariation) were considerably stronger than those for different dimensions of the same stop (within-category covariation). The existence of between-category covariation supports a uniformity principle that restricts the mapping from phonological features to phonetic targets in the sound system of each speaker. This principle was formalized with factor analysis, in which observed covariation derives from a lower-dimensional space of talker variation. Knowledge of between-category phonetic covariation could facilitate perceptual adaptation to novel talkers by providing a rational basis for generalizing idiosyncratic properties to several sounds on the basis of limited exposure.

A wide range of reduction phenomena have been described in the literature as predictability effects, in which more predictable units (i.e. words, syllables, vowels) are reduced in duration or other acoustic dimensions relative to less predictable units. The goal of the current study was to critically evaluate these predictability effects on vowel duration in read speech to explore the extent to which they reflect a single underlying phenomenon. The results revealed shorter vowel duration for words with high phonotactic probability, for high-frequency words (in clear speech only), and for words in plain lab speech relative to clear speech. However, the results also revealed qualitatively different effects of three measures of contextual probability (cloze probability, written trigram probability, and spoken trigram probability). Greater spoken trigram probability predicted longer vowel duration, contrary to expectations, and this effect was limited to high-frequency words in first mentions and in plain speech. Cloze probability and written trigram probability exhibited even more complex interactions with other predictability measures. These results provide evidence for fundamental differences in these measures of predictability, suggesting that a more nuanced perspective on predictability effects and the mechanisms underlying them is necessary to account for the complexity of the empirical data.

It has long been noted that language production seems to reflect a correlation between message redundancy and signal reduction. More frequent words and contextually predictable instances of words, for example, tend to be produced with shorter and less clear signals. The same tendency is observed in the language code (e.g. the phonological lexicon), where more frequent words and words that are typically contextually predictable tend to have fewer segments or syllables. Average predictability in context (informativity) also seems to be an important factor in understanding phonological alternations. What has received little attention so far is the relation between various information-theoretic indices – such as frequency, contextual predictability, and informativity. Although each of these indices has been associated with different theories about the source of the redundancy-reduction link, different indices tend to be highly correlated in natural language, making it difficult to tease apart their effects. We present a computational approach to this problem. We assess the correlations between frequency, predictability, and informativity, and assess when these correlations are likely to create spurious (null or non-null) effects depending on, for example, the amount of data available to the researcher.

Recent evidence suggests that the phonetic realization of linguistic units is sensitive to informational context. For example, the duration of a word is shorter when it is probable given the following word. Word-specific phonetic variation is unexpected according to modular/feedforward models. We consider various challenges to identifying the loci of informational effects on phonetic implementation – do they arise in production, perception, memory, or some combination? Section 2 addresses a theoretical issue: what are the right measure(s) of predictability/informativity? An urgent direction for future work is to understand what kinds of context matter and why. Section 3 reviews second-mention reduction and other non-local discourse effects, which strongly suggest a production locus (rather than arising in speech perception or memory). Important future directions include modeling discourse/topic in corpus studies, and experimentally assessing the role of nonlocal context in perception and memory. Section 4 addresses the role of computational modeling. We call for integrated, implemented end-to-end models which include speech perception, lexical representation, and speech production components.

There is now abundant evidence that phonetic forms are shaped by probabilistic effects reflecting predictability or informativity . We outline a number of challenges for such work, where theoretical claims are often based on small differences in acoustic measurements, or interpretations of small statistical effect sizes. We outline caveats about the methods and assumptions encountered in many studies of predictability effects, particularly regarding corpus-based approaches. We consider the wide range of factors that influence patterns of variability in phonetic forms, taking a broad perspective on what is meant by “the message” in order to show that predictability effects need to be considered alongside many others, including indexical and prosodic factors. We suggest a number of ways forward to extend our understanding of the form-predictability relationship.

Much previous research on distributional learning and phonetic categorization assumes that categories are either faithful reproductions or parametric summaries of experienced frequency distributions, acquired through a Hebbian learning process in which every experience contributes equally to the category representation. We suggest that category representations may instead be formed via error-driven predictive learning. Rather than passively storing tagged category exemplars or updating parametric summaries of token counts, learners actively anticipate upcoming events and update their beliefs in proportion to how surprising/unexpected these events turn out to be. As a result, rare category members exert a disproportionate influence on the category representation. We present evidence for this hypothesis from a distributional learning experiment on acquiring a novel phonetic category, and show that the results are well described by a classic error-driven learning model (Rescorla, R. A. & A. R. Wagner. 1972. A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement. In A. H. Black & W. F. Prokasy (eds.), Classical conditioning II: Current research and theory , 64–99. New York, NY: Appleton-Century-Crofts).

