Current and Recent Research
Word Frequency Effects in Sound Change as a Consequence of Perceptual Asymmetries
We present a computational exemplar-theoretic model for the perception/production loop, with which we explore how items of different lexical frequencies may respond to changes in discriminability in the phonological space during sound change. The model predicts high-frequency words to change at a faster rate than low-frequency words when the change decreases discriminability, at a slower rate than low-frequency words when the change increases discriminability, and at the same rate as low-frequency words when the change does not affect discriminability.
Diagnosing Change in a Sparse, Bursty Variable: Eh in Pākehā English
I investigate change in the degree of use of the discourse particle eh among Pākehā (white) New Zealanders. Diagnosing change in eh is made difficult by the fact that it is both sparse – not used at all by many speakers – and bursty – infrequent among most users, but excessive among a few. Thus, many speakers may fail to use eh in speech observed in a corpus not because they do not use it at all (in general), but simply because they do not use it enough for it to appear. To address this difficulty, and allow change-in-progress to be distinguished from age-grading, I introduce a new statistical method into the quantitative linguistics toolbox: zero-inflated negative binomial regression. I find quantitative evidence that eh has spread from indigenous Māori to white Pākeha over time, and qualitative evidence that this spread was facilitated by the decrease in social stigma attached to Māori.
Unsupervised Morphological Analysis in a Language with Reduplication
We present an extension of the Morfessor family of unsupervised morphological analysis algorithms that can recognize phonological reduplication. We show that the extension improves accuracy of morphological analyses in Māori, and we conduct a detailed comparison to human analyses that demonstrates the reasons for this improvement and the remaining limitations. We conclude by discussing the general advantages offered by incorporating treatment of reduplication into morphological analyzers.
An Adult Protolexicon
We investigate implicit vocabulary learning by adults who are exposed to a language in their ambient environment. Most New Zealanders do not speak Māori, yet are exposed to it throughout their lifetime. We show that this exposure leads to a very large proto-lexicon - implicit knowledge of the existence of morphemes without any associated meaning. Non-speakers of Māori are able to generalize over this lexicon to generate sophisticated intuitions about the well-formedness of Māori words, in a manner that is indistinguishable from fluent Māori speakers.
Quantitative Insights into Māori Word Structure
We develop quantitative insights into Māori word structure, based on morphological segmentations provided by an unsupervised machine learning system and by fluent speakers of Māori. We analyse over 18,000 words from the Te Aka dictionary, and we find that fewer than 10% of them are unambiguously monomorphemic. Thus, the vast majority of words in Māori were potentially formed via morphological processes. We present a statistical breakdown of the morphological patterns in Māori, in which we quantify and rank the productivities of the morphological processes from the literature. By presenting this breakdown both over word types in the dictionary and over word tokens in corpora of running speech, we obtain probabilistic strategies for Māori word formation, as well as an indication of which strategies are likely to yield words that will be widely used by speakers.
Talker and Stereotype in the False Recall of Spoken Words
We explore the implicit biases that link the use of African American Vernacular English (AAVE) to negative and violent stereotypes, and how these biases interact with linguistic memory. In a DRM task, we find that listeners more often falsely remember hearing words such as gun and gang from an AAVE speaker than from a (white) General American English speaker. We argue that meaning is constructed from both words and voice, and that the strength of a memory for a word is mediated by the word's consistency with stereotypes inferred from the corresponding voice.
Faithfulness and Exceptionality in Polish Stress
We describe new facts about the placement of secondary stress in inflected exceptional stems and compounds in Polish, which are incompatible with existing analyses based in monostratal Optimality Theory. We propose an alternative analysis based in a multistratal approach to OT which not only accounts for these new facts, but also efficiently harnesses the similarity in stress placement across different morphophonological domains.
Syntactic Embedding and Visibility of Morphophonological Structure
I present experiments that investigate gradience and variation in the propensity to suppress the English possessive morpheme /z/ with possessors ending in the plural morpheme /z/. I find that suppression becomes less necessary as the number of syntactic brackets between the two /z/s increases. I sketch formal implementations of this finding in representational and derivational morphophonological frameworks, and I argue that it implies that morphophonological processing must be guided by syntactic (or corresponding prosodic) structure in a stochastic fashion.