Adjudication Surveys - Graduate Student and Postdoctoral Fellow Research Day 2023

A range of studies suggest that translation is a valuable pedagogical tool for language learning (e.g., Beiler & Dewilde, 2020; Canagarajah, 2013; Galante, 2021; Sneddon, 2012). Nevertheless, the process of translation and the identity of the translator are invisible in much research on translation. To address these gaps in the field, the author drew on a self-study of her translation of English children’s stories into Bangla and used narrative inquiry methods to investigate the process of translation and the negotiation of identity. Drawing theoretically on Norton’s work on identity (Darvin & Norton, 2015, 2023; Norton, 2016), the author investigated how she navigated her translator identity with respect to investment, capital, and ideology. The author developed a “continuum of equivalence” model to capture the shifts in her identity as a translator in the translation process (Author, 2022). By discussing translation challenges, the author illustrates how her linguistic and cultural capital, her understanding of ideologies, the structure of power of translation guidelines, and her habitus influenced her investment in the translation process and, thus, her translator identity. Drawing on her study, the author makes the case that the model could enhance the use of translation as a pedagogical tool to promote language learners’ critical language awareness, multicultural knowledge, and creative thinking. This study makes visible the identity of the translator, an important stakeholder in the promotion of multilingualism in language education internationally.

Many Indigenous scholars assert that Indigenous languages are inseparable from the land. Loss of language and loss of land are often interconnected, and reclaiming an ancestral language is one way in which communities can continue to steward their land [1]. Despite the digital divide that persists in many Indigenous communities, digital technologies are also being used by communities to foster relationships with the land and facilitate language learning [2]. The purpose of this research is to explore the potential use of digital technologies to support Indigenous language reclamation and land stewardship initiatives, using a community-based research approach that actively engages community members as collaborators at every stage [3]. In this presentation, I will discuss the development of the Tlingit Language and Land App, a project developed collaboratively with the Taku River Tlingit First Nation Land Guardians [4] as an extension of the existing TRTFN Place Names web platform [5], which hosts language, stories, and a map of traditional place names. Because there is no access to internet or cell service on the land, the Tlingit Language and Land App must be operational offline. This project complements existing work on land use and place names reclamation that is currently underway in the community. It will provide a resource for information on place names that community members can access easily, and will contribute to language revitalization by providing a new way of visualizing language on the landscape and by encouraging language learners to engage with language on the land.

Decades of infant neuroimaing research has illuminated the brain correlates of speech perception [1], however, there is a gulf between an infant’s experiences and how these experiences are studied. Most studies are highly constrained, with structured screen-based tasks [1]. We aimed to bridge this gulf by measuring infant’s brain responses to mother’s speech during a session of free play. Nineteen 5- to 7-month-olds (13 female) engaged in face-to-face play with their mother, while we recorded the interaction using cameras and infants’ brain activity using a functional near-infrared spectroscopy (fNIRS) system. Mothers were free to play as usually at home while keeping the infant relatively still and playing with only one toy at a time. The recording continued for as long as the infant wanted to play (M = 13.4 min). Thirteen infants played for over 5 minutes (minimum criteria) and contributed high quality fNIRS data. We manually coded the video-recordings marking each bout of infant-directed speech, adult-directed speech, and adult labelling behaviours on the fNIRS signal. Ongoing work in our lab aims to quantify the strength and location of brain responses to the perception of speech. To the best of our knowledge, this is the first infant fNIRS study to measure the neural correlates of speech processing during free play. Our results will eludicate whether previous findings translate to real-life, naturalistic situations. The devised methods to collect and analyse infant brain responses in unstructured scenarios will open the door for more ecologically valid and culturally sensitive research.

This study examined L2 speech self-assessment from a Complex Dynamic Systems Theory (CDST) perspective. CDST prioritizes inquiry through the lens of interconnected parts that undergo constant, dynamic, and nonlinear change (Hiver and Al-Hoorie, 2019). The theory provides a promising new direction for L2 acquisition research (Larsen-Freeman, 2020), though it has yet to be applied to the topic of L2 self-assessment. Traditional research on L2 speech assessment, using static global ratings, has produced mixed results regarding the degree of alignment between learners’ self-assessment and assessment by external judges (Li and Zhang, 2021). To address this, the study analyzed whether self-assessment can be characterized as a complex and dynamic process, and compared time-aligned, continuous assessment with global assessment in terms of degree of agreement between speaker and listener ratings. Twenty-four L2 speakers of English performed two extemporaneous speaking tasks. They self-assessed their recordings for fluency and comprehensibility, both dynamically (moving a computer joystick in response to perceived changes in the speech) and globally (with a one-time overall rating). A cohort of 40 L1 English listeners assessed the recordings using the same procedures. In a debrief interview, participants elaborated on the attributes of the speech that influenced their assessment. In contrast with previous research, L2 speakers’ self-assessment, both global and dynamic, was similar to assessment by listeners. Analysis of the dynamic ratings underscored the complex and fluctuating nature of comprehensibility and fluency assessment, by both the listeners and the speakers themselves-- though there were substantial individual differences in the extent of the dynamicity.