This paper analyzes the vocative truncation pattern in Korean from the viewpoint of Message-Oriented Phonology (MOP), which capitalizes on the idea that sound patterns are governed by a principle that makes message transfer effective. In the traditional naming pattern, Korean first names consist of a generation marker and a unique portion, and the order between these two elements alternates between generations. To derive vocative forms, the generation marker is truncated, and the suffixal [(j)a] is attached to the unique portion. We argue that MOP naturally predicts this type of truncation. As the generation marker is shared by all the members of the same generation, the generation marker is highly predictable and hence does not reduce uncertainty about the intended message. To achieve effective communication, predictable portions are deleted. Our analysis implies that MOP is relevant not only to phonetic implementation patterns, but also to morphophonological patterns. It also provides support for MOP based on data from a non-Indo-European language. Finally, we aim to integrate insights of MOP with a more formal proposal like Optimality Theory, by relating the predictability of a contrast to the ranking of the faithfulness constraint that it protects, following the spirit of the P-map hypothesis.

Recent studies in Message Oriented Phonology (MOP) have provided increasing evidence that informativity plays a non-trivial role in linguistic behavior. This paper provides a case study of MOP focusing on the durational contrast of singleton and geminate consonants in spoken Japanese. In modern Japanese, short consonants (singletons) and long consonants (geminates) are lexically contrastive, and the durational properties of these consonants are affected by a variety of factors. This provides a useful test of the assumptions of MOP. Based on the assumption that the higher the informativity, the more robustly the contrast is phonetically implemented, this study examines the hypothesis that the durations of singletons and geminates increase or decrease according to the informativity of their durational contrast. The study confirms that (i) the distribution of singletons and geminates is affected by the manner of articulation and positional differences (morpheme-initial, medial, and final); (ii) the distributional differences follow from the informativity of contrasts as represented by Shannon’s entropy; and (iii) the durational contrast is enhanced by the presence or absence of a minimal pair.

Many studies report shorter acoustic durations, more coarticulation and reduced articulatory targets for frequent words. This study investigates a factor ignored in discussions on the relation between frequency and phonetic detail, namely, that motor skills improve with experience. Since frequency is a measure of experience, it follows that frequent words should show increased articulatory proficiency. We used EMA to test this prediction on German inflected verbs with [a] as stem vowels. Modeling median vertical tongue positions with quantile regression, we observed significant modulation by frequency of the U-shaped trajectory characterizing the articulation of the [a:]. These modulations reflect two constraints, one favoring smooth trajectories through anticipatory coarticulation, and one favoring clear articulation by realizing lower minima. The predominant pattern across sensors, exponents, and speech rate suggests that the constraint of clarity dominates for lower-frequency words. For medium-frequency words, the smoothness constraint leads to a raising of the trajectory. For the higher-frequency words, both constraints are met simultaneously, resulting in low minima and stronger coarticulation. These consequences of motor practice for articulation challenge both the common view that a higher-frequency of use comes with more articulatory reduction, and cognitive models of speech production positing that articulation is post-lexical.

Probabilistic phonetic reduction is widely attested in a variety of languages, acoustic domains, and interpretations of predictability. Less well-studied is the categorical effect of probabilistic segment deletion, which in principle is subject to similar pressures. This paper presents the results of an exploratory study into patterns of segment deletion in corpora of spontaneous speech in English and Japanese. Analysis at the word level reveals that words with more phonemes and higher-frequency words tend to have more of their segments deleted. Analysis at the phoneme level reveals that high-probability phonemes are more likely to be deleted than low-probability phonemes. For Japanese only, this analysis also shows effects of word length, frequency, and neighborhood density on deletion probability. Taken together, these results suggest that several large-scale patterns of probabilistic segment deletion mirror the processes of phonetic reduction and apply to both languages. Some patterns, though, appear to be language-specific, and it is not clear to what extent languages can and do differ in this regard. These findings are discussed in terms of our understanding of the universality of proposed predictability effects, and in terms of probabilistic reduction more broadly.

A diverse set of empirical findings indicate that word predictability in context influences the fine-grained details of both speech production and comprehension. In particular, lower predictability relative to similar competitors tends to be associated with phonetic enhancement, while higher predictability is associated with phonetic reduction. We review evidence that these in-the-moment biases can shift the prototypical pronunciations of individual lexical items, and that over time, these shifts can promote larger-scale phonological changes such as phoneme mergers. We argue that predictability-associated enhancement and reduction effects are based on predictability at the level of meaning-bearing units (such as words) rather than at sublexical levels (such as segments) and present preliminary typological evidence in support of this view. Based on these arguments, we introduce a Bayesian framework that helps generate testable predictions about the type of enhancement and reduction patterns that are more probable in a given language.