Recently, neural natural language models have attained state-of-the-art performance on a wide variety of tasks, but the high performance can result from superficial, surface-level cues (Bender and Koller, 2020; Niven and Kao, 2020). These surface cues, as the “shortcuts" inherent in the datasets, do not contribute to the task-specific information (TSI) of the classification tasks. While it is essential to look at the model performance, it is also important to understand the datasets. In this paper, we consider this question: Apart from the information introduced by the shortcut features, how much task-specific information is required to classify a dataset? We formulate this quantity in an information-theoretic framework. While this quantity is hard to compute, we approximate it with a fast and stable method. TSI quantifies the amount of linguistic knowledge modulo a set of predefined shortcuts -- that contributes to classifying a sample from each dataset. This framework allows us to compare across datasets, saying that, apart from a set of "shortcut features", classifying each sample in the Multi-NLI task involves around 0.4 nats more TSI than in the Quora Question Pair.

Spanish and English differ in vowel inventories (5 vowels vs. 12, respectively) and Spanish vowels are shorter in duration, and higher or lower in F1 and F2, than their English counterparts. (Cobb & Simonet, 2015; Colantoni et al., 2015; Hualde, 2014; Menke & Face, 2010; Schwegler et al., 2010) This study investigated differences in Spanish vowel duration, F1 and F2 between L2 Spanish-L1 English and native Spanish speakers. It was hypothesized that, compared to native Spanish speakers, L2 learners (especially lower proficiency) would produce stressed /e/, /i/, /o/ and /u/ with longer duration and unstressed /a/ with shorter duration, /a/, /o/ and /u/ with higher F1 and /e/ and /i/ with lower F1, and all 5 vowels with higher F2.  

Data from the University of Toronto Romance Phonetics Database (Colantoni & Steele, 2004) were analyzed for 3 L2 Spanish-L1 English learners (1 intermediate, 1 advanced, 1 near-native) and 3 L1 Spanish speakers, for 50 words (10 per vowel, 5 stressed, 5 unstressed).   Results showed that, for L2 learners at all proficiency levels, compared to L1 Spanish speakers: Duration may be longer for stressed /a/ only, and equal for other vowels, but F1 and F2 may be higher or lower for all vowels.  

This study provides some evidence that for Spanish vowels in L2 learners with L1 English, vowel, stress and proficiency level combined may influence duration, F1 and F2, and that these L2 learners may achieve target-like production more quickly for some parameters (duration) than for others (F1, F2).

This oral presentation presents quantitative research on a Mandarin phonetic difference which results in a pragmatic meaning change. The topic of the research is a suprasegmental change in terms of pitch and particle duration with the meaning variation of agreement and disagreement associated with the Mandarin discourse particle en. Theoretically, en will be produced with shorter duration and falling pitch when it is used with the meaning of agreement. On the other hand, it will be produced with longer duration (longer than 350ms) and rising pitch when it has the disagreement meaning. However, informal observation suggests that Mandarin speakers from different regions (south vs. north) pronounced this particle differently. This talk presents a study of the variation in the pronunciation of en among speakers from four Chinese cities (two from the south: Guangzhou and Shanghai; two from the north: Beijing and Shenyang). The data of the particle en is collected by the virtual sociolinguistic interview with Mandarin speakers living in either China or Canada. The duration, maximum pitch, mean pitch and the tone shape were measured and analyzed. The results of the research indicate that the variation might not be due to the regional difference (i.e. south vs. north), while the difference between speakers from four different cities is more obvious and straightforward. Shanghai and Shenyang speakers used both suprasegmental features, duration and pitch. They produced the short particle with the falling tone to produce the particle with the agreement meaning. In contrast, they lengthened the particle and used the level tone to indicate the disagreement meaning with the particle. Speakers from other two cities only lengthened the duration of the particle when the particle is used with the disagreement meaning. They did not have changes to the pitch.

