This thesis investigates properties of two reflexive pronouns in Late Archaic Chinese (LAC) and Early Middle Chinee (EMC), 自 zì and 己 jǐ, and demonstrates how syntactic generalisations about them are motivated by discourse-functional considerations.

I argue that zì has six major uses: signaling coreference between subjects and (1) objects, (2) possessors of objects, and (3) arguments inside complement clauses, and (4) emphasising the agent against plausible alternatives, (5) expressing the subject’s status as the affectee, and (6) indicating that the action had no external cause. I characterise the prototypical use of zì with four properties: the referent of zì is (1) centrally relevant to the predicate, (2) the same as the subject, (3) unexpected in its role in the predicate, and (4) contrasted with other possible potential referents in its role. I also discuss deviations from these properties.

Turning to jǐ, I identify three main uses: (1) coreference between subject of a complement-taking verb and a referential form inside the complement, (2) coreference between a referent in an agent-nominalised clause and a non-agent referent in surrounding context, and (3) salient contrasts between the ‘self’ and ‘other’. I characterise these uses through a common discourse profile: the referent of jǐ is either (1) the perceiver of a situation or, in EMC, (2) a discourse topic, and in both LAC and EMC, (3) competes with another highly salient referent in the discourse context.

I show that although the reflexives’ individual properties are crosslinguistically common, the constellations of properties and the resultant division of labour between the reflexives are typologically unique. I also argue that, consistent with Ariel’s (2008) account, both reflexives are used in contexts of intermediate discourse accessibility: some factors push the referent’s accessibility up, and others lower it. Yet, unlike Ariel, I argue that the two reflexives differ not in terms of overall accessibility level, but in the sources of accessibility they are sensitive to, supporting a multidimensional view of accessibility, rather than collapsing sources of accessibility into a single value as is traditionally done.

Recent usage-based approaches to linguistic theory have claimed that linguistic processing is driven by domain-general cognitive abilities which operate on a rich memory store that retains all the details of every experience with language, extracting patterns from these experiences purely on the basis of regular patterning. It has also been claimed that these mechanisms operate similarly on all levels of linguistic structure and at all stages of the human lifespan. Taken together, these claims imply the hypothesis that any dimension of experience can influence the linguistic knowledge and behavior of any language user. This dissertation tests this hypothesis. A series of experiments were conducted, each consisting of a prime phase and a test phase, using native English-speaking participants. In the prime phase, participants were exposed to combinations of linguistic structures (active and passive voice) and nonlinguistic contextual elements (colors, background music, sounds, or physical environments). The linguistic and nonlinguistic components of the experiences so created bore no semantic relationship to one another, but the pattern of cooccurrence between them was completely regular and reliable, such that, for each experimental participant, a particular syntactic voice always occurred in a particular nonlinguistic context. Participants then performed a picture description task, in which each picture was accompanied by one of the nonlinguistic contexts to which they had previously been exposed. The hypothesis was that, when describing each picture, participants should be more likely to use the syntactic voice which had previously been associated with the nonlinguistic context which accompanied the picture. This hypothesis was not supported by the data. Instead, it was found that the only consistently significant factor influencing the syntactic voice of participants' responses was the syntactic voice of their own previous responses: people were more likely to keep using whatever voice they had already been using. In addition, in every experiment, it was found that the results were most accurately characterized by an extremely simple model using only subject-specific and picture-specific baseline response rates. These results suggest that usage-based theories have been overly optimistic in asserting that regular patterns of experience alone are sufficient to explain linguistic knowledge and behavior. Instead, it is argued that more specific constraints on linguistic processing mechanisms are needed in order to provide a fullfledged, causal account of how people's experiences affect their mental representations of language.