Many newcomers to Canada experience significant difficulties as they adjust to life in their new communities, with few being more challenging than their journey to become English speakers. While Canada’s espousal of a welcoming, multicultural national identity, as well as its federally funded Language Instruction for Newcomers to Canada (LINC) program, suggest cohesive social integration, the historical consequences of the discourse of neoliberal multiculturalism and exclusionary language ideologies, such as the norm of monolingualism and standardization (which currently inform the LINC program), have only served to assimilate and marginalize newcomers. The abrupt shift to virtual and hybrid learning models in response to the COVID-19 pandemic only complicated and exacerbated many of the issues that LINC students were already facing. Using the research methodologies of participant observation and semi-structured interviews, this project offers an ethnographic analysis of a LINC class located in Surrey, BC, one that was contending with precisely the new reality of English language learning that COVID-19 had created. This study concludes that, despite the challenges of online learning, the incorporation of digital technologies into the LINC class has created a new space outside of the classroom wherein a transformative, translingual pedagogy may take root. With careful implementation, the adoption of a translingual pedagogy has the potential to both work against the problematic discourses perpetuating within the LINC program, as well as improve the students’ English language learning outcomes by providing increased opportunities for digital literacy socialization.

Self-supervised speech models have been used in a variety of tasks. However, it is still unclear whether the algorithms behind these models exhibit human-like perception or not (Schatz & Feldman 2018, Millet et al. 2019). One important feature of human perception is known as categorical perception (CP), which refers to the general tendency of a sudden change in identification at the perceptual boundary and a higher cross-category identification accuracy than within-category, even when the differences between stimuli are equal in a continuous manner, like speech (Liberman et al. 1957). The most tested speech sounds of CP are voicing of English stops (voiced: /b, d, g/; voiceless: /p, t, k/) as they can be easily modified through the acoustic feature known as Voice Onset Time (VOT). When VOT vary along a continuum, native English listeners tend to categorize stops below a certain threshold as voiced and stops above it as voiceless. The specific values for such categorical boundaries differ depending on the place of articulation for the English stops (/g/-/k/ > /d/-/t/ > /b/-/p/; Lisker & Abramson 1970, Nakai & Scobbie 2016). This project evaluates whether the speech models’ performance in categorizing the voicing of English stop consonant continua also show human-like categoricalness. The pre-trained wav2vec2 model (Baevski et al. 2020) was fine-tuned to perform classification task on speech continua based on the voicing of word-initial stops. Preliminary results show that the model also exhibits categorical perception, while values of such categorical boundaries differ from human perception.

Aspiration is a well documented feature of sound systems cross-linguistically but is under-researched in Salish languages. Many Salish grammars offer brief statements about aspiration, indicating only whether it occurs, often with no evidence (Montler, 1986; van Eijk, 1997). This project investigates the aspiration system of Nɬeʔkepmxcin (Thompson River Salish), a Northern Interior Salish language. Unlike other Salish languages, a detailed pattern of aspiration is outlined in the Nɬeʔkepmxcin grammar (Thompson & Thompson, 1992). Within the proposed system, stops are categorized as “unaspirated before vowels and resonants, but often somewhat aspirated before a spirant and regularly before another stop. In syllable-final position they are strongly aspirated” (Thompson & Thompson, 1992: 4). No evidence is provided for this split and clear definitions of the meanings of somewhat, regularly, and strongly are not provided. To assess the described system, data were collected from three fluent speakers of Nɬeʔkepmxcin in weekly Zoom meetings. The first stage of this project measured release burst durations of voiceless stops in the “unaspirated” and “strongly aspirated” contexts. Voiceless stops in these positions show significant differences in release burst duration with shorter release bursts before vowels (32ms) and longer release bursts in word-final stops (149ms; t(162) = 19.22, p < 0.001). Data collection for the “somewhat” and “regularly” aspirated contexts is ongoing, but preliminary measurements suggest that the descriptors used by Thompson & Thompson (1992) correspond to release burst duration rather than providing a description of regularity in the use of aspiration by speakers.