In this work, we evaluate various existing dialogue relevance metrics, find strong dependencies on the dataset, often with poor correlation with human scores of relevance, and propose modifications to reduce data requirements while improving correlation. With these changes, our metric achieves a new state-of-the-art on the HUMOD dataset. We achieve this without fine-tuning, using only 3750 unannotated human dialogues and a single negative example. Despite these limitations, we demonstrate competitive performance on three datasets from different domains.

I am fourth year PhD student in the department of Computer Science where I design technology for informal language and culture learning contexts. In my current project, I am uncovering the needs for language and culture support in multigenerational immigrant families. In multigenerational immigrant families, everyone is a lifelong learner. Grandparents must learn to foster social connection with their grandchildren despite language and culture barriers, while grandchildren seek to learn their heritage language and culture to better connect with their grandparents. Educational support tools are sparse as language learning apps are often Eurocentric in design and do not fit in the existing routines of immigrant families. In my talk, I will discuss an ongoing project that addresses the two main shortcomings of existing technology. First, I employ human-centered approaches that are better suited for uncovering the practices of marginalized immigrant populations. Second, I design and evaluate a tool that fits within the routines of immigrant grandparents and grandchildren by employing the familiar activity of storytelling. I will also provide a brief overview of my PhD objectives, which examine how technology mediation can support the learning process, whether through social collaboration, AI-based support, or novel modes of interaction. In my PhD, I bridge the domains of computer science, education, language learning, and psychology to inform technology design. At this talk, I will also touch on my future research projects and opportunities for collaboration with researchers in the language learning space.

The Royal Bank of Canada is not "royal" in the sense of having any affiliation with the Crown, and yet the following is not a contradiction: "the Royal Bank of Canada is not a royal bank". Although names like "Royal Bank of Canada" have lexical content, there is no expectation that they have compositional meaning. In contrast, a noun phrase like "royal family" does presuppose that the referent be royal. Names are like idioms in that they do not necessarily reflect the sum of their parts: just as the "Royal Bank of Canada" does not have to be royal, "letting the cat out of the bag" does not have to involve cats or bags. Furthermore, there seems to be a cognitive threshold between a description ("royal bank of Canada") and a label ("Royal Bank of Canada") that mirrors the distinction between a literal expression ("we freed the cat from the bag") and a figurative one ("we let the cat out of the bag").

My research argues that naming is part of the syntactic derivation of a lexical item ("bank" → "Bank"), the same as number ("bank" → "banks") and determination ("bank" → "the bank"). The effects of naming can be observed with both inflectional and derivational morphology. When naming precedes number inflection, regularization occurs ("maple leaf" → "Maple Leaf" → "Maple Leafs"), but when naming follows number inflection, irregular morphology is preserved ("timberwolf" → "timberwolves" → "Timberwolves"). With derivational morphology, demonyms show that naming follows the addition of lexical material ("New York" → "New Yorker") but precedes that of grammatical material ("Bronx" → "Bronxite", not "the Bronx" → "*the Bronxite"). This account of naming as a syntactic process captures what is otherwise puzzling data and predicts how naming affects the interpretation of a lexical item. 

Previous virtual care literature within speech-language pathology has primarily focused on validating the virtual use of intervention programs. Fewer studies have addressed the validity of conducting virtual assessments, particularly standardized assessment of oral language and literacy abilities in children. Additionally, limited practical recommendations are available for clinicians and researchers on how to conduct these assessment measures virtually. Given the rapid rise in virtual care and research as a result of the COVID-19 pandemic, clinicians and researchers require guidance on best practices for virtual administration of these tools imminently. Methods: We (a) completed a narrative review, and (b) conducted semi-structured interviews and recommendations rating process with a group of 12 clinicians, students and researchers who had administered standardized language and literacy assessments with a variety of monolingual and multilingual school-aged children, with and without speech and language difficulties, in clinical and research settings. Results: We have outlined recommendations for clinicians and researchers to guide their use of standardized language and literacy assessments in virtual care, across six key themes: candidacy for virtual assessment, communication and collaboration with caregivers, technology and equipment, virtual administration, ethics, consent and confidentiality, and considerations for bilingual populations were identified as a result and were used to develop a set of recommendations. Conclusions: This paper is one of the first to share such recommendations. We hope the current recommendations will facilitate future clinical research in this area, and that this paper will act as a basis for the development of formal Clinical Practice Guidelines.