Recent psycholinguistic research has demonstrated that our knowledge of words includes fine-grained information about the syntactic contexts in which they are likely to participate. In these studies, the syntactic distribution of a word is defined as a probability distribution of its occurrences in various dependency roles. For example, nouns may be more or less likely to serve as the subject of a verb or as the head of an adjective modifier. In contrast to the constraint-based representations of mainstream generative theories, these syntactic distributions are gradient and probabilistic, situating words within a rich, multidimensional syntactic space.At the same time, a growing body of research has identified patterns of systematicity within and among features of the lexicon that reflect cognitive and communicative pressures on learning, memory, production, and perception. For example, the same patterns of clustering and association that are observed for lexical features are also known in the psycholinguistic literature to facilitate aspects of language acquisition and use, yet these tendencies are held in check by pressures toward distinctiveness that are crucial for perception in particular. Using corpus data from forty-eight languages, this dissertation represents the first investigation into how syntactic distributional information is patterned in the lexicon. First, I ask how syntactic representations cluster within the multidimensional syntactic space. I find that the more frequent words have denser orthographic, semantic, and—crucially—syntactic neighborhoods. At least in phonology, having many near neighbors is known to facilitate learning, memory, and production. By analogy, it seems as though the most frequent words are also the most syntactically optimized. Next, I ask if syntactic distributions participate in non-arbitrary relationships with other features of the lexicon, such as semantics and phonology. My analysis shows that, while the meanings of words are correlated positively with both their phonological forms and syntactic distributions, phonology and syntax do not share any significant correlation below the level of word class. This surprising result suggests new ideas for how functional pressures may be negotiated at different hierarchical levels within a particular lexical feature. Finally, I ask whether the syntactic distributions of words can shed light on the organization and function of other grammatical phenomena such as grammatical gender systems. I demonstrate that, while semantically and phonologically similar words are more likely to be grouped within genders, syntactically similar words are more likely to be distributed across genders. I interpret this result as a design feature of language, with grammatical gender serving to disambiguate syntactically similar words. Taken as a whole, the studies within this dissertation paint a novel picture of the role of syntax within the architecture of the lexicon. In some ways syntactic representations pattern similarly to semantic and phonological representations, and yet in other ways syntax seems to have a unique role relative to these other lexical features. The syntactic distributions of words reflect a kind of functional negotiation seen elsewhere in the lexicon, exhibiting both clustering and dispersion in different domains. The balance of such design features within the lexicon lend support to the idea that language structure is evolved for efficient use.

Multiword expressions (MWEs) play a critical role in language acquisition, production, and comprehension, serving as integral components of linguistic formulaicity. Despite their significance, the identification, analysis, and integration of MWEs in corpus-driven and corpus-based research have largely remained peripheral within the discipline. Most advancements in discovering, identifying, and incorporating MWEs into linguistic analyses have primarily emerged from the computational branch of the field. Accordingly, the majority of algorithms designed for the tasks of discovering and identifying MWEs have primarily focused on the outcome itself, privileging performance over interpretability. This dissertation introduces mMERGE (multidimensional MWE Extraction through Recursive Grouping of Elements), a novel algorithm that integrates linguistic insights and computational methods to address these challenges.

At the core of mMERGE are five well-known but underutilized information-theoretic dimensions—frequency, type frequency, dispersion, entropy, and association measures—designed to reflect key patterns of language use and cognitive processing. Furthermore, the algorithm implements these dimensions to address two critical challenges. First, many measures, particularly dispersion, are correlated with frequency, leading to a bias against infrequent MWEs. For instance, traditional dispersion metrics often fail to account for infrequent MWEs that are well-distributed across a corpus, such as domain specific MWEs or idiomatic expressions, thereby underestimating their representativeness. Second, existing measures frequently conflate directionality by assuming bidirectional attraction between words. This approach overlooks asymmetry in certain expressions such as give up or in truth, where one element strongly attracts the other, but not vice versa. mMERGE explicitly addresses these limitations, by using a dispersion measure that partials out the effect of frequency, as well as bidirectional association and entropy measures.

Built on recursive grouping and implemented in the Julia programming language, mMERGE achieves reasonable computational efficiency, enabling its application to the BROWN corpus over 20,000 iterations. Through a combination of human annotation and predictive modeling, the results not only demonstrate the algorithm's capability to uncover diverse MWE categories, such as idioms, compounds, and lexical bundles, but also reveal their distributional profiles and establish connections with the cognitive processes underlying language use.

Finally, a key application explored in this dissertation is MWE-augmented keyword analysis, which demonstrates how integrating MWEs into keyness metrics can enhance the interpretation of linguistic patterns in a given corpus. By examining changes in keyword rankings and distributions, this analysis underscores the impact of MWEs on corpus-based studies of register, domain specificity, and discourse organization.