Postural maintenance in the vocal tract (VT) is achieved through tonic muscle activity (TMA) that counteracts gravity (e.g. when maintaining jaw occlusion) [1]. Substantial research identifies gamma (γ) and high-gamma (Γ) waves in the sensorimotor cortex (SMC) underlying transient speech movements [2] yet little is known about neural control of VT postures. In gross motor skills (GMS), beta waves (β) correspond to inhibition and state maintenance integral to posture; e.g. β decreases at movement onset, increases at movement termination, and corresponds to postural complexity: the TMA necessary to counteract gravity during postural maintenance [3,4]. Here we use public electrocorticography (ECoG) data [2] to assess β/γ/Γ contributions to speech and VT posture. Participants with cortical electrodes (inferior SMC) spoke CV syllables separated by brief interstimulus-rest intervals. Brief rest results in interspeech posture (ISP) where VT articulators are maintained in efficient positions for speech [5]. We compared mean β/γ/Γ power across speech and ISP, and found strong anti-correlations of β with γ/Γ. We show that β responds to fine motor skill (FMS) movement onset/termination and corresponds to inhibition of discrete phonemic movements and maintenance of VT posture. Next, we isolate the effect of postural complexity by comparing VT postures in three orientations agonistic or antagonistic to gravity (upright, supine, and upside-down via an inversion table). Mean β power is compared across counterbalanced posture/orientations. We predict β will increase in proportion to postural complexity, following a unified approach to postural control across GMS/FMS.

In Hong Kong Cantonese, /n/ is categorically produced as [l] syllable-initially in a diachronic sound that has been observed since the early twentieth century (Ball, 1907, Cheng et al., 2022). In this sound change, words such as nou5 腦 “brain", historically pronounced with /n/ are pronounced with [l], producing homophones with pre-existing /l/-initial words, such as lou5 老 “old". Sociolinguistic work has shown that these pronunciations bear social weight ([n] pronunciations are prestige variants, while [l] pronunciations are socially stigmatized; Cheng, 2017; Pan, 1981) but little work has examined the consequences of these sound changes for speech perception and lexical processing. I test Cantonese talkers on the perception, recognition, and encoding of these sound change pronunciation variants through four online experiments. An immediate repetition priming paradigm with [l] targets (Experiment 1) demonstrates recognition equivalence between [n] and [l] forms, in spite of clear phonetic sensitivity to [n] and [l] in AX discrimination (Experiment 2a) and categorization tasks (Experiment 2b). A long distance repetition priming task (Experiment 3) establishes recognition equivalence in long term recognition as well, suggesting listeners encode both pronunciation variants by dually mapping them to a single lexical representation (Samuel & Larraza, 2015). These data suggest that regular exposure to sound change variants supports perceptual flexibility to multiple phonetic forms on a language-specific basis. This study contributes to our understanding of the sound change and uniquely situates the study of phonetic variation in the context of diachronic sound change variants.

This is an ongoing project whereby examining soundtracks of speakers of Tsimshian English doing a monologue, and from that, previously claimed phonological processes present in literature will be justified. Additional findings of processes not included in the literature will also be presented. Analyses will be done to explain the phenomena. The research questions of this study include: what are the distinctive phonetic and phonological features present in this dialect? And why and how do these features arise? This project is based on the identified phenomena in Mulder’s (1982) work, which was proposed through an examination of his fieldwork data. The results from this analysis will be useful in understanding the grammar of the Tsimshian language itself while providing insights into how language contact occurs in this scenario. The hypothesis is that there will be systematic phonological differences and variations in the data collected and analyzed, which were motivated by the grammatical structure of the Tsimshian language and limiting the adaptation of the English words. The study was performed by collecting and looking into audios of speakers of the dialect from an online archive and analyzing tokens of interest acoustically and phonologically. The topic of vowel lengthening before voiced and voiceless stops is the main focus at this stage of the study, where the lengths of vowels in these two environments from the data will be analyzed. Consonant variations and analyses were also anticipated and are potential interests for further research in this project.

While opening/closing of the velopharyngeal port (VPP) in speech has been much studied, the speed of these movements has been largely overlooked. The present study compares opening/closing velocities of the VPP in French and English, testing relation to distance traveled and speech-related vs. physiological movements. Running speech samples from 9 Quebecois French speakers and 4 Canadian English speakers were obtained from the Université Laval X-ray videofluorography database [Munhall et al., 1995 J Acoust Soc Am, 98(2), 1222-1224]. Using ImageJ software, we tracked VPP opening/closing movements during two types of events: phonologically nasal segments and rest intervals between chunks of speech. We calculated velocity of VPP opening/closing during these events and analyzed the data using linear mixed-effects models to identify differences between the nasals and rest intervals as well as for any cross-linguistic differences. Results indicated that: 1) VPP closure was faster following English nasals than French nasals; 2) VPP opening was faster than closure for rest intervals in French; 3) VPP opening/closing speeds were faster coming into/out of rest position than into/out of nasals in French. Preliminary cross-language observations support a correspondence between velocity and distance. [work supported by NIH and NSERC].

Autocomplete is a task where the user inputs a piece of text, termed prompt, which is conditioned by the model to generate semantically coherent continuation. Existing works for this task have primarily focused on datasets (e.g., email, chat) with high frequency user prompt patterns (or focused prompts) where word-based language models have been quite effective. In this work, we study the more challenging setting consisting of low frequency user prompt patterns (or broad prompts, e.g., prompt about 93rd academy awards) and demonstrate the effectiveness of character-based language models. We study this problem under memory-constrained settings (e.g., edge devices and smartphones), where character-based representation is effective in reducing the overall model size (in terms of parameters). We use WikiText-103 benchmark to simulate broad prompts and demonstrate that character models rival word models in exact match accuracy for the autocomplete task, when controlled for the model size. For instance, we show that a 20M parameter character model performs similar to an 80M parameter word model in the vanilla setting. We further propose novel methods to improve character models by incorporating inductive bias in the form of compositional information and representation transfer from large word models.

Introduction: Dyslexia is a neurodevelopmental reading disorder associated with deficits in motion perception [1]. These deficits have been attributed to vulnerability in the dorsal visual processing stream [2] involved in motion, attention, and visually-guided action [3]. Contrary to this traditional view, an alternative framework proposes a lateral stream, that includes visual areas MT and lateral occipital complex (LOC) [4]. This lateral stream is proposed to mediate both motion perception and language processing [5], and its function may be disrupted by dyslexia [6]. To investigate whether the brain’s functional connection is altered in either or both visual streams in dyslexia, we examined functional connectivity between regions of interest (ROIs) in the dorsal and lateral streams. Methods: Resting-state fMRI data collected from children with dyslexia (N=11) and typical readers (N=17) in grade 3 were analyzed [7]. Whole-brain ROI-to-ROI functional connectivity between the time series of ROIs in the dorsal stream (intraparietal sulcus, visual areas 3d, 6, and 7) and the lateral stream (MT, anterior and posterior LOC) was compared between the groups. Results: Compared to typical readers, children with dyslexia showed weaker functional connectivity between the ROIs located in the dorsal stream, and between the ROIs in the lateral stream. Conclusions: Our results showed weaker within-stream connections both in the dorsal and lateral visual streams in dyslexia, suggesting that the dorsal stream vulnerability hypothesis may require reconsideration. Future research will investigate the functional connectivity between the dorsal and lateral streams and language networks (e.g., inferior frontal gyrus and superior temporal gyrus [8]) in dyslexia.

It has been a well-established concept that body posture involves tonic muscle activation and it acts against gravity (Ivanenko & Gurfinkel, 2018). While body posture is fundamental to movement in gross motor skills, it remains unclear on what the role of posture is in fine motor skills. For example, current speech motor control models have not considered the effect of gravity on speech postures and movements (Parrell et al., 2019). Fine motor skills have been shown as sensitive to gravitational changes. For instance, Shamei et al. (2022) examined astronauts’ speech immediately before and after exposure of microgravity in space, and found a lowered vowel space after landing. The findings suggest that the posture of the jaw and the tongue adapted to microgravity during flight as tonic muscle activation might not be necessary to counteract gravity in such an environment. Their findings further indicate that similar to gross motor control, posture in fine motor skills also counteracts gravity and is an essential component of fine motor skills. Due to the similarity, we aim to build a unified approach to posture control. Inspired by a well-established posture-movement coordination parallel model from Massion et al. (2004), we built a speech posture-movement parallel control model in Simulink. This model consists of two parallel control systems for vocal tract structures, one for postural control, and the other for movement control. A simplified model will be presented and simulation results will be compared to activations of muscles maintaining a smiling posture and muscles producing bilabial tokens.

French liaison is a highly idiosyncratic external sandhi process in which segment alternations occur at the morpheme boundary of certain morphosyntactic contexts (e.g. the plural determiner phrase in which “les” is pronounced [le] in 'les tables' ‘the tables’ (C-initial) but [lez] in 'les amis' ‘the friends’ (V-initial)). Over the years, multiple competing representational accounts have developed to explain the alternating liaison consonant at the juncture between two morphemes (see Cote (2011) for an overview). One of the properties of liaison that any comprehensive analysis should take into account is the fact that it has several types of exceptions. Following Tessier & Jesney (2020), this study aims to sketch a learning account that explains the regular cases as well as the exceptions. We will consider two research questions: i. what kind of representational account to adopt that captures the speaker’s phonological knowledge of the liaison consonant's behavior both in the regular and exceptional cases (cf. Smith, 2015, Smolensky & Goldrick, 2016) and ii. what type of grammar can learn those representations through one of the learning algorithms such as an error-driven gradual algorithm (Boersma & Pater, 2016). In the next step, after building the learning model we will conduct an artificial learning experiment with nonce words to evaluate the predictions that the learning model makes not only about the real words in the lexicon but also about the novel words (Tessier & Jesney, 2022). The results of such a study will shed light on learning of exceptions along with regularities in phonological theory in addition to learning of French liaison in particular.