Reading, a uniquely human process, involves the interpretation of written language. Difficulties in reading or reading disabilities (RD) come about when a child struggles with a part of this process, specifically the sounding out/recognition of words. The etiology of RD is not well known; however, researchers hypothesize that genetic variants lead to disruptions in neuronal migration, which affects connectivity in language centers in the brain. To contribute to understanding the genetics of reading, we previously performed a genome-wide association study (GWAS). We found top variants were in genes involved in neuronal migration or implicated in other neurodevelopmental disorders, such as Autism Spectrum Disorder (ASD). Using this information, we then performed a hypothesis-driven (HD) analysis (HD-GWAS) where we upweighted variants, separately, based on their implication in either neuronal migration or ASD. This analysis was conducted on a family-based RD-selected cohort from Toronto and a large meta-analysis from the GenLang Consortium. For the Toronto sample, no variants reached statistical significance; however, we identified that the joint contribution of ASD genes significantly contributed to word reading. For the GenLang sample, we identified a significant variant on chr 21 (q21.1). This locus was previously identified by traditional GWAS and not up-weighted, confirming the robust association. Our results suggest that ASD risk genes are enriched in word reading; however, this finding was limited to the Toronto sample. The relationship between RD and ASD has not been thoroughly investigated, but both are language-based disorders, suggesting a possible common neurobiological link. 

Language is inherently flexible: people continually generalize over observed data to produce creative linguistic expressions. This process is constrained by a wide range of factors, whose interaction is not fully understood. We present a novel study of the creative use of verb constructions “in the wild”, where we use computational modeling techniques on data from a very large social media corpus.  

Our first experiment confirms on this large-scale data the important interaction of category variability and item similarity within creative extension. More specifically, we show that people will be more stringent in extending a “low variability” construction, only permitting coinages that are very similar to attested exemplars, but will extend a “high variability” construction more freely, attending less to the similarity of a novel coinage. This effect has previously been demonstrated in an artificial language learning experiment (Suttle & Goldberg, 2011). Here, we test this hypothesis on our constructions in the wild, finding support for the interaction of variability and similarity in actual creative language use.  

Our second experiment confirms the novel hypothesis that low-frequency exemplars may play an important role in generalization. This builds on work on morphology which has shown that a high proportion of low-frequency examples (specifically, singletons) is a signal to language users that a construction is productive (Baayen & Lieber, 1991; Pierrehumbert & Granell, 2018). We go beyond this by showing that low-frequency exemplars indicate not only that a construction generalizes, but where in semantic space the generalization is happening.

Speech communication involves production—which maps discrete phonemes to infinite distinct articulatory realizations—as well as perception, which maps physical signals back to the phonemes. Differences between the realizations of phonemes are often abstracted as distinctive features with respect to articulatory gestures. Distinctive features have generally been manually devised. Here, we conduct unsupervised feature extraction from human behavioral data, including both perception and production errors. 

Previous works on perceptual confusions (e.g., Miller & Nicely, 1955) have adapted information theoretic metrics for evaluating how well a priori features are transferred from acoustic signals to perceived phonemes. We extend such logic to unsupervised feature extraction based on mutual information. In our experiment, features are extracted based on English production errors from the Fromkin Speech Error Database and English perception errors from Miller & Nicely (1955). Production and perception confusion data show different patterns and yield two rather different feature sets. By concatenating the two feature sets, we obtain a feature set that is better correlated with both an English distinctive feature set and cross-linguistic phoneme dissimilarities, suggesting that phonological classes are related to both perceptual and articulatory similarity.