This research contributes to corpus and computational linguistics by offering a method that balances the demands of efficiency and model interpretability, while addressing the limitations of existing approaches. Beyond theoretical implications, mMERGE offers practical versatility, enabling adaptation to a wide variety of corpora, regardless of their size, domain specificity, or language.

This dissertation investigates how Machine Learning can inform the description of endangered languages through three case studies on Kolyma Yukaghir (Yukaghiric, Russia). Its primary motivation is to investigate how computational approaches can enhance the language description process, which remains primarily based on qualitative analyses. This dissertation aims to demonstrate how various Machine Learning techniques can be effectively applied to different aspects of language structure, even with opportunistic corpora common in endangered language contexts. The research is structured around three case studies, each employing different supervised and unsupervised learning methods to analyze specific aspects of Kolyma Yukaghir grammar.

The first study examines the discourse distributions of the form ʎe- using random forests. This form is glossed with various labels in existing documentation materials, including ‘copula’, ‘discourse particle’, and ‘that one’. Through a combination of qualitative discourse analysis and quantitative modeling, the study establishes two main functions for ʎe-: copula and placeholder. The random forest modeling reveals distinct distributional patterns for each function, suggesting that a polysemy analysis is unlikely. This study thus offers a proof of concept of how supervised learning methods like random forests can be helpful in the context of language description to generalize beyond qualitative analyses and establish the discourse preferences of different elements.

The second study investigates the assertion system employing random forests and k-nearest neighbor algorithms. Current syntax-based analyses fail to account for a variety of discourse uses of assertion forms. Through close readings of narratives and Machine Learning modeling, the study argues that the form -meE is deployed in events to which the speaker wants to direct special attention. This study demonstrates how supervised learning methods like k-nearest neighbor can aid data exploration in language description and help identify grammatical features that may havegone unnoticed in the qualitative analysis of morphosyntactic phenomena.

The third study applies k-means and hierarchical clustering techniques to explore lexical categorization. Kolyma Yukaghir displays various degrees of lexical flexibility, with roots performing referring, predicating, and modifying functions, often without clear signs of derivation. Roots and affixes are clustered based on their co-occurrence patterns, and the results suggest a coarse split between nominal and verbal tendencies for roots, whereas affix patterns reflect both morphosyntacticand morphophonological phenomena. In short, this study showcases how unsupervised learning methods can be helpful to discover nuanced patterns of root and affix behavior while controlling for frequency effects, and how they can help identify the relevant number and identify of lexical classes in a language.

The findings of this dissertation demonstrate that Machine Learning methods can effectively work with the types of datasets typically available in language documentation projects–small, opportunistic corpora with limited genres and registers. Unsupervised learning algorithms appear to be valuable tools to identify latent patterns in the data, which can then generate hypotheses about speakers’ linguistic choices. These hypotheses can be subsequently tested using supervised learning models, even with small datasets coded for multiple predictors. It is important to stress, however, that such methods need to be paired with qualitative analyses that guide their application and interpretation. The strength of Machine Learning techniques for language description lies precisely in this combination of quantitative and qualitative lenses.

In sum, the dissertation offers a proof of concept for incorporating Machine Learning methods into the toolkit of language description. It advocates for a broader scope of languages analyzed under a quantitative lens and offers examples of how to incorporate quantitative validation with these methods into current qualitative analyses.