The COVID-19 pandemic has brought unprecedented challenges to French Immersion students this past year. While the switch to online learning continues to be challenging for all learners, French Immersion students face particular challenges since they are obtaining academic instruction in their non-dominant language. In addition, children do not get the same opportunities to practice their social language in an online learning format. Fourteen parents of elementary school children in French Immersion were interviewed in a semi-structured qualitative approach using a video-conferencing platform. Parents were asked their thoughts and beliefs concerning the French Immersion program and its instruction during the COVID-19 pandemic. Specifically, parents were asked about their child(ren)’s access to technology, any benefits or challenges associated with online learning, how they felt online teaching could be improved, and how much support they felt that they received from their child(ren)’s school and teacher. Overall, many parents felt that the French Immersion program lacked proper funding and resources to support their child(ren)’s French language learning during the pandemic. Many parents shared that they were forced to look outside the school for online language learning resources to supplement the curriculum, such as online educational applications or community resources. Other parents expressed frustration with having to help teach their children French, despite not speaking it themselves. Overall impressions from the interviews are discussed and suggestions regarding future directions are proposed. 

When speakers use a referring expression, such as “the dog” or “the small black dog”, their goal is for a listener to identify the intended object. This is challenging because speakers and listeners often have different perspectives – e.g., at a dog park, their respective views may include different subsets of the dogs. To address this, speakers engage in audience design, adapting their productions to the listener’s perspective despite the extra cognitive effort required.

Surprisingly, Heller and Stevenson (2018) found that speakers took the listener’s perspective into account in referring expressions, even when that wasn’t necessary for referential success. Why would speakers expend cognitive effort to adapt to the listener unnecessarily? Are speakers not “rational”? Or was the task easy enough that speakers simply had sufficient cognitive resources to adapt even when unnecessary, as would be suggested by the resource-rational framework (Lieder & Griffiths, 2019)? 

We tested this with a web-based experiment using three seconds of time pressure as a proxy for cognitive load, and found that participants performed audience design even when unnecessary. Do speakers not manage their cognitive resources rationally? Or is it that speakers are resource-rational, but they manage resources by adopting a strategy tailored to an overall situational context, rather than determining a strategy for each sentence produced? We’re undertaking a follow-up experiment to test this by varying the proportion of trials where perspective-taking is and isn’t necessary to investigate if speakers in these different conditions adapt their perspective-taking strategies to the make-up of the trials.

Previous work (Brownlow & O’Dell, 2006) has posited that online forums can help autistic people find their voice, as these spaces present fewer communication barriers. However, it is unknown whether there are differences in how autism spectrum disorder (ASD) influences communicative interactions online. Here, we build on work in which automatic language processing techniques have been deployed to detect cognitive differences in large samples of writing. Specifically, we investigate linguistic markers of autism in social media text, and hypothesize that such markers, even when subtle, are detectable at scale.

To do so, we manually filtered self-reported autistic users and compared their comment history with a random sample of Reddit users, using data from the Reddit Pushshift dataset (Baumgartner et al., 2020). We compare their language over linguistic and psycholinguistic features drawn from studies of communication in ASD. After extracting values for these variables, we statistically modelled our data using a linear mixed-effects model. 

Our model found significant differences between the two sets of authors. We found that autistic authors use more impersonal, third-person pronouns than our random sample, and significantly less first-person singular and second person pronouns. We also find that, on average, autistic authors have lower vocabulary diversity and write shorter posts. 

These results demonstrate that autistic authors who participate in online forums show some detectable differences in their language use. Our work is a first step in empirical investigation of how autistic individuals engage in online spaces, which will facilitate future work on their interactions on such platforms.

We examine the development of Hindi from Sanskrit using spectral analysis techniques. These methods, such as Principal Component Analysis (PCA), are exceptionally good at drawing out inconsistencies and discrepancies in a dataset. Language change is especially suited to this task as we are interested in what changes and what remains constant.

We contrast Panini's karaka role scheme - comparable to modern thematic roles - against grammatical case using spectral analysis. We use IIIT Hyderabad's Hindi corpus annotated in a Paninian scheme. Panini's karaka roles classify the correspondence between Sanskrit's rich morphology and verbal arguments. Sanskrit is obviously not Hindi, so any adaptation of this scheme will inadequately capture modern Hindi grammar.

Our analysis produces several noteworthy cases. First, there is a clear relation between the agent role and genitive markers, and the locative role and location markers. Further, we note the failure to place the ergative marker, which has no precedent in Sanskrit, and the postposition 'ko', which has merged several roles and cases. Finally, we draw a connection between seemingly disjoint markers as evidence for missed analysis.

Other work in language change has focused on modelling the system of change (Niyogi and Berwick, 1996; Kodner and Falco, 2018). These approaches are useful in deciphering the process and mechanisms of change, but are less suited to uncovering specific grammatical connections and discrepancies between related languages. We hope this work contributes to the methodology of the latter. 

Introduction

In multilingual, sub-Saharan African countries, children often attend school taught in a language they do not speak (Heugh, 2011). Bilingual education may improve academic outcomes and bilingual children have shown language and literacy advantages (Jasińska & Petitto, 2018), but little is known about the impact of bilingual language environments in multilingual, low-and-middle-income countries. How do home and school language environments impact children’s language and literacy development in rural Côte d’Ivoire? We compared local language (Attié, Abidji, Baoulé, Bété) and French language and literacy outcomes of children attending French-only or bilingual schools from local language-only or bilingual households. We predicted that children with bilingual versus monolingual experience would perform better. We also explored whether French-only and bilingual schools differ in quality, which may mitigate potential advantages of bilingual education.  

Method

Children (N=830; 6-14 years) completed local language and French language and reading tasks. Teachers (N=51) completed a questionnaire measuring school quality indicators (RTI International, 2009; Woodcock et al., 2001).

Results

Children from bilingual households scored higher on local language and French reading tasks while overall, children attending French-only schools had higher language and reading scores. French-only schools had better infrastructure and bilingual school teachers lacked training and materials in local languages.

Discussion

While our results suggest bilingual home environments were advantageous for children’s language and literacy skills, French-only schools were associated with better scores. Bilingual schools were lower quality than French-only schools, likely contributing to poorer outcomes. For children to benefit from bilingual education, education must be high quality.

There are mixed findings regarding the presence of a cognitive advantage in bilinguals. It has been proposed that executive function and attention are positively affected in bilingual populations. In this meta-analysis we tested whether bilingual children exhibit an advantage when it comes to auditory attention, specifically. We reviewed articles that use standardized tests to assess auditory attention in monolingual and bilingual children. We synthesized data from 1102 monolingual and 1172 bilingual children with an age range of 5 to 14 years. Multiple meta-regression model analysis was performed, suggesting that test measures and participant age were significant predictors of the pooled effect size. Specifically, a small bilingual advantage was only found in accuracy measures (g=0.09, 95% confidence interval = [-0.12, 0.30]), but not in response latency (g=-0.34, 95% confidence interval = [-0.57, -0.11]). Taken both measures together, however, bilingual children did not outperform their monolingual peers on auditory attention tests (g=-0.09, 95% confidence interval = [-0.27, 0.09]). Interestingly, a further subgroup analysis of age reveals a developmental change in auditory attention: a bilingual advantage does not emerge until early adolescence while compared to young children. In addition, studies on atypical populations were reviewed. In conclusion, our results suggest that should there be a bilingual advantage in auditory attention tests, it is small and contextual.

Many children who experience ischemic stroke come from bilingual backgrounds. Whether bilingual exposure affects post-stroke development is unclear. Our research evaluates the effects of bilingual vs. monolingual exposure on linguistic/cognitive development post-stroke. An institutional stroke registry and medical charts were used to gather clinical variables and outcome measure performance data on 237 children across 3 stroke-onset groups: 0-28 days, 1 to 12 months and 13 months to 18 years. We used the Pediatric Stroke Outcome Measure (PSOM) administered at several timepoints post-stroke, to evaluate cognitive and linguistic development. While no main effect of language group was found in the pediatric stroke patients on the PSOM subscales, on language expression subscale, among children with a stroke onset between 1 and 12 months of age, bilingual children had better post-stroke performance compared to monolinguals within 10 years after stroke onset. Overall, we did not find any detrimental effects of bilingual learning environment on post-stroke language development. Further, it is a possibility that a bilingual learning environment facilitates some aspects of language development – if the stroke occurs between 1 and 12 months of age, and abilities are assessed within 10 years of post-stroke.