This dissertation addresses the methodology used in classifying speech rhythms in order to resolve a long-standing linguistic conundrum about whether languages differ rhythmically. There is a widespread perception, both among linguists and the general population, that some languages are stress-timed and others are syllable timed. Stress-timed languages are described as having less-regular rhythms, as syllable durations vary according to the placement of stress in the phrase. Meanwhile, syllable-timed languages are described as displaying less variation in rhythm, which syllable durations being more regular. This dissertation quantitatively evaluates these described rhythmic differences in Spanish,Portuguese, and English. The first chapter introduces speech rhythms and reviews past literature on their perception and production. The second chapter evaluates a widely used metric of speech rhythms, the PVI, and determines that it is not effective in distinguishing between two dialects of Spanish. The third chapter compares the speech rhythms of Mexican and Chicano Spanish. This chapter concludes that Chicano Spanish is more restricted in its vowel duration variability, while Mexican Spanish employs both highly variable durations (i.e. stress-timed) and highly uniform durations (i.e. syllable-timed). The fourth chapter describes a perception study used to compare the speech rhythms of Spanish, English, and Portuguese, and shows that these languages' rhythms do not always group according to language. In the fifth chapter, I describe a study of the production of the same utterances initially used in the perception experiment; this allows an analysis of what prompts the perceptual differences in speech rhythm described in Chapter Four. The sixth and final chapter discusses the implications and applications of these findings and gives direction for further investigation. Although both production and perception studies of speech rhythms have been performed in the past, my dissertation expands these methodologies by combining production and perception data is a single analysis. I use perception data to relatively classify the rhythms of utterances through low-pass speech filtering, then analyze the production of these data computationally to provide a more complete perspective of what prompts differences in speech-rhythms and how Spanish, Portuguese, and English data relate rhythmically. Thus, my dissertation is thorough, while still addressing traditional rhythm metrics and employing current computational methodology. It seeks to challenge linguists' methodologies in quantitatively addressing speech rhythms, and to further clarify the position of Spanish, Portuguese, and English on the speech rhythm continuum.

Extensive literature demonstrates that, with extended contact and high integrative motivation, L2 speakers gradually acquire L2 nonstandard variation over time as a result of interaction with speakers of the target language (Dewaele & Regan 2001, Dewaele 2002, Nagy et al. 2003, Raish 2015, Salgado-Robles 2011). Yet, some learners fear that adopting nonstandard variation would be inauthentic (Kinginger and Farrell 2004, Ringer-Hilfinger 2012, Trosset 1986), and various pedagogical models and scholarly works advocate against the dangers of overaccommodation, arguing that native speakers expect language learners to set a linguistic target more socially prestigious than the vernacular common to the L1 community (Andreasson 1994, Auger & Valdman 1999, Christophersen 1973, Saville-Troike 2003, Valdman 1988). The question of how native speakers perceive L2 use of nonstandard variation has received little attention in terms of empirical studies, and results have been conflicting among the few works which have investigated the issue. Some have found that L1 speakers respond negatively to such language use (Prodromou 2007, Ruivivar & Collins 2018, Swacker 1976) while others indicate the opposite (Beaulieu 2016; George 2013, 2014, 2017).

The present study sought to clarify these conflicting findings by investigating the attitudes of L1 speakers of Peninsular Spanish to the use of colloquial lexical items by L1 speakers of English. The study utilized a matched guise methodology (Lambert et al. 1960), asking over 200 participants to evaluate audio samples of L1 and L2 speakers using colloquial language which was selected according to a corpus analysis of Peninsular Spanish. Participants assessed the speakers according to factors such as linguistic proficiency, the appropriateness of language use according to context, and personality traits corresponding with the dimensions of status, solidarity, and dynamism common to language attitudes research (Giles & Billings 2004, Zahn & Hopper 1985). The results of the study found that the use of colloquial language did not quantitatively correlate with significantly higher or lower measures of linguistic proficiency on a broad level, despite its use being highly salient to participants. In terms of measures of contextual appropriateness and personality ratings, the study found that L2 speakers experience many of the same disadvantages suffered by native speakers when using nonstandard language—for example, being downgraded on personality characteristics associated with status (successful, intelligent, etc.)—but they do not have the same level of access to corresponding advantages, such as increased ratings of traits indicating solidarity (sociable, friendly, etc.). While these results verify to some degree the existence of a double standard in attitudes toward nonstandard language use according to the speaker’s L1, the lack of concrete benefits available to L2 speakers using colloquial language constituted a bigger drawback than any outright negative associations made with its use; what negative associations did exist were by and large those experienced by L1 speakers, and these were not so negative as to justify theoretical arguments in the literature advocating the wholesale avoidance of its production. Instead, the results of this study suggest that so long as L2 users are aware of the potential risks associated with using such language, there seems to be no reason to discourage its